automated reactive thermal evaporation system for transparent conductive coatings automated reactive thermal evaporation system for transparent conductive coatings m. fernandes a,b , y. vygranenko a,b , m. vieira a,b , g. lavareda b,c , c. nunes de carvalho c,d , a. amaral d . a isel-instituto superior de engenharia de lisboa, instituto politécnico de lisboa, lisboa, portugal b cts-uninova, caparica, portugal c departamento de ciência dos materiais, universidade nova de lisboa, caparica, portugal d physics and engineering of advanced materials, universidade de lisboa, lisboa, portugal mfernandes@deetc.isel.ipl.pt abstract — this work presents a fully automated plasmaenhanced reactive thermal evaporation system (rf-perte) that can be used for the deposition of transparent metal oxide films with high reproducibility of their electrical and optical properties. the developed hardware/software platform enables the full control over the critical deposition conditions such as mass flow of oxygen, process pressure, current flowing through crucible and rf-power. for indium oxide films on glass substrates a resistivity of 9×10 -4 ω-cm and a transmittance of 90% in the visible spectral range were achieved without substrate heating. the system is also suitable for the deposition of transparent conducting coatings on a wide range of plastic substrates, for applications in the field of flexible sensors or solar cells. in particular, we have successfully deposited indium oxide on pen (polyethylene naphthalate) sheets with electrical and optical properties approaching the ones of the films deposited on glass substrates. keywords: tco, deposition, transparent electronics. i. introduction transparent conducting oxide (tco) layers on polymeric substrates are an important component of flexible electronics [1]. substrate materials such as polyethylene terephthalate (pet) or polyethylene naphthalate (pen) are being considered for opto-electronic devices due to their high transparency and low cost [2]. the technological challenge is that tco coating should be deposited at low temperatures, desirably below the glass transition point for these plastics. several vacuum techniques such as dc and rf sputtering, ion beam-assisted evaporation, and arc-discharge ion plating have been used for deposition of indium-tin oxide (ito) on polymeric substrates [3-6]. however, the properties of lowtemperature amorphous ito on plastics are substantially inferior in comparison to crystalline ito grown on glass substrates at high temperatures. to overcome the limitation on growing tco films with satisfactory electrical and optical properties on plastic substrates, we have developed a radiofrequency plasma-enhanced reactive thermal evaporation (rfperte) technique [7]. this work reports on an automated rfperte system, which is suitable for deposition of in2o3 based coatings on unheated polymeric substrates. ii. deposition system fig. 1 shows a general design of the rf-perte system. the system is based on a bell jar type vacuum chamber with a diffusion and mechanical pump vacuum pumping group. the typical configuration for thermal evaporation is used, with a distance between the tungsten boat and the substrate holder of 32 cm. for plasma assistance, an rf-electrode in the form of a copper ring is placed at half-way between crucible and substrate holder. an electrically-driven shutter placed about 15 mm below the substrate holder, shields the substrate from oxygen plasma and impurities from the starting of the evaporation process. the oxygen injection into chamber is controlled by a smarttrak 100 series mass flow controller. a genesystm series programmable regulated power supply and a cesartm generator, model 136, are used as dc and rf power sources, respectively. fig. 1. general overview of the evaporation system a. controller hardware all electronic units and electrical parts are linked to a microprocessor based control unit either through analog, digital or communication (rs-232) ports. the control unit is connected to a personal computer through usb interface that enables the system control using dedicated software. in order to keep the controller board as simple as possible, the analog to digital converter (adc) included in the microcontroller was used. as seen in fig.1, an analog input is used for the i-etc: isel academic journal of electronics, telecommunications and computers cetc2016 issue, vol. 3, n. 1 (2017) id-3 http://journals.isel.pt mailto:mfernandes@deetc.isel.ipl.pt measurement of the pressure signal coming from a pfeiffer pkr251, the 10 bit resolution of the existing adc was not sufficient for the determination of the process pressure with enough precision in the range of interest. to overcome this issue, without having to add an external higher resolution adc, a two-step subranging architecture, similar to a pipelined adc, was implemented using two inputs of the existing 10bits converter [8]. in the implemented architecture, depicted in fig. 2, the pressure signal is converted by the microcontroller´s adc in a two-step sequence with the aid of a differential amplifier. in the first step, the voltage from the gauge is measured on one adc input, then this value minus a fraction is subtracted from the pressure signal and amplified 64x. the residue amplifier is fed with an analog signal generated using pulse width modulation technique and a low pass filter. the amplified residue is then captured on the second adc input and the pressure valued calculated from the two adc values. with this simple method the resolution achieved enables the accurate control of the pressure during the process. fig. 2. two-step subranging architecture used. b. software the control software was developed targeting the following tasks: the programmed control of all electronic units; real time monitoring of all process variables and their recording for post-analysis and documentation; automated control of critical process parameters; and the use of recipe files for process reproducibility. the user interface is intuitive and self-explanatory, with the functional blocks for each piece of equipment, in a way that a rapid look is sufficient to get information about system status. a view of the user interface is presented in fig. 3. a pressure chart is also shown and updated in real time allowing the user to follow the pressure variations. in a typical deposition sequence, when the initial values of the process parameters are reached, the program enables deposition in full-auto mode by starting a timer, and then a dedicated subroutine opens the shutter, holds the process pressure at a given value and adjusts the evaporation rate by varying the current through the tungsten boat. when the time has run out, the shutter will be closed and with some delay dcand rf-values will be set to zero to end the process. adjusting the deposition conditions such as pressure, gas flow rate, rf-power, and evaporation rate the deviation from stoichiometry  in in2o3- can be varied in wide range thus producing conducting, semiconducting and insulating films [9,10]. fig. 3. print screen of the user interface. c. results to produce transparent and highly-conductive inox films, an optimal balance between the evaporated metal mass and absorbed volume of oxygen should be reached. this balance can be evaluated by measuring the pressure at steady oxygen flow, pumping speed, and rf-power. fig. 4 shows that the pressure in the chamber decreases, when the evaporation of indium starts, and then stabilizes when the evaporation rate becomes constant. the deviation from stoichiometry in inox was found to be related to the difference between the initial pressure, pin, and deposition pressure, pdep, which is determined by the evaporation rate. the evaporation rate depends on multiple factors such as the dc power applied to the crucible, process pressure, amount of metal still in the crucible, etc. to stabilize the evaporation rate, the dc current through the crucible is automatically adjusted to keep the deposition pressure constant. this approach enables the reproducible deposition of metal oxide films with required stoichiometry simply by setting the differential pressure ∆p value. 0 200 400 600 800 3.0x10 -4 4.0x10 -4 5.0x10 -4 6.0x10 -4 7.0x10 -4 8.0x10 -4 p in p re s s u re ( m b a r) time (sec) shutter open metal evaporation  p p dep . fig. 4. pressure in the chamber during the deposition. m. fernandess et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-3 i-etc: isel academic journal of electronics, telecommunications and computers a series of in2o3- films deposited on glass was obtained by varying the differential process pressure to study their electrical and optical properties. the films were deposited on 1-mm-thick borosilicate glass substrates at pdep = 4.1×10 -4 mbar with p varying from 1.6×10 -4 to 2.7×10 -4 mbar. the oxygen flow rate and rf-power where kept constant at 15 sccm and 70w, respectively. the transmittance spectra of the films are shown in fig. 5. here, one can see that the transmittance in the green-blue spectral range decreases with increasing p, this effect can be ascribed to the growing number of atomic-scale defects as a result of oxygen deficiency. fig. 5. transmittance spectra of in2o3- layers deposited on glass at various differential process pressures. to test the properties of the materials under low temperature deposition conditions, in2o3- films were also deposited on pen (q65fa-100 µm, teijin-dupont). the obtained material is highly conductive and transparent. a resistivity of 9×10 -4 ωcm was achieved under optimized deposition conditions, in line with the literature [11]. the observed resistivity difference between the layers deposited on glass and on pen substrates is within the range of the measurement error. for glass and pen substrates with about 100 nm thick in2o3- coatings, the obtained peak values of visible transmittance were 90% and 85%, respectively (see fig. 6). fig. 6. transmittance spectra of 104 nm thick indium oxide layers on glass and pen substrates. iii. conclusions a simple technique for preparing undoped, conductive and transparent thin films of indium oxide has been developed using the rf-perte method. the resistivity of 9×10 -4 ω.cm was achieved for coatings on pen and glass substrates. inox films on glass and pen substrates show 90 and 85% peak values of transmittance in the visible spectral range, respectively. process automation proved to allow stable deposition conditions and high reproducibility of the fabricated film characteristics acknowledgements the authors are grateful to the portuguese foundation of science and technology through fellowship sfrh/bpd/102217/2014 and to instituto politécnico de lisboa (id&ca project-solwin) for financial support of this research. references [1] cheng c, wagner s, overview of flexible electronics technology. in: wong ws, salleo a. editors. flexible electronics: materials and applications, new york: springer; 2009. p. 1–28. [2] macdonald mk, looney et al. latest advances in substrates for flexible electronics. journal of the sid 2007;15(12): 1075. [3] tseng k-s, lo y-l. effect of sputtering parameters on optical and electrical properties of ito films on pet substrates. appl. surf. sci. 2013; 285-157. [4] kim d-h et al. thickness dependence of electrical properties of ito film deposited on a plastic substrate by rf magnetron sputtering. appl. surf. sci. 2006; 253 409. [5] meng y, et al. molybdenum-doped indium oxide transparent conductive thin films, j. vac. sci. technol. a 2002; 20-288. [6] niino f, hirasawa h, kondo k. deposition of lowresistivity ito on plastic substrates by dc arc-discharge ion plating. thin solid films 2002; 411-28. [8] behzad razavi and bruce a. wooley, “a 12-b 5msample/s two-step cmos a/d converter,” ieee j. solid-state circuits, vol. 27, no. 12, dec. 1992, pp. 1667-1678. [9] frank g, kauer e, köstlin h, schmitte fj, transparent heat-reflecting coatings for solar applications based on highly doped tin oxide and indium oxide, solar energy materials 1983;8:4-387. [10] lavareda g, nunes de carvalho c, fortunato e, ramos ar, alves e, conde o, amaral a, transparent thin film transistors based on indium oxide semiconductor, j noncrystalline solids 2006; 352-2311. [11] s.j. wakeham, et al. low temperature remote plasma sputtering of indium tin oxide for flexible display applications, thin solid films 2009; 518:4-1355. m. fernandess et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-3 i-etc: isel academic journal of electronics, telecommunications and computers a flexible amorphous silicon photovoltaic module for portable electronics a flexible amorphous silicon photovoltaic module for portable electronics y. vygranenkoa, m. fernandesa,b, p. louroa,b, m. vieiraa,b acts-uninova, 2829-516 caparica, portugal belectronics, telecommunications and computer engineering department, isel, lisbon, portugal yvygranenko@deetc.isel.ipl.pt mfernandes@deetc.isel.ipl.pt plouro@deetc.isel.ipl.pt mv@isel.ipl.pt abstract — this article reports on a monolithic 10 cm × 10 cm area pv module integrating an array of 72 a-si:h n-i-p cells on a a thin polyethylene-naphtalate substrate. the design optimization and device performance analysis are performed using a two-dimensional distributed circuit model of the photovoltaic cell. experimental results show that the shunt leakage is one of the factors reducing the device performance. using the lbic technique, the multiple micro-shunts in the n-i-p n-i-p cell were detected. the mechanism of electrical shunts formation is proposed and discussed. keywords: solar cells, thin films, amorphous silicon, flexible substrate. i. introduction flexible hydrogenated amorphous silicon (a-si:h) solar cells on thin plastic substrates are of great interest for a wide variety of engineering applications. flexible devices can be rolled for transportation, installed on curved surfaces, and they are less likely to be damaged by mechanical friction and vibrations. these advantages make possible for portable electronic devices to cover part of their power demand from sunlight. regardless of the intense research efforts directed to the development of flexible a-si:h solar cells on low-cost plastic substrates, these devices still exhibit considerably lower performance in comparison to that for glass-based equivalents. the major technological challenge is deposition of dopedand undoped a-si:h layers with required electronic properties at temperatures lower than that for solar cells on glasses or metal foils [1-3]. the increased shunt leakage in si:h cells on the plastic substrate is another challenging issue [4]. in this contribution, we report on a monolithic a si:hbased photovoltaic module utilizing the 100 µm thick polyethylene-naphtalate (pen) substrate. the impact of the shunt leakage on the device performance and nature of observed micro-shunts in the p-i-n cells are under discussion. ii. device design and fabrication figure 1 shows a photograph of the developed pv module and cross-sectional view of two cells connected in series. the module of 10 cm × 10 cm area consists of 72 rectangular cells on the pen substrate. the individual cells are connected in series forming eight rows with connection pads. the novelty in the device design is in the backside encapsulation and front buffer silicon-oxynitride layers incorporated for device integrity. the a-si:h n-i-p cell integrates the al/cr and zno:al layers as the back and top electrodes, respectively. two 0.3 mm wide al fingers are symmetrically placed on the zno:al electrode to reduce the emitter resistance. the photosensitive area is 1 × 0.8 cm2. in the developed fabrication process, three shadow masks are used in sputtering steps to form the bottom and top electrodes, and top metallization. the first step is sputtering of al/cr layers through mask #1. here, cr serves as an adhesion layer providing a low contact resistance at the metal/n-layer interface. then, the n-i-p stack is deposited using a 13.56 mhz pecvd system at 150 oc substrate temperature. the deposition conditions are reported elsewhere [5]. the formation of zno:al top electrodes is performed by sputtering through shadow mask #2 at 140 oc substrate temperature. to perform via opening, the n-i-p stack is selectively etched through shadow mask #3 in the reactive ion etching (rie) system. finally, the al fingers and contact pads are sputtered through shadow mask #3. a) b) c) fig. 1. (a) photograph of the pv module, and (b) crosssectional and (c) top views of two cells connected in series. i-etc: isel academic journal of electronics, telecommunications and computers cetc2016 issue, vol. 3, n. 1 (2017) id-7 http://journals.isel.pt iii. device modelling the design optimization and device performance analysis are performed using a two-dimensional distributed circuit model of the photovoltaic cell [6]. the circuit simulator spice was used to calculate current and potential distributions in a network of sub-cell circuits, while matlab was used to create so-called “net list” file, to readout data from the output file, and to visualize the results. figure 2a shows a simulated j-v curve. the indicated solar cell performance characteristics are in a good agreement with that for optimized a-si:h solar cells under am1.5 illumination conditions. figure 2b shows simulated potential distribution across the emitter at a maximum power point (vmax = 0.722 v, imax = 4.8 ma). here, the voltage drop a) 0.0 0 2 0.4 0.6 0.8 1.0 0 1 2 3 4 5 6 c ur re nt (m a ) voltage (v) cell area = 0.5 cm2 v max = 0.722 v i max = 4.8 ma v oc = 0.8839 v i sc = 5.6 ma ff = 0.7027 pce = 7.0 % b) c) fig. 2. (a) simulated potential distribution, (b) current density distribution, and (c) joule losses in the tco electrode and metal finger under am1.5 illumination conditions. across the metal finger is 12 mv and voltage variation across the transparent electrode is up to 45 mv at 0.2 and 60 ω/sq. sheet resistances of metal and tco layers, respectively. the lateral current flow in the tco and top metal layers is shown in fig. 2c. the increased current density at the metal finger corners is observed here. the ‘corner effect’ causes the local increase in joule losses. iii. device characterization a. current-voltage characterization device characterization shows that the open circuit voltage and fill factor are lower than the calculated values due to existence of resistive shunts in some cells in the module (fig. 3a). current-voltage characteristics of all individual a-si:h p-i-n cells in the module were measured and analyzed to better understand the issue. figure 3(b) shows comparison of dark i-v curves for cells with various magnitudes of the shunt leakage current. for both curves, the shunt current varies linearly with biasing voltage, what can be represented as a shunt resistance (rsh). the statistics on shunt resistances is shown in figure 4, where the number of cells with certain ranges of shunt resistance in the module is shown in terms of its probability. here, the shunt resistance varies in wide range from 10 to 105 ω. a) 0 2 4 6 8 10 12 0 2 4 6 8 c ur re nt (m a ) bias voltage (v) i sc = 7.52 ma v oc = 11.5 v i max = 6.32 v v max = 7.34 v ff = 54% b) fig. 3. (a) current-voltage characteristics of the module section (18 individual cells connected in series) under am1.5 conditions. (b) current-voltage characteristics of a-si:h n-i-p cells with lowand high shunt leakage. -1.0 -0.5 0.0 0.5 1.0 -1 0 1 2 3 4 c ur re nt (m a ) voltage (v) r sh = 901 ω r sh = 32 kω y. vygranenko et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-7 i-etc: isel academic journal of electronics, telecommunications and computers 101 102 103 104 105 0.00 0.04 0.08 0.12 0.16 p ro ba bi lit y shunt resistance (ω) fig. 4. probability histogram showing distribution of cells with various shunt resistances in a module. b. lbic experiments the performed laser beam induced current (lbic) experiments show the presence of multiple shunts. the ac component of the output voltage was measured when the device impedance was low (< 1 kω). figure 5a shows a lbic signal map indicating the shunt position manifested as decreased lbic signal. in the absence of shunts the signal is constant over the whole area, excluding the metal finger area due to light blockage. using the developed spice model, the lbic signal was also simulated considering a single shunt at different locations (fig. 5b). modelling shows that the shunting microdefect affects the lbic signal from an area a) b) 0 2 4 6 8 10 0.0 0.5 1.0 1.5 2.0 lb ic s ig na l ( x1 04 v ) position (mm) rms=100 ω ι ph = 1 μa fig. 5. (a) lbic signal for a cell with multiple shunts. (b) calculated lbic signals for a 100 ω shunt at different locations. much larger than its actual size since current from the surrounding area is shunted through the defect. the impedance of the cell depends on the micro-shunt position leading to various signal levels. the magnitude of signal also depends on the sheet resistance of the tco layer. this is in good agreement with the experimental lbic results. c. nature of electrical shunts to understand the nature of electrical shunts, we analyzed the change in the surface roughness of all layers comprising individual solar cells throughout fabrication process. afm surface scan of the a-si:h layer has revealed random surface peaks with height up to 150 nm (fig. 6). the density of these defects is much higher than what is expected from external contamination. multiple experiments have confirmed that the defects appear after deposition of dielectric or semiconductor layers, or even after sputtering of metal layers on the pen substrate. the reason is that cyclic oligomers, which are present in pen, can migrate to the surface forming crystals, if the film is held at temperatures >100°c for tens of minutes [7]. the observed substrate defects may form microscopic pinholes, which are filled with highly conductive top layer material at the final process step, resulting in ohmic shunts. fig. 6. surface chart of the a-si:h film on the polyethylenenaphtalate (pen) substrate. iv. conclusion this study shows that the shunt leakage is a major factor reducing the performance of a si:h solar cells on pen plastic films. the presence of multiple ohmic micro-shunts is confirmed by the lbic technique. the formation of microshunts is attributed to surface defects in plastic foils, which are thermally induced during the device fabrication. acknowledgements the authors are grateful to the portuguese foundation of science and technology through fellowship sfrh/bpd/102217/2014 for financial support of this research. y. vygranenko et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-7 i-etc: isel academic journal of electronics, telecommunications and computers references [1] x. bories-azeau, w. a. macdonald, and j. m. mace. latest developments in polyester films for thin film photovoltaic applications. proc. 5th world conference on photovoltaic energy conversion, 6-10 september 2010, valencia, spain. [2] d.d. fischer et al. development and investigation of thin film solar cells on flexible substrates using very high frequency plasma enhanced chemical vapor deposition (vhfpecvd) technique. proc. 29th european photovoltaic solar energy conference and exhibition, 2010. [3] j. k. rath, m. brinza, liu y., a. borreman, and r.e.i. schropp. fabrication of thin film silicon solar cells on plastic substrate by very high frequency pecvd, sol. energ. mater. sol. cell., vol. 94, pp.1534 -1541, 2010. [4] m. fortes, e. comesana, j.a. rodriguez, p. otero, and a.j. garcia-loureiro. impact of series and shunt resistances in amorphous silicon thin film solar cells, solar energy, vol. 100, pp. 114-123, 2014. [5] k. h. kim, y. vygranenko, d. striakhilev, m. bedzyk, j.h. chang, a. nathan, t.c. chuang, g. heiler, and t. tredwell. performance of a-si:h n–i–p photodiodes on plastic substrate, j. non-crystal. solids, vol. 354, pp. 2590-2593, 2008. [6] y. vygranenko, m. fernandes, m. vieira, a. khosropour, and a. sazonov. a distributed spice model for amorphous silicon solar cells. energy procedia, vol. 60, pp. 96 – 101, 2014. [7] macdonald, m. k. looney et al. latest advances in substrates for flexible electronics. journal of the sid 15, pp.1075 -1084, 2007. y. vygranenko et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-7 i-etc: isel academic journal of electronics, telecommunications and computers an implementation on gnuradio of a new model to isdb-tb using fbmc an implementation on gnuradio of a new model to isdb-tb using fbmc j. almeida, c. akamine, p. b. lopes programa de pós graduação em engenharia elétrica e computação, universidade presbiteriana mackenzie, brazil jefferson.a.br@ieee.org akamine@ieee.org paulo.lopes@mackenzie.br abstract — the modulation technique filter bank multi carrier (fbmc) is an alternative widely studied in order to replace orthogonal frequency division multiplexing (ofdm) in wireless telecommunications systems. the fact that fbmc does not use cyclic prefix and uses polyphase filters allows improvements in bit rate, bandwidth efficiency and robustness against multipath channel impairments. this implementation can bring advantages to digital tv case in comparison to the traditional ofdm based systems because of the need to transmit higher resolution videos such as 4k and 8k. this article presents a study on the use of fbmc in integrated system digital broadcasting transmission b (isdb-tb), developing an application on gnuradio environment, analysing bit error rate (ber) and power spectrum curves in a multipath channel. keywords: fbmc, ofdm, isdb-tb, gnuradio, polyphase filters. i. introduction the ofdm is one of the most used modulation techniques in telecommunication systems. however, the fbmc has become an alternative in order to improve the bit rate and bandwidth efficiency of wireless systems [1]. when ofdm is replaced by fbmc, the improvement in bandwidth utilization (bits/sec/hz) may be up to 25% [2] as there is no need for cyclic prefix. at the same time, there is a 20% increase in robustness to interferences resulting in better bit error curve [3]. the fbmc innovation is the fact that the modulation is accomplished using a filter bank that splits the spectrum in narrow bands. however, the channel equalization becomes hard and it is necessary to implement a little more complex equalization methods. for those reasons, this article presents a modified model of isdb-tb using fbmc, which was implemented on gnuradio software defined radio environment with different channel estimators. the ber curves and frequency domain are presented to evaluate the obtained improvements with different channel estimators. this article consists of seven sections. the first introduces the theme of the research. the second details the fbmc modulation method. the section three presents isdb-t characteristics. in the fourth, it is showed the pilot based estimation method. the number five describes a developed model of isdb-tb using fbmc on gnuradio. in the sixth, the comparison results between estimation methods are presented, obtained through the simulations and finally the relevant conclusions to the initial objectives. ii. filter bank multi carrier fbmc can be understood if figure 1 is considered. the data to be transmitted is split into m different paths in a filter bank arrangement. the resulting signal s(k) is expressed in (1) in which am,n is the symbol, gm the filter response shifted, and n the time position.      1 0 , ) 2 ()( m m n mnm nm kgaks (1) mathematically, each filter can be expressed by (2). 𝐵𝑘 𝑓 = 𝐻 𝑓 − 𝑘 𝑀 = ℎ𝑖𝑒 −𝑗 2𝜋𝑖 𝑓− 𝑘 𝑀 𝐿−1 𝑖=0 (2) where f is the shifted frequency, l is the number of filter coefficients, and m is the number of sub channels. applying the z transform, polyphase decomposition, and if 𝑊𝑀 = 𝑒 − 2𝑗𝜋 𝑀 , then we can find (3). fig. 1 fbmc modulator 𝐵𝑘 𝑧 = 𝑊𝑀 −𝑘𝑝 𝑧−𝑝𝐻𝑝 (𝑧 𝑀 )𝑀−1𝑝=0 (3) it is possible to show (3) in a matrix version, such can be seen in (4), that represents the same system showed in figure 1. 𝐵0 (𝑧) ⋮ 𝐵𝑀−1 (𝑧) = 1 ⋯ 1 ⋮ ⋱ ⋮ 1 ⋯ 𝑊𝑀 −(𝑀+1)2 𝐻0(𝑧 𝑀 ) ⋮ 𝑧−(𝑀−1)𝐻𝑀−1(𝑧 𝑀 ) (4) as non-orthogonal filters usually compose the filter bank, offset quadrature amplitude modulation (oqam) is used. hence, the symbols are transmitted in a staggered way in order to keep the orthogonality among adjacent carriers. the complete system, transmitter and receiver, can be implemented through a combination of a polyphase filter bank and fft blocks as explained in [4]. iii. integrated services digital broadcasting terrestrial version b isdb-tb is the terrestrial digital tv standard adopted by 15 countries in latin america and africa. isdb-tb can be transmitted on 6, 7, or 8 mhz channels with 13 segments that are multiplexed in ofdm blocks and 1 segment as guard band as can be seen in figure 2. i-etc: isel academic journal of electronics, telecommunications and computers cetc2016 issue, vol. 3, n. 1 (2017) id-10 http://journals.isel.pt mailto:jefferson.a.br@ieee.org mailto:akamine@ieee.org mailto:paulo.lopes@mackenzie.br fig. 2 segmentation of isdb-tb channel the 13 segments can be arranged for operation with one segment, standard definition, or high definition modes. the standard allows the use of qpsk, 16qam, and 64qam modulations in each particular ofdm layer. it uses 4992 data carriers, 625 pilot tones and an 8k ifft with zeros insertion to complete the frame. the cyclic prefix values can be 1/4, 1/8, 1/16, or 1/32 of the useful ofdm symbol time. we can use three different modes showed in table i [5]. table i. isdb-tb transmission parameters mode parameters carriers useful carriers pilots useful time ifft length 1 1405 1248 157 0.252ms 2048 2 2809 2496 313 0.504ms 4096 3 5617 4992 625 1.008ms 8192 iv. pilot based channel estimation in order to estimate the transfer function of the channel, pilot tones are sent in the isdb-tb. the scattered pilots have an amplitude of either +4/3 or -4/3. these values depend on a pseudorandom binary sequence (prbs) with polynomial , positioned in every 12th sub channel. the starting subcarrier is either the 0th, 3rd, 6th, or 9th according to the ofdm or fbmc symbol order [5]. after finding the transfer function of the pilots (hp) by (5), where y(k) and x(k) (always different from zero) are the amplitudes of received and transmitted pilot respectively, an interpolation method, such as linear or cubic, is applied with the purpose of estimating the other subcarrier responses. this operation can be done not only in the frequency domain but also in the time domain. )( )( )( kxp kyp khp  (5) in the linear case, it is used (6). 𝐻 𝑘 = 1 − 𝑎 𝐻𝑝 𝑘 + 𝑎𝐻𝑝 𝑘 + 1 (6) where a is a constant determined by the relation between the distance of the carrier until the pilot and the distance until the next pilot. on the other hand, applying the cubic interpolation, (7) is used. 𝐻 𝑘 = 𝐴 𝑎 𝐻𝑝 𝑘 + 𝐵 𝑎 𝐻𝑝 𝑘 + 1 + 𝐶 𝑎 𝑧 𝑘 + 𝐷 𝑎 𝑧 𝑘 + 1 (7) where a(a), b(a), c(a), and d(a) are constants related to a, and z(m) is the second order derivation obtained by the pilot information [6]. v. isdb-tb using fbmc implementation the complete isdb-tb proposed system employing fbmc is depicted in figure 3. as it can be seen, the source generates data which are modulated, added zeros, staggered (pre oqam) and multiplied by beta. then, an ifft and synthesis filters are applied. the resulting signal is sent through the channel. at the receiver, the signal passes through the analysis filters and fft. then, the opposite steps performed at the transmitter are accomplished. for this implementation, a system model was created on gnuradio utilizing the c++ programming language, in order to form a flow graph. this model was constructed in a way that both the traditional ofdm and fbmc-based isdb-tb can be simulated in real time. the configuration used was the full segment, mode 3, 64qam, 4992 useful subcarriers, and 1/16 cp in the ofdm system. four channel estimators were implemented: two of them only in the frequency domain with linear or cubic interpolation, one in the frequency and time domain with cubic method [7] and the last one specifically to fbmc in the frequency domain with cubic using 3-tap equalization [4]. vi. results in order to obtain these results it was used a 64-qam random source. then, by analyzing the spectrum representation in figure 4, it is possible to observe that the spectrum of fbmc has a decay of about 70db more than ofdm. fig. 4 fbmc and ofdm spectra figures 5, 6, and 7 present the ber curves found for the introduction of gaussian noise in a brazil a digital tv channel model [8], which encompasses 6 paths with 0, 0.15, 2.2, 3.05, 5.86, and 5.93 microseconds of delay and 0, 13.8, 16.2, 14.9, 13.6, and 16.4db of attenuation respectively, using different estimator for fbmc and ofdm systems. the fig.3 schematic of isdb-tb fbmc j.almeida et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-10 i-etc: isel academic journal of electronics, telecommunications and computers estimators used were chosen according to most common applied on ofdm and fbmc systems. from figure 5, 6 and 7, the fbmc based system presents a better robustness to gaussian noise in a multipath channel even in the absence of a cyclic prefix, using the same estimators. other important aspect is the fact that the equalization using 3-taps designed in order to avoid the adjacent channel interference in the case of fbmc allows a great improvement in the robustness to impairments, resulting in a bit rate increase of 6% for cp equal to 1/16. 0 5 10 15 20 25 10 -3 10 -2 10 -1 10 0 snr b e r fbmc-linear-estimator-frequency(f) ofdm-linear-estimator-frequency(f) fig. 5 ber curves using linear interpolation for estimation on frequency 0 5 10 15 20 25 10 -5 10 -4 10 -3 10 -2 10 -1 10 0 snr b e r fbmc-cubic-estimator-frequency(f) ofdm-cubic-estimator-frequency(f) fig. 6 ber curves using cubic interpolation for estimation on frequency 0 5 10 15 20 25 10 -5 10 -4 10 -3 10 -2 10 -1 10 0 snr b e r fbmc-cubic-estimator-frequency(f)-time(t) ofdm-cubic-estimator-frequency(f)-time(t) fbmc-cubic-estimator-frequency(f)-3tap fig. 7 ber curves using cubic different estimators the table ii contains a summary with the obtained results. table ii. summary of ber results sistema linear estimator (f) (10-2) cubic estimator (f) (10-3) cubic estimator (f/t) (10-3) 3-tap estimator (10-3) fbmc 22db 23db 22db 21db ofdm 24db 23,5db 22,5db vii. conclusion this study has proved that the implementation of isdb-tb digital tv system employing fbmc instead of ofdm is viable and brings some advantages. among them the increase of bit rate and bandwidth efficience in terms of bits/sec/hz around to 25% if the cp is 1/4. in addition, due to the decay of spectrum, reducing the interferences among adjacent sub channels, there is an improvement in terms of ber in multipath channels. when the estimators are analyzed, it is possible to verify that 3 tap equalization it is necessary to cancel the interference among adjacent subchannels, which increases the robustness to interferences with a little more computational complexity. references [1] bellanger, m. et al. a filter bank multicarrier scheme running at symbol rate for future wireless systems. in: wireless telecommunications symposium (wts), 2015. [s.l.: s.n.], 2015. p. 15. [2] bellanger, m. et al. ofdm and fbmc transmission techniques: a compatible high performance proposal for broadband power line communications. in: power line communications and its applications (isplc), 2010 ieee international symposium on. [s.l.: s.n.], 2010. p. 154-159. [3] farhang-boroujeny, b. ofdm versus fillter bank multicarrier. signal processing magazine, ieee, v. 28, n. 3, p. 92-112, may 2011. issn 1053-5888. [4] bellanger, m. fbmc physical layer: a primer. phydyas, january, p. 1-31, 2010. disponível em: <http://www.ictphydyas.org/teamspace/internal-folder/fbmcprimer\_06-2010.pdf>. [5] associação brasileira de normas técnicas. nbr 15601: televisão digital terrestre sistema de transmissão. rio de janeiro, 2008. [6] kang, s. g.; ha, y. m.; joo, e. k. a comparative investigation on channel estimation algorithms for ofdm in mobile communications. ieee transactions on broadcasting, v. 49, n. 2, p. 142-149, june 2003. issn 0018-9316. [7] ishini, a. k., akamine, c.; 2009. técnicas de estimação de canal para o sistema isdb-tb. revista de radiodifusão. issn print: 19814984. issn online: 2236-9619. v.3. doi: 10.18580/radiodifusao.2009.3.46. [8] itu radiocommunication study groups: document 6e/temp/131-e, guidelines and techniques for the evaluation of dttb systems, 19 march 2003. j.almeida et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-10 i-etc: isel academic journal of electronics, telecommunications and computers floating sensor platform for the monitoring of water quality in urban and white-water environments floating sensor platform for the monitoring of water quality in urban and white-water environments h.c. neitzerta, m. baltesb, r. schmittb, p. lovisic, g. landia, a. cuomod, d. guidad adepartment of industrial engineering, salerno university, italy bdepartment of electrical engineering, trier university of applied sciences, germany cconsortium cugri, salerno university, italy ddepartment of civil engineering, salerno university, italy neitzert@unisa.it michael.baltes@gmx.de romanschmitt90@gmx.de glandi@unisa.it acuomo@unisa.it dguida@unisa.it abstract — in the present paper the project of an embedded solution for the realization of a floating sensor platform for the monitoring of the water and ambient quality in a flowing water environment is described. first results regarding the monitoring of the water conductivity and the ambient noise level under harsh environmental conditions in a karstic river and in the final part of a river going towards the mediterranean sea are presented. it is further discussed how this kind of system can be modified in order to serve as urban waterway multisensory platform, adding important features like connectivity, energy harvesting and determination of the platform position. keywords: water quality, ambient noise, remote sensing, water conductivity, embedded sensor, smart city. i. introduction one of the most important constituents of smart city networks are autonomous sensing solutions. a great variety of nowadays developed sensors are dealing with road traffic and air pollution control. however, another important city ambient, that needs to be monitored, are the urban and suburban water ways and the recreation areas with lakes and rivers. water quality is one of the most important parameters, that influences the wellness of the urban population. some first solutions of continuous pollution monitoring based on the use of buoys [1] and also the application of a smart phone for water monitoring [2] have been reported. the realization of an autonomous multisensory buoy with a wide range of sensor functions, including salinity, sea water temperature and turbidity has been reported in literature, however in this case the sensor platform had no space restrictions [3]. this was also the case for an aerial wireless coastal buoy [4]. here we will show first results of a small, simple and mechanical very stable sensor platform for the monitoring of important water and ambient properties during floating operation on a river and then discuss the needed future developments in order to achieve autonomous operation and connectivity of the sensor platform. ii. sensor platform realization a. embedded sensor platform the described sensor platform has been originally developed for a non-urban application. the initially planned and realized application was the monitoring of a small whitewater river in karstic environment with long underground passages without possibility of wireless contact to the sensor platform. one of the main focuses during the development was therefore a very stable mechanical design, which can even survive passages in wild water and over high water-falls. i-etc: isel academic journal of electronics, telecommunications and computers iot-2018 issue, vol. 4, n. 1 (2018) id-3 http://journals.isel.pt fig. 1. block-diagram of the electronic circuits of the a) actually developed and b) of the future design of the floating water sensor platform the electronics of this first version of the embedded sensing platform is schematically shown in fig. 1a. the core element is an arduino1 embedded system with multiple analog inputs and a connected sd card unit for long-term data storage. arduino based sensor solutions are popular due to the easy to learn programming language. for example, an indoor environmental quality sensing system has been developed, that was based on an arduino embedded system with voltron software for data handling and analysis [5]. in this first sensor platform version three sensing elements have been used. the measured entities are water temperature, ambient acoustic sound level and water conductivity. the power supply by a simple 9v block battery was sufficient for some hours of data taking. regarding the first two sensors, commercial sensor solutions have been used. the water temperature has been measured using a mcp 9701 bandgap reference type electrical temperature sensor glued into the lower part of the plastic case of the sensor and for the ambient acoustic noise measurements, a loudness sensor type “seed studio grove sound sensor” has been used. this type of sensor had been added in order to distinguish calm water and turbulent water – including waterfalls and cataracts – during the floating of the sensor platform in the investigated mountain river. in an urban environment the determination of the noise level in recreational area waters could also be of potential interest. the last type of sensor, a water conductivity sensor, however, has been our own development and will be described more in detail. fig. 2. main parts of the realized floating water sensor platform, as are: 1. arduino uno with sensor shield, sd card unit and attached loudness sensor 2. 9v block battery 3. bottom part of the sensor shielding with conductivity sensing electrodes, temperature sensor and protective plastic shielding 4. metallic ring as weight for ensuring floating operation 5. upper closure of the sensor shielding an exploit of the sensor components is shown in fig. 2, the main electronic board, with the arduino1, the sd card unit and the sensor connection shield is put within a plastic case with the temperature and the water conductivity sensor integrated into the bottom part of the plastic case. both sensors are protected by an additional plastic ring in order to withstand even hard falling on rocks within a typical whitewater river environment. the efficient floating in the river and the maintaining of an upright position within the water has been obtained by adding a cylindrical metallic ring into the plastic case with a specific weight. after connecting the 9v block battery to the electronics, the case is closed by screwing the top closure tight and the measurements start. because in the case of the white water river with underearth passages no radio-frequency connection has been possible, this first version of the sensor platform has been realized just with a local data storage on a sd card. in this case for the data recovery, the platform has to be physically recovered and opened after the measurement campaign. in fig.1b the block diagram of a possible future evolution of the floating type sensor platform is indicated. in addition to the various ambient sensors (e.g. ph-sensor, water co2 concentration sensor, optical turbidity sensor) a gps system allows for the determination of the sensor position during data taking and the gps-unit allows to send the data, which is eventually also locally stored on the platform, in real time for direct feedback. this future version of the platform could also be made completely autonomous by adding an energy harvesting option with electrical energy storage in a rechargeable battery or a supercapacitor. for this option, however, it would be needed to utilize an embedded system with far less power consumption than the arduino type. b. conductivity sensor water conductivity is an important parameter, which has to be monitored (see for example [6].) regarding the conductivity sensor, developed by two of the authors, a more detailed description regarding the geometry, measurement principle and calibration procedure will be given. a variety of techniques have been proposed for water conductivity measurements. for example, a capacitive technique has been reported [7] also optical spectroscopy techniques are often proposed [8]. due to the geometrical constraints of the small sensor platform dimensions these measurement techniques could not been used in our case and we decided to use conventional metallic contact based measurements. a drawback of this techniques is of course the poor long-term stability due to electrochemical modifications of the used electrodes. in order to minimize these problems, a pulsed voltage signal, generated with the arduino (shown in fig. 3b) with a pulse length of only 12 ms and a repetition rate of 1 hz has been applied to the electrodes, which together with a series calibration resistor forms a voltage divider. the voltage, measured over the measurement electrodes, is then fed into one of the analog input channels of the arduino with a maximum input voltage amplitude of 5 v. the sensor has been calibrated using a commercial type “hanna hi 99300 ec/tds” conductivity meter as reference instrument and adding successively small amounts h.c. nietzert et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt of salt to the distilled water (the measurement setup is shown in fig. 3c). the sensor calibration curve is shown in fig.3a and it can be seen, that with the chosen resistance value for the series resistor, a large range of conductivity values can be monitored with a good sensitivity in the expected typical conductivity range of river water. fig. 3. conductivity sensing configuration with a) calibration curve (inset: photo of the sensor electrodes), b) the electrical measurement pulse and c) a photo of the experimental setup for the calibration procedure iii. experimental monitoring results a. water conductivity monitoring during the floating on the river irno even if the sensing platform originally has been developed for the characterization of a white-water river in a karstic environment in the cilento region south of salerno, we started testing it in the urban environment of salerno. in particular, we let it float in the last 300 m of the irno river, ending in the salty tyrrhenian sea. some months before this monitoring test, salerno university students under guidance of prof. d. guida measured the water conductivity at different points on the beach, very close to the final part of the irno river with a commercial water conductivity meter. in fig. 4a in the map of the fraction of salerno, where the irno is meeting the tyrrhenian sea is shown. fig. 4.a) map of the investigated region with the irno river (indicated by the blue circle) going to the tyrrhenian sea near salerno harbor and b) sketch of the irno near seaside area with indicated measurement points on the beach. table i water conductivity values at the indicated positions in fig.4 point conductivity (s/cm) 0 665 1 669 2 670 3 674 4 677 5 700 6 715 7 725 8 715 9 695 10 715 11 42000 h.c. nietzert et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt 12 42000 13 695 fig. 5. conductivity sensor output voltage monitoring during the floating of the sensor platform for the last 300m towards the open sea. in fig.4b a schematic drawing of the river and the nearby shoreline is given, where the numbers are indicating the measurement positions, where the students measured the electrical conductivity values, listed in table 1. we see that rather constant values are measured up to point 10 and only the following 2 measurements show conductivity values typical for the salt concentration as present in the tyrrhenian sea. the monitoring results of the conductivity sensor output voltage during the floating of the sensor platform for the last some meters before entering the open sea are demonstrated in fig.5. it should be reminded, that high output voltages correspond to low electrical conductivity values (see the above shown calibration). it should be mentioned that only some few erroneous data points have been removed for clarity, but that no filtering procedure has been applied to the data. it can be observed that in the beginning a very smooth trace is observed with a conductivity sensor output voltage value of about 2.33 v. the corresponding electrical conductivity value, obtained by using the calibration curve, shown in fig.3, agrees very well with the values measured by the students before (see table 1) in the river near measurement positions. successively we see a slightly noisier signal with varying conductivity due to the arrival of sea water waves in the river ending and the beginning of mixing of river and sea water. and finally the sensor output voltage drops to much lower values of about 0.65. this is out of the range of the prior done calibration of the conductivity sensor and the salinity most probably reaches values near the sensor close to the open sea values. the sensor platform had been afterwards with the help of an attached cord again teared back inside the irno river. the sensor signal increases again exactly to the value as in the beginning of the shown monitoring trace. this is an indication for the good reproducibility of the sensor data and for the absence of short term sensor characteristics drift. g. monitoring of the ambient noise level during the sensor floating in a white-water mountain river as another example of successful data taking during sensor platform floating, in fig.6 the monitoring of the loudness sensor output voltage, which is proportional to the measured sound amplitude, is shown. fig. 6. loudness sensor output voltage monitoring during the floating of the sensor platform on the bussento river in the cileto mountains. in this case the sensor has been tested under extreme conditions in the bussento river in the cilento region south of salerno. the monitored part of the river was characterized by cataracts and water falls, which presence is clearly evidenced in the monitoring trace seen in fig.6. iii. conclusions a compact and robust floating measurement platform for the monitoring of water and ambient parameters has been realized using an arduino embedded system for the data taking and storage. a water electrical conductivity sensor has been developed, calibrated and tested successfully for the characterization of a river estuary, showing clearly the mixing process of sweet and salt water when the measurement platform was entering the mediterranean sea. the results have been compared to measurements done with a commercial water conductivity meter in a separate measurement campaign. references [1] g. griffo, l. piper, a. lay-ekuakille and d. pellicano. design of buoy station for marine pollutant detection, measurement, vol. 47, pp. 1024-1029, 2014. h.c. nietzert et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt [2] s. dutta, d. sarma and p. nath. ground and river water quality monitoring using a smartphone-based ph sensor, aip advances, vol. 5, 057151, 2015. [3] s. sendra, l. parra, j. lloret and j.m. jiménez. oceanographic multisensor buoy based on low cost sensors for posidonia meadows monitoring in mediterranean sea, journal of sensors, vol. 2015, article id 920168 (23 pages), 2015. [4] l.o. wahidin, i. jaya and a.s. atmadipoera, design, construction, and stability test od aerial wireless coastal buoy, iop conf. series: earth and environmental. science, vol 176, article id 012041 (10 pages), 2018. [5] m. karami, g. v. mcmorrow and l. wang. continuous monitoring of indoor environmental quality using an arduino-based data acquisition system, j. building eng., vol. 19, pp. 412-419, 2018. [6] b.d. thomas, t.g. thompson and c.l. utterback. the electrical conductivity of sea water, iecs journal of marine science, vol. 9, pp.28-34, 1934. [7] s. bhat. salinity (conductivity) sensor plate capacitors, ph.d thesis, university of south florida, 2005. [8] o. esteban, m. cruz-navarrete, a. gonzález-cano and e. bernabeu. measurement of the degree of salinity of water with a fiber-optic sensor, applied optics, vol. 38, no. 25, pp. 5267-5271, 1999. h.c. nietzert et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt semiconductor device for detection of fret signals semiconductor device for detection of fret signals p. louro 1,2, m. vieira1,2,3, m. a. vieira1,2, a. karmali4, m. fernandes1,2 1 isel-adeetc, lisboa, portugal 2 cts-uninova, caparica, 2829-516, caparica, portugal 3 dee-fct-unl, quinta da torre, monte da caparica, 2829-516, caparica, portugal 4 cieb-isel, lisbon, portugal keywords: optoelectronics, fret, visible spectral range, a-sic:h. abstract: the transducer consists on a semiconductor device based on two stacked-i-n heterostructures, that were designed to detect the emissions of the fluorescence resonance energy transfer between fluorophores in the cyan (470 nm) and yellow (588 nm) range of the spectrum. this research represents a preliminary study on the use of such wavelength sensitive devices as photodetectors for this kind of application. the device was characterized through optoelectronic measurements concerning spectral response measurements under different electrical and optical biasing conditions. to simulate the fret pairs a chromatic time dependent combination of cyan and yellow wavelengths was applied to the device. the generated photocurrent was measured under reverse and forward bias to readout the output photocurrent signal. different wavelength biasing light was also superimposed. results show that under reverse bias the photocurrent signal presents four separate levels each one assigned to the different wavelength combinations of the fret pairs. if a blue background is superimposed the yellow channel is enhanced and the cyan suppressed while under red irradiation the opposite behavior occurs. so under suitable biasing light the transducer is able to detect separately the cyan and yellow fluorescence pairs. an electrical model, supported by a numerical simulation supports the transduction mechanism of the device. 1 introduction fluorescence resonance energy transfer (fret) is a process involving the radiationless transfer of energy from a "donor" fluorophore to an appropriately positioned "acceptor" fluorophore. the distance over which fret can occur is limited to between 1-10 nm, and hence this technique is used to demonstrate whether two types of molecules, labelled with a donor-fluorophore and a receptor fluorophore, occur within 10 nm of each other. the determination of the distance between two molecular species is important in the field of medical and biological applications. this technique has grown in popularity due to the emergence of various fluorescent mutant proteins with shifted spectral properties [1, 2]. in this paper we use wavelength sensitive devices operating in the visible range to detect the cyan and yellow optical signals involved in fret. the advantage of this type of transducer relies on the intrinsic optical filtering properties of the transducer associated with the photodetection ability. thus, the use of this type of integrated devices discards the need of optical/mechanical filters to separate the wavelength of the emitted spectra and the used of an additional photodetector for each optical signal, in this paper we have used a purified preparation of glucose oxidase (ec 1.1.3.4) from aspergillus niger [3]. the fluorescent signals obtained by excitation at 450 nm were measured with a spectrofluorimeter. changes in the peak values of fluorescence in the cyan and yellow region can be used to detect the presence of glucose. the sensor was studied also having this application in consideration. 2 experimental details the element sensor is a glass/ito/a-sic:h (p-i-n)/ a-sic:h(-p) /si:h(-i)/sic:h (-n)/ito double heterostructure produced by pecvd. deposition i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-2 http://journals.isel.pt/index.php/iajetc conditions are described elsewhere [4]. the thickness (200 nm) and the absorption coefficient of the front photodiode (based on a-sic:h) are optimized for short wavelength collection and long wavelength transmittance, and the thickness of the back one (1000 nm) adjusted to achieve full absorption in the intermediate wavelength region and high collection in the long wavelength spectral section. as a result, both front and back diodes act as optical filters confining, respectively, the short and the long wavelength optical carriers, while the intermediate ones are absorbed across both [5]. in fig.1 the device configuration is depicted figure 1: device configuration. dual monochromatic light beams of 470 nm and 588 nm are directed to the device to simulate, respectively, the emitted cyan and yellow fluorescent signals of glucose oxidase from aspergillus niger [1]. 3 results and discussion 3.1. glucose oxidase fluorescence spectrum in fig. 2 it is displayed the fluorescence emission spectrum of from aspergillus niger prepared with purified enzyme (specific activity of 125 u/mg protein) dissolved in 20 mm phosphate buffer ph 7.0 (5 mg/l). a wavelength of 450 nm was used as excitation source. the emission spectrum shows fluorescence in the range 480-600 nm with a peak located at 520 nm. the deconvolution of the experimental data gaussian fit originates two peaks located at 518 nm and 562 nm, which corresponds respectively to the cyan and yellow emissions. 480 500 520 540 560 580 600 0,0 0,1 0,2 0,3 0,4 glucose oxidase experimental data gaussian fit gauss fit peak at 518 nm gauss fit peak at 562 nm f lu or es ce nc e in te ns ity ( a. u. ) wavelength (nm) figure 2: fluorescence emission of glucose oxidase. 3.2. device spectral sensitivity figure 3 displays the spectral photocurrent, measured along the visible spectrum, under reverse and forward bias without and with optical light bias of different wavelengths. 400 450 500 550 600 650 700 0 2x10-3 4x10-3 6x10-3 8x10-3 1x10-2 λ yellowλcyan p ho to cu rr en t ( μ a ) wavelength (nm) bias -8 v +1 v off 624 nm 588 nm figure 3: device spectral photocurrent under reverse and forward bias without and with red and yellow optical bias. results show that without optical bias (dark lines) the use of reverse bias influences the photocurrent of the cyan wavelength emission, while for the yellow emission it is independent on the electrical voltage applied to the device. on the other hand, under optical bias, either red (624 nm) or yellow (588 nm), the photocurrent at – 8 v exhibits the opposite behaviour. its magnitude remains constant at 470 nm and at 588 nm it decreases. thus, the device sensitivity to the wavelengths of interest, i.e., 470 nm and 588 nm, can be tuned either by optical and electrical bias. p. louro et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 3.3. device output photocurrent signal the photocurrent signal obtained at reverse and forward electric bias with and without red optical bias (624 nm) is displayed in fig. 4 for both emission optical signals (470 nm and 588). 0,0 0,5 1,0 1,5 2,0 2,5 0,0 2,0x10-7 4,0x10-7 6,0x10-7 8,0x10-7 1,0x10-6 +1v -8 v 470 nm cyan emission p ho to cu rr en t ( a ) time (ms) a) 0,0 0,5 1,0 1,5 2,0 2,5 0,0 5,0x10-7 1,0x10-6 1,5x10-6 -8 v +1 v p ho to cu rr em t ( a ) time (ms) 588 nm yellow emission b) figure 4: output photocurrent signal obtained at reverse (-8v, solid lines) and forward (+1 v, dash lines) bias without (dark lines) and with under red background light (red lines) with the: a) cyan (470 nm) and b) yellow (588 nm) emission signals. as already seen in fig. 3, the cyan emission signal (fig. 4 a) shows a strong dependence on the applied bias of the device. under red optical bias the signal is amplified with a magnitude factor around 2. the yellow emission signal (fig. 4b) shows a weaker dependence on the applied voltage and a strong attenuation under red optical optical bias. a chromatic time dependent wavelength combination (3000hz) of both emission signals (470 nm and 588 nm), was used to simulate the fret emission spectrum in the device. the output photocurrent, with (color lines) and without (dark lines) optical background light are displayed in figure 5 a) and b) under forward (dash arrows) and reverse (solid arrows) voltages, respectively. the reference level was assumed to be the signal when all the input optical signals channels off. at the top of the figure, the individual optical signals are displayed to guide the eyes in relation to the different on-off states. 0,00 0,25 0,50 0,0 5,0x10-7 1,0x10-6 1,5x10-6 624 nm φ l =0 emission signals 470 nm 588 nm +1 v p ho to cu rr en t ( a ) time (ms) a) 0,00 0,25 0,50 0,0 1,0x10-6 2,0x10-6 3,0x10-6 4,0x10-6 -8 v emission signals 624 nm φ l =0 470 nm 588 nm p ho to cu rr en t ( a ) time (ms) b) figure 5: output photocurrent signals without (фl=0) and with red optical bias (624 nm) at: a) forward bias (+1 v) and b) reverse bias (-8 v). the optical signals waveforms are shown at the top of the figure. under forward bias (fig. 5a) the photocurrent signals with or without optical bias are similar and their waveform follows the yellow fluorescence signal, due to the lower sensitivity of the device to the cyan light (figures 3 and 4). this feature allows immediate decoding of the yellow fluorescence signal. under reverse bias (fig. 5b) the photocurrent signal is more complex. each waveform is 4-level encoding (22) due to the different combinations of the input optical signals. p. louro et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc once the yellow signal is decoded by the use of forward bias, the other fluorescence signal can be obtained from the photocurrent signal at -8v taking into account the magnitude dependence on the applied bias (figure 4). to recover the cyan and yellow emission intensities, red optical bias was used. under red irradiation the yellow channel is quenched and the cyan signal is amplified. so the highest two levels in the photocurrent signal under 624 nm (figure 6b) correspond to the presence of the cyan light and the lowest two levels to its absence. the yellow channel is decoded under forward bias (fig. 6a). by using this simple algorithm the emission spectra can be recovered, in real time, and its ratio can be correlated with the distance between the fluorophores. 3.4. electrical model based on the experimental results and device configuration an optoelectronic model was developed [6]. the device was modeled by a two single-tuned stages circuit with two variable capacitors and interconnected phototransistors through a resistor (fig. 6). two optical gate connections ascribed to the different light penetration depths across the front and back phototransistors were considered to allow independent yellow and cyan optical signals transmission. figure 6: simplified ac equivalent electrical circuit of the device. the operation is based upon the following principle: the flow of current through the resistor connecting the two front and back transistor bases is proportional to the difference in the voltages across both capacitors (charge storage buckets). so, it uses a changing capacitance to control the power delivered to the load acting as a state variable filter circuit. in figure 7 the simulated current without and under red backgrounds is displayed (symbols) using the same test signal of fig. 5. the input channels are also displayed (lines). to simulate the red background, current sources intensities (input channels) were multiplied by the on/off ratio between the input channels with and without red optical bias. good agreement between experimental and simulated data was observed. 0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,0 2,0 i(off)r 1 =1kω r 2 =5kω i(on) α r =0.4 α b =1.75 φ=0 λ=624 nm p ho to cu rr en t ( μ a ) time (ms) figure 7: simulated photocurrent signal (symbols), input channels (dash lines) and experimental signal (solid lines): under reverse bias (-8 v) and red optical bias (624 nm). the four expected levels, under reversed bias, and their reduction under red irradiation, are clearly seen. under red background the expected optical amplification of the cyan channel and the quenching of the yellow one were observed due to the effect of the active multiple-feedback filter when the back diode is light triggered. 5 conclusion we present a new device semiconductor based in an a-sic:h p-i-n/p-i-n heterostructures for the detection of optical signals near the cyan and yellow regions of the visible spectrum, which can be used for the detection of the emission signals used in the fret technique. two different modulated optical signals were used to simulate the emission signals of the fluorophores used in fret. the output emission spectrum was analyzed by reading out the photocurrent generated by the device. the transducer mechanism was i4 q2 i1 i2 q1 c2 r2 r1 (v) c1 i3 p. louro et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc explained by an electrical model supported by a numerical simulation. future work comprises the use of lower power intensities for the simulated emission signals in order to reach the same range of the ones produced during the fret phenomenon. in a further stage tests with emission fluorescence signals from different samples of glucose oxidase must also be done. acknowledge this work was supported by ptdc/eeaelc/111854/2009. references [1] d.a lavan, terry mcguire, robert langer, nat. biotechnol. 21 (10), 2003, 1184–1191. [2] d. grace, medical product manufacturing news, 12, 2008, 22–23. [3] k. karmali, a. karmali, a. teixeira, m. j. marcelo curto “assay of glucose oxidase from penicillium amagasakiense and aspergillus niger by fourier transform infrared spectroscopy”. (2004) analytical biochemistry 333, 320-327. [4] m. vieira, a. fantoni, m. fernandes, p. louro, g. lavareda and c.n. carvalho, thin solid films, 515, issue 19, 2007, 7566-7570. [5] p. louro, m. vieira, yu. vygranenko, a. fantoni, m. fernandes, g. lavareda, n. carvalho mat. res. soc. symp. proc., 989 (2007) a12.04. [6] m. a. vieira, m. vieira, j. costa, p. louro, m. fernandes, a. fantoni, in sensors & transducers journal vol. 9, special issue, december 2010, pp.96120. p. louro et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc inverted-l antenna (ila) design using fractal for wlan usb dongle inverted-l antenna (ila) design using fractal for wlan usb dongle qi luo school of engineering and digital arts, university of kent, uk qiluo@ieee.org j.r.pereira instituto de telecomunicações/universidade de aveiro jrp@ua.pt h.m.salgado inesc porto, faculdade de engenharia, universidade do porto, porto, portugal hsalgado@fe.up.pt keywords: inverted-l antenna, fractal, usb dongle, wlan. abstract: this work presents an inverted-l antenna design using the fractal geometry for dual band wlan (2.4/5.2ghz) usb dongle application. the proposed antenna has the advantages of compact size, wide operation bandwidth and easy fabrication. the experimental results show that it has a s11<-10 db bandwidth from 2.25 to 2.60 ghz and 5.06 to 5.62 ghz. the radiation performances of the proposed antenna in free space and when connected to a laptop computer were also studied in this work. the proposed antenna was designed and optimized by using ansoft hfss v13. 1 introduction there are several existing reported works concerning about the design of wlan antennas for usb dongle applications. in [1], two internal multiband pifa antennas were proposed for umts and wlan applications for a usb dongle. although both of them have compact size and can operate at multiple frequency bands, they are rather complicated to fabricate and the use of a short pin means that the size of the ground plane will play a significant role in determining the resonant frequencies. in [2], one printed monopole antenna was designed for wlan usb dongle. this antenna employed the meander-line to reduce the occupied volume of the radiation element. however, this antenna can only operate at 2.4 ghz band, which is not enough for nowadays dual band wlan applications. in [3], one two armed printed monopole for wlan has been proposed. the use of multiple arms can provide an alternative solution in designing a multiband printed monopole. yet, its size seems inappropriate for a usb dongle. in the previous study [4], we have utilized the fractal to design a compact printed monopole antenna for dual-band wlan usb dongle. in that design, the antenna has a planar structure and the feeding port is located at the end of the substrate. the radiation element with the feeding line and the ground plane are respectively printed at the top and bottom side of i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-14 http://journals.isel.pt/index.php/iajetc the substrate, which constitutes a typical approach for a printed monopole antenna design. this might be not convenient for most of the industry designs as other components, such as rf module, also need to be mounted on the same ground plane. the aim of this work is to use the same fractal-based structure reported previously to design an inverted-l antenna (ila), which has been widely used in the design of antennas for portable devices, allowing for the antenna element and ground plane to be printed on the same side of the substrate. this facilitates the integration of the antenna into industrial products. however, one inherent drawback for ila is that it always has low input impedance [5]. the typical method that can be employed to solve this problem for an ila is to short the antenna element to the ground, which can increase the input impedance of the antenna. then the antenna becomes an invertedf antenna (ifa), whose input impedance is easier to be matched. as will be presented later, the proposed fractal ila can be easily matched without shorting the antenna to the ground plane and the experimental results show that this antenna exhibits broad operational bandwidth (s11<-10 db) from 2.25 to 2.60 ghz and 5.06 to 5.62 ghz. moreover, the radiation performances of the proposed antenna when connected to a laptop computer were also investigated by doing numerical simulations in hfss. 2 antenna structure since the objective of this study is to design a printed fractal monopole antenna for wlan usb dongle applications, based on the industrial requirement, the overall size of this antenna including the ground plane is chosen to be 20 mm × 60 mm and the available space for antenna design is limited to 20 mm × 10 mm. fig. 1 shows the proposed ila using the fractal geometry. this antenna is designed on roger 4003 with thickness of 0.8 mm. in this design, as can be seen in the fig. 1, the antenna element and the ground plane are printed on the same side of the substrate. the bottom of the substrate has no copper. additionally, the feeding point lies on the left edge of the ground plane and this leaves enough space to mount other hardware components on the system ground plane. figure 1: proposed ila based on the fractal geometry a variation of the koch fractal was used to the design of the proposed multiband printed monopole antenna. fig.2 shows the first three iterations of the koch fractal geometry. in this work, the second iteration of the fractal is employed in the designing of the antenna because for higher iterations, in order to describe it properly the microstrip line needs to be very narrow, which can in turn decrease the radiation efficiency of the antenna due to the conductor losses. the depth d shown in fig.2 is a parameter that is used to construct the koch fractal. this value can be varied as required, and in this work the value of d is set to be 1/5 of the line length at each iteration. figure 2: the first three iterations of koch fractal 3 simulation and experimental results 3.1 parametrical studies the antenna component needs to be taken as an integrated part of the entire layout of the transceiver. it is predictable that after assembling the overall product and connecting the usb dongle to a device, for instance a laptop computer, the radiation q. luo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-14 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc performance of the antenna will be affected. therefore, parametrical studies of the proposed antenna are necessary, which can aid the tuning of the resonant frequency of the antenna in the final stage of product design. two key parameters, l0 which is related to the overall length of the fractal and l1, which is length of the horizontal meander line, are chosen as the variables. these two parameters are labeled in fig. 3 (a). fig. 3 (b) and fig.3 (c) present the simulated return loss of the proposed fractal ila with different values of l0 and l1, respectively. it is found that the overall length of the fractal determines the resonant frequencies at both bands while the length of l1 has major influences on the higher band. by utilizing these findings, the resonant frequencies of the antenna can be tuned to the desired resonant frequency according to the requirements. in the final prototype, the values of these parameters were chosen to be l0=14mm, l1=8.1mm and l2=1.6mm. figure 3: parametrical studies for the proposed antenna: (a) the layout of the antenna element; (b) simulated return loss of the antenna with different value of l0; (c) simulated return loss of the antenna with different value of l1 3.2 measurement results the proposed antenna was fabricated and fig. 4 shows a photo of this prototype under return loss measurement. fig. 5 shows the comparison of the return loss between simulation and measurement results. there is a good agreement between the simulated and measured return loss, both of which confirm that this antenna has a s11<-10 db bandwidth from 2.25 to 2.60 ghz and 5.06 to 5.62 ghz. figure 4: photo of the fabricated prototype during return loss measurement (a) q. luo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-14 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 5: comparison of the simulated and measured return loss of the fractal ila fig. 6 shows the measured radiation patterns of this fractal ila at 2.4ghz at both e-and h-plane. figure 6: measured radiation patterns of the printed multiband ila antenna at 2.4ghz 3.3 frequency reduction of the proposed antenna in order to show the effectiveness of antenna size reduction by using the fractal geometry, one more inverted-f antenna (ifa) was designed. this antenna uses the same substrate and the radiation element was also confined within the required area of size 20 mm × 10 mm, as shown in fig.7. figure 7: the structure of a typical inverted-f antenna fig.8 compares the simulated return loss of this ifa with the proposed fractal ila. as can be observed from fig.7, the simple ifa can only operate at 3.85 ghz. compared to the ifa, besides exhibiting a dual band operation, the proposed ila antenna can resonate at 2.3 ghz, which represents a frequency decrease of 40%. figure 8: comparison of the simulated return loss between the proposed fractal ila and a typical inverted-f antenna e plane h plane q. luo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-14 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 3.4 simulated radiation patterns fig. 9 presents the simulated 3d radiation pattern of the proposed antenna in ansoft hfss. it can be seen that at both bands, the h plane (phi=0 degree) is almost omnidirectional while the e plane (phi=90 degree) exhibits broad side radiation patterns. figure 9: the simulated radiation patterns of the proposed antenna in free space (phi=0 degree: dashed line; phi=90 degree: solid line) the simulation results shows that the proposed antenna has maximum gain of 1.9 dbi at the lower band and 3.3 dbi at higher band with radiation efficiency higher than 95% at both bands. 4 antenna performance with the existence of laptop computer the influence of the laptop on the radiation performance of the antenna was studied by simulating the usb dongle attached to a laptop computer that is modeled as finite conductive material for simple consideration. when the usb dongle is connected to the laptop, it is equivalent to extend the size of the ground plan of the antenna to a much larger one. as the simulation software package used in this study, ansoft hfss, is a based on the finite element method (fem), in which the calculation is proportional to the size of the overall domain, therefore, to save simulation time and computation memory, the size of the laptop is truncated to half of the size of the real laptop computer. the simulation model is presented in fig.10. the laptop is modeled by using two copper plates vertically joined together to mimic the case when the laptop is opened. each of the copper plate is 20 cm long and 10 cm wide, which is approximately half of the size of a netbook pc. figure 10: simulation model of the usb dongle connecting to the laptop computer in hfss 4.1 return loss fig.8 shows the comparison of the simulated return loss of the proposed antenna with and without connection to the laptop computer. as expected, when connecting to the laptop, the proposed antenna has some frequency shift at both bands. at the lower band, the amplitude of the return loss has degraded by more than 10 db. after resizing the fractal, both resonant frequencies are tuned to the desired frequency bands, as shown in fig.11 by black solid line. after tuning, the simulation results show that the proposed antenna (connected with the laptop) exhibits a vswr 3:1 bandwidth, which is the standard accepted by most portable devices manufactures, over 2.36 to 2.54 ghz and 4.96 to 5.84 ghz. this covers the entire frequency bands for ieee 802.11 a/b/g applications. f=5.3ghz f=2.3ghz q. luo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-14 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 11: simulated return loss of the proposed antenna with and without the laptop fig.12 presents the comparison of the simulated return loss of the proposed antenna with the laptop computer of different sizes. as can be seen from this figure, the size of the laptop computer affects in some degree the resonant frequencies of the antenna. however, above 15 cm the resonant frequencies of the antenna are no longer influenced by the length particularly at the lower band. this proves the validity of the simple simulation model of the laptop computer. figure 12: simulated return loss of the proposed antenna with laptop computer of different length 4.2 radiation pattern fig.13 and fig.14 presents the comparison between the simulated radiation patterns of the proposed fractal ila in free space and connected with the laptop at 2 and 5 ghz band, respectively. it can be observed that there is a large influence from the laptop to the radiation patterns of the antenna. with the existence of the laptop computer, the radiation patterns of the antenna become more directive in certain angle. however, according to the simulation results, the radiation efficiency of the antenna at the lower band decreased to 60% while at the higher band, the radiation efficiency reduced to 50%. compared to the case that it is radiating in the free space, the radiation efficiency at both bands has dropped by more than 40%. figure 13: comparison of the simulated radiation patterns of the proposed antenna with (blue dashed line) and without (black solid line) connecting to the laptop computer at 2 ghz band phi=0degree phi=90 degree phi=0 degree q. luo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-14 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 14: comparison of the simulated radiation patterns of the proposed antenna with (blue dashed line) and without (black solid line) connecting to the laptop computer at 5 ghz band 5 conclusion and discussion in summary, in this study a fractal ila using the 2nd iteration of the koch fractal geometry combined with the meander line has been proposed. this antenna exhibits wide operation bandwidth, moderate gain and high radiation efficiency. the radiation performance of the proposed antenna connecting to a laptop computer is also studied. it is found that connecting the usb dongle to the laptop computer can affect the resonant frequencies of the antenna and greatly change its radiation patterns. the frequency shifting of the proposed antenna caused by the laptop can be tuned by resizing the fractal, which has been justified by the simulation results. acknowledgement qi luo acknowledges the support for a scholarship from fundação para a ciência e tecnologia. the authors also thank instituto de telecomunicações, aveiro for fabricating the antenna prototype. references [1] w.c.su and k.l.wong, "internal pifas for umts/wlan/wimax multi-network for a usb dongle", microwave and optical technology letters (2006), 2249-2253. [2] c.c.lin, s.w. kuo and h.r. chuang, "a 2.4-ghz printed meander-line antenna for usb wlan with notebook-pc housing", ieee microwave and wireless components letters (2005), 546-548. [3]y. song, y.c.jiao and ed., "compact printed monopole antenna for multiband wlan applications", microwave and optical technology letters (2008), 365-367. [4] q. luo; j.r.pereira and h. salgado, "fractal monopole antenna for wlan usb dongle", proceeding of loughborough antennas and propagation conf., loughborough(2009), uk, 245-247. [5] z.n.chen, “note on impedance characteristics of lshaped wire monopole antenna”, microwave and optical technology letters (2008), vol.26, no.1, 365367. phi=90 degree q. luo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-14 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc wireless communication based on chirp signals for lora iot devices wireless communication based on chirp signals for lora iot devices v. fialhoab, f. azevedoa aárea departamental de engenharia electrónica e telecomunicações e de computadores adeetc, isel, portugal b centre of technology and systems – cts, uninova, fct, portugal vfialho@deetc.isel.ipl.pt fazevedo@deetc.isel.ipl.pt abstract — this paper presents the study of chirp signals for wireless communications between internet of thing devices used on low power wide area networks. up and down chirp concept is introduced as well as the chirp spread spectrum concept. a computationally efficient symbol decoding method is presented based of discrete fourier transform as an alternative to typical coherent detection. the proposed lora simulation model is implemented in matlab allowing the communication system evaluation based on bit error rate and packet error rate. keywords: chirp signal, spread-spectrum, iot, spreading factor, bit error rate packet error rate, signal to noise ratio. i. introduction internet of things (iot) concept emerged several years ago supporting the actual topic of smart cities. actually, every object (thing) can be monitored by transmitting information about a specific context and environment. the main paradigm of iot is the ability to connect to internet with a low cost, limited infrastructure and a very long life time battery, e.g. low power wide area network lpwan [1]. several standards support wireless communication between iot devices and each device to internet, using gateways [1]. each standard has different specifications, however they can be divided in two major categories: short range (less than 100 meters radius) and long range (greater than 1000 meters radius). first category includes protocols such as ieee 802.15.4, zigbee and bluetooth. second category includes system architecture and protocols similar to cellular network [1][2][3]. from the several lpwan standards presented in [2][3], sigfox® [4] and lora (long range modulation), defined by lora® alliance [5], are spreading faster than others lpwan [6]. both protocols operate across instrumental, scientific and medical frequency bands (ism). sigfox® standard allows a limited uplink and downlink messages per day. in order to access the information, a subscription fee is paid [4][6]. lora® is an open protocol, defined by the nonprofit association lora® alliance [5]. this protocol allows the creation of a private network with no data traffic limitation, packet size and no subscription fee need to be paid [6]. the frequency band allocated for european union is 868 mhz with configurable bandwidth. one last feature of lora® protocol is the duty-cycle, which may assume values between 0,1% to 1% of total transmitting time [5]. figure 1 presents a simplified lora® architecture in which are represented the end-device managed by a microcontroller, lora® gateway, data base and a personal computer to access the stored data. µc µc µc lora device lora device lora device lora gateway lora gateway lora gateway internet/cloud lora server fig. 1. lora simplified network concept. this paper presents the study of lora® wireless radio communication between the end device and the gateway. the modulation is based on chirp signals. therefore, the chirp spreading spectrum modulation concept is presented, such as a computationally efficient method for demodulation and decoding chirp signals. this method is based on discrete fourier transform (dft). a simulation model developed in matlab is presented, allowing the performance assessment in presence of additive white gaussian noise (awgn). the obtained simulation results are based on bit error rate (ber) and packet error rate (per). ii. lora chirp spreading spectrum modulation lora® wireless communication is based on chirp spreadspectrum (css) modulation scheme [5][6][7][8][9]. chirp signal with in-phase (i) and quadrature (q) is given by (1), 𝑠 𝑡 = e j (2π f𝑐 t+2π 𝛽 2 𝑡2 ) (1) where fc is frequency carrier, β is frequency variation slope given by the ratio between bandwidth (bw) and time symbol (tsymb), as expressed by (2) [7][9] 𝛽 = 𝐵𝑊 𝑇𝑠𝑦𝑚𝑏 , (2) time symbol is given by (3) 𝑇𝑠𝑦𝑚𝑏 = 2𝑆𝐹 𝐵𝑊 . 𝐶𝑅, (3) i-etc: isel academic journal of electronics, telecommunications and computers iot-2018 issue, vol. 4, n. 1 (2018) id-6 http://journals.isel.pt mailto:vfialho@deetc.isel.ipl.pt mailto:fazevedo@deetc.isel.ipl.pt where sf is the spreading factor, e.g., number of bits per encoded symbol, and cr the code rate [7]. figure 2 presents two signals normalized to tsymb: up-chirp (β>0) and a down-chirp (β<0). assuming bw=f1-f0 with f1 > f0 positive slope is obtained, otherwise slope is negative. fig. 2. up and down chirp signals normalized to tsymb. chirp frequency variation can be represented as depicted in figure 3. this graphical analysis allows tsymb estimation within a specific bw. it is also possible to infer the chirp slope. time freq. f 1 = f c +bw/2 f c bw t symb up-chirp down-chirp f 0 = f c -bw/2 fig. 3. spectrogram representation: up-chirp and down-chirp. according to lora® specification the bw, sf and cr may assume the values presented on table i [7]. table i lora® modulation main parameters bandwidth (bw) 125 khz to 500 khz spreading factor (sf) 7 to 12 coding rate (cr) 1 to 4 lora packet, presented in figure 4, contains a preamble, composed by six up-chirps and two down-chirps configured with a specific sf and bw values. data field contains optional header payload and payload cyclic redundancy check (crc) value [6]. the cr value in this filed may change between the values presented in table i. preamble data fig. 4. lora packet composed by preamble and data fields. a. css symbol coding and modulation lora® css technique allows symbol generation depending on the sf value, for a specific bandwidth. if only chirp slope variation were used to encode digital data, only two different symbols were generated. in css symbol coding and modulation is based on sf value within a fixed bw. as analogy to digital modulation, sf corresponds to the number of bits per symbol. therefore, according to (3), with sf=7 it is possible to encode 128 symbol during a tsymb within a given bw. in order to distinguish each symbol, using a chirp signal, a cyclic shift (cs) is used according to (4) 𝑐𝑠 = 𝑠𝑣 2𝑆𝐹 𝑇𝑠𝑦𝑚𝑏𝑜𝑙 , (4) where sv corresponds to the symbol value (0 to 2sf-1), and the reference symbol is an up-chirp with sv=0. this shift encoding increases the number of symbols in a factor of 2sf. for sv=16, the deviation from reference chirp is 16/128 of tsymbol. figure 5 show the spectrogram estimation for two sv values: 16 and 112 for a sf=7. fig. 5. chirp cyclic shift for two different symbol with sf=7. as conclusion, cs correspond to a time delay referenced to tsymbol. this characteristic enables the demodulation and decoding of 2sf symbols with a computationally efficient method [8]. b. lora® symbol demodulation and decoding lora® css demodulation process is based on a typical coherent demodulation concept, however the reference signal used for decoding is a down-chirp signal with the same sf used in the modulation process [8][9]. considering css a digital modulation and assuming perfect time and frequency synchronization, as well as symbol source generator with uniform distribution, the optimum receiver can be mathematically described as presented in [8]. for a computationally efficient implementation (discrete signals), the method proposed by [8] consists on the product of the received signal with a down-chirp, corresponding to the complex conjugate of a chirp with cs=0. figure 6 presents signal operation the block diagram of the proposed method, where d[n] is given by d[n] = r[n] . s*[n], (4) v. fialho et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt where * denotes s[n] complex conjugate, e.g. a down-chirp with sv=0. r [n] d [n] | dft(d [n]) |2 d [n] max ( d [n] ) sv=index(max ( d [n] )) sv s* [n] t symb fig. 6. lora® css demodulation concept. d[n] is the dft square modulus of the signal d[n]. symbol value, sv, corresponds to the index of the maximum value of d[n]. as result of operation described in figure 6 applied to the chirp signals generated by the cs presented in figure 5, the results are shown in figure 7. fig. 7. decoded lora® symbols: 0, 16 and 112. since sf is 7, r[n] is composed by 128 samples (0 to 127). therefore the cyclic shift, cs, imposed by coding process will impose the correct symbol decoding. iii. lora® wireless simulation scenario in this section it is presented the simulation system developed in matlab which implements css modulation described in section ii. the simulator block diagram is presented in figure 8. symbol mapping is generated according to gray code which ensures that adjacent symbols are mapped to bit patterns differing in one position only [9]. the output chirp signal, cs, depends on spreading factor (sf), frequency carrier (fc) and bandwidth (bw). the channel model is based on awgn with a configurable snr. bit sequence 2sf symbols cs chirp signal awgn channel demod. and decod. sf f c bw snr fig. 8. lora® css matlab simulation model. surrounded block with the dashed line, composed by the demodulator and decoder, is implemented with the algorithm proposed in section ii-b and represented in figure 6. as mentioned in section i, in eu, lora® frequency band is allocated in 868 mhz, however the frequency carrier used for the simulations where smaller in order to ensure that simulation time would be feasible. therefore all the results are normalized to fc. a. simulation results – spreading factor the spectrogram of three up-chirps with 125 khz bandwidth is presented in figure 9 for a snr=10 db. the sf values used for each chirp signal are 7, 8 and 9 with cr = 1. these configurations lead to a tsymb of 1ms, 2ms and 4ms. in conclusion, increasing sf a factor by one, the time symbol duration doubles, according to expression (3). fig. 9. lora® css symbols with sf 7,8 and 9. b. simulation results – symbol demodulation and decoding figure 10 presents lora® css spectrogram preamble with sf=7, composed by six up-chirps and two down chirps [7] followed by symbols 0, 16 and 112. fig. 10. lora® css packet spectrogram with sf = 7 with symbols 0, 16 and 112. for each received css symbol, r[n], the demodulation and decoding process presented in figure 6 is performed. after discarding the preamble, d[n] signals for symbols 0, 16 and 112 are presented in figure 11. as it turns out, index of d[n] maximum value corresponds to the transmitted sv. v. fialho et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt fig. 11. css dft d[n] signals for symbols 0, 16 and 112. figure 12 presents lora packet for snr=-10 db. as depicted, css power spectral density is less defined due to awng noise influence. nevertheless up and down chirps can be identified. the noise impact on css modulation is described in next section. fig. 12. . lora® css packet spectrogram with sf = 7 for snr = -10db. iv. lora model evaluation: ber and per to infer system evaluation under noise condition, matlab simulation scenario generates 20000 symbols based on binary sequence with a uniform distribution, grouped in 2sf values with gray coding. figure 13 show first fifty coded and decoded symbols for snr=-10 db with a sf=7. the obtained results denote the errors between the coded and decoded symbols. fig. 13. first 50 lora® css decoded symbols (sf=7) with snr=-10 db. awgn impact on lora® css modulation is evaluated based on ber. this evaluation is preformed comparing the matlab binary sequence and the decoded one. the obtained results are compared with work [8], under the same noise conditions, as presented in figure 14. fig. 14. simulation results based on ber for sf 7, 8, 9 and work [8]. assuming system reference ber=10-3, for each sf increase, the snr decrease 3db. therefore, if channel noise conditions decreases, to achieve the same performance, the sf needs to increase. assuming css link budget project, to achieved greater distances, sf value must increase [7], however symbol rate decreases. another lora® css modulation evaluation performance is based on per values. this is a useful parameter, since several lora® modules allow per measurement such as sx1276 [10]. lora® packet preamble is used for synchronization, therefore if one of the six decoded symbols is incorrect, the packet is discarded, and an error is incremented. for per simulation 10000 packets where generated and each preamble analyzed. the obtained values are presented in table ii. table ii per simulation results for different sf values per sf snr=-13 db snr=-11 db 7 38,7% 19,1% 8 9,5% 0,8% 9 0,02% n.a based on the obtained values it is possible to infer that per is approximately similar to ber, under the same css modulation parameters and noise channel conditions. v. conclusions a chirp spreading spectrum modulation scheme for lora® wireless communication has been presented, based on computationally efficient decoding method (dft). a simulation model is presented which enables the chirp signal configuration as well as channel noise conditions, which allows system performance assessment based on ber and per. the obtained results show that css modulation configuration parameters results on a trade of on binary rate, channel noise conditions and, consequently, the link budget between lora module and the gateway. v. fialho et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt the main advantage of css modulation comparing with other binary modulations is the occupied bandwidth, computationally efficient symbol decoding and power consumption. for future work we propose to improve the simulation model in order to evaluate signal interference between modules. therefore we need to adapt the proposed model to lora® channel access and, under these conditions, evaluate ber and per. for experimental validation it will be used sx1276 modules [10]. references [1] k. nolan, et al. an evaluation of low power wide area network technologies for the internet of things, in proc. of iwcmc 2016–ieee international wireless communications and mobile computing conference, paphos, cyprus, sep. 2016, pages 439-444. [2] s. al-sarawi, et al. internet of things (iot) communication protocols: review, in proc. of icit 2017–ieee 8th international conference on information technology, amman, jordan, may. 2017, pages 685690. [3] o. khutsoane, et al. iot devices and applications based on lora/lorawan, in proc. of iecon 2017–43rd annual conference of the ieee industrial electronics society, beijing, china, nov. 2017, pages 6107-6112. [4] sigfox, “about sigfox.” [online]. available: http://www.sigfox.com/en/#!/about. [5] “lora technology.” [online]. available: https://www.lora-alliance.org/. [6] l. angrisani, et al. lora protocol performance assessment in critical noise conditions, in proc. of rtsi 2017–ieee 3rd international forum on research and technologies for society and industry, modena, italy, sep. 2017. [7] lora™ modulation basics, an1200.22 application note-[online]available: https://www.semtech.com/uploads/documents/an1200.22 .pdf.. [8] l. vangelista, et al. frequency shift chirp modulation: the lora modulation, in ieee signal processing letters, vol.24, no. 12, pages 1818-1821, dec. 2017. [9] d. croce, et al. impact of lora imperfect orthogonality: analysis of link-level performance, in ieee communications letters, vol. 22, no. 4, pages 796-799, apr. 2018. [10] semtech wireless, sensing & timing, sx127x datasheet, [online] available: https://www.semtech.com/uploads/documents/ds_sx12 76-7-8-9_w_app_v5.pdf. v. fialho et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt https://www.semtech.com/uploads/documents/an1200.22.pdf https://www.semtech.com/uploads/documents/an1200.22.pdf facial features tracking system for adapted facial features tracking system for adapted human-computer interface rafael santos, arnaldo abrantes, pedro m. jorge electronics, telecommunication and computer engineering department, instituto superior de engenharia de lisboa, portugal a ​37178@alunos.isel.pt ​ ​aabrantes@deetc.isel.pt ​ ​pjorge@deetc.isel.pt abstract — ​this work describes the implementation of an eye gaze tracking system for a natural user interface, based only on non-intrusive devices, like intel realsense camera. this camera has depth and infrared information which make the extraction of face features more robust. through image processing, the implemented system is able to convert face elements to the corresponding focus point on the computer screen. two approaches were taken, one using the eye gaze and the other using nose tracking to move the cursor. preliminary tests show promising results. keywords: eye gaze; nose tracking; image processing; human-computer interaction; 3d camera. i. introduction the advance of technology has been motivating the development of new human-computer interfaces (hci) [4]. the act of looking at a screen is part of most natural interaction processes. however, the information that the eye gaze can provide is still not entirely exploited on current hci applications. gathering and processing the eye gaze from a user to interact with the computer is a topic already studied [3,4], but mostly based on specific technologies that are not available in mass market devices, such as, laptops or tablets. currently, most devices are equipped with a webcam, which can be used to collect visual information and provide feedback. however, this technology is not specific to acquire the information needed for eye gaze detection and tracking, missing in quality and sample rate. even if those requirements were met, light condition can limit eye gaze accuracy. to avoid some detection problems, a camera with infra-red light and depth information is used. this work describes a system to detect eye gaze based on intel realsense f200 camera, enabling a more natural form of human-computer interaction.the manufacturer expects that this camera can replace the generic webcam and there are already some laptops that include it. realsense sdk not only allow us to access some useful camera functions but with a faster and better quality. to help on image processing tasks, opencv [5] is used with some useful functionalities, like, haar cascade, hough transform or kalman filter. the implemented system also includes other face elements, like the nose, that could replace eye movements in the proposed hci. this papers is organized as follows: section ii presents the state of art of eye gaze systems. the implemented system is described in section iii. eye gaze results are evaluated in section iv and section v concludes this paper. 1 ii. state of the art eye gaze is a natural form of interaction, and identifying where a person is looking allows a computer to interact in a more human way [1,6]. however, replicating this procedure automatically in terms of human-computer interaction is not simple. eye gaze has been the subject of several studies over the past years [1,2]. rayner and pollatsek [7] presented the first work that used electro oculography to measure retina movements to monitoring the user while reading. in 2003, duchowski [8] proposed a similar method, where it was used a metallic hoop on contact lens to measure the electromagnetic field variations created by retina movement. morimoto and mimica [9], presented in 2004, a study where they used several techniques for eye gaze detection, where the main purpose was to develop interactive applications, based on eye structure and corneal reflex. however, none of these systems achieved good results, and the best results are obtained using specific and expensive hardware/equipment. new systems are being developed that use non-intrusive and more affordable devices. among these systems, the best performances are obtained with a source of infrared light [3]. however, distrust of infrared light exposure motivated the development of eye gaze detection and tracking systems that use current technology, such as webcams [2]. ​these systems still have limitations associated with head movement compensation [2,3]. it is also necessary to improve real time processing algorithms and hardware [3]. at the end of 2014, intel presented a new camera with depth, infrared and color information. the intel realsense is meant to work on tablets, computer and even smartphones. with this technology, the work proposed here is meant to update the one presented in [5]. iii. eye gaze detection system the main goal of the proposed work is to implement a new interaction system between human and computer with a state of art equipment. this method uses facial metrics to perform eye gaze tracking. it is desirable that the system comply some requirements like the use of commercial hardware, real time execution, enough precision for daily use and easy and fast calibration method. the implemented system can be described by the block diagram of figure 1. the first block, image acquisition, is where the system acquires information from the camera; acquire information block is where the system processes the acquired data; based on the features points extracted from the face, the system estimate the cursor position on the screen display, performed by cursor positioning block; finally, in i-etc: isel academic journal of electronics, telecommunications and computers cetc2016 issue, vol. 3, n. 1 (2017) id-1 http://journals.isel.pt mailto:37178@alunos.isel.pt mailto:aabrantes@deetc.isel.pt the last block, perform actions, the system will perform some actions requested by the user, besides cursor positioning, like clicking or moving a window. figure 1system block diagram. a. image acquisition the system starts with image acquisition from intel realsense camera (image acquisition block). it can be used two streams: color and depth images (see figure 2). figure 2 – rgb image and depth image. b. acquire information based on the acquired image, the system processes the informations to extract the required features from the face of the user, such as, head orientation, eye, pupil and nose positions. this block is implemented in two different operation modes: (i) as a common webcam and (ii) as a 3d camera. in mode (i) it was used only the color stream from the realsense camera with some image processing algorithms [3,5]. those algorithms consist on using haar cascade to find out the face position and eyes position. it also uses hough transform to find pupil circle and kalman filter to smooth the eye movements. mode (ii) use the infrared, depth and color information stream from the same camera. the camera sdk fuses those information to provide a more accurate and robust face features. however, the results obtained with the webcam mode were not very different from the other works presented in the previous section. figure 3 presents two frames where is possible to conclude that pupil tracking with traditional methods has not enough accuracy. figure 3 looking forward and looking up right. since realsense camera has depth, infrared and color information, it is expected that better eye gaze tracking results can be obtain. for that propose this system uses the real sense sdk to acquire the required features. for face pose, euler angles were used. on figure 4 an example is shown. figure 4 euler angle applied to aeronautics. after face pose information, the reference points were acquired. the system uses points from both eyes and nose (see figure 5 and 6). figure 5 eyes points. figure 6 nose point. although the sdk allows to acquire the reference points, their detection over time is very noisy. even with the eye without movement the reference points have some variations on the position. to reduce the noise, a kalman filter was used. figure 7 kalman filter applied to eye points. (x is the time and y is the pupil horizontal coordinate) figure 7 shows the results of the kalman filter apply to the pupil position (green line) along eye movements around the screen. it is possible to see that, with kalman filter, the detection is smoother (yellow line). however, kalman filter prediction causes some delay. nevertheless, the kalman prediction allows the system to keep eye tracking even with eye's natural blinking. r.santos et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-1 i-etc: isel academic journal of electronics, telecommunications and compute c. cursor positioning with face features, the system can process the information and execute actions. at this step, the system uses the previous information to determine the focus of attention point and position the cursor on the screen. the main objective is to use the eye as a mouse (illustrated in figure 9). figure 9 cursor movement. for eye positioning several methods were tested: ● eye as a mouse working with incremental position based on eye movements (up, down, left and right). see illustration in figure 10. the results were not as good as expected since we cannot make eye stop like a mouse. this approach has not enough precision; ● the cursor position is controlled by the pupil centre related with the eye. however, this method requires a calibration procedure to adjust the pupil position to the screen dimensions. for this propose a calibration pattern composed by 9 points is shown to the user (see figure 11). he should look at them in a predefined order and those eye positions are acquired. finally, this information is used to relate eye and cursor positions. as the pupil movements are too small compared with screen dimensions, this approach has some limitations; ● the position of the pupil relative to the limits of the eye defines the position of the cursor relative to the screen (see figure 12). for this method the minimum and the maximum value for pupil centre variation are calculated. those values define screen boundaries. good results are achieved but the method needs to be improved due to small pupil movements and sensibility. figure 10 cursor movement as mouse differential. figure 11 calibration screen. figure 12 eye with screen reference. however, the results are not as good as expected. so, to define the position of the cursor on the screen, another approach was implemented. last approach is used, but now the cursor position is controlled by the nose position relative to a predefined window. figure 13 illustrate this approach. figure 13 nose with screen reference. this method obtains the best results, allowing the user to control the cursor position on the screen with good accuracy. d. perform actions once the cursor is controlled with enough precision, some actions can be implemented, allowing the user to interact with the computer. the implemented actions are: 1. eye blink to simulate left and right mouse click; 2. mouth open to simulate left mouse click. the first action (1) needs the eye circular reference points to determine if the eye is open or closed, as shown in figure 14. with those points, a minimum and maximum distances are calculated. the minimum distance would means that the eye is closed, and maximum that the eye is open (see figure 14). r.santos et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-1 i-etc: isel academic journal of electronics, telecommunications and compute figure 14 eye closed minimal distance and eye opened maximum distance. good results were achieved as the system is able to detect 7 clicks of 10 tries. with some training, the user is able to get 9 clicks of 10 tries, from the system. the second action (2) was implemented to complement nose cursor positioning. as the user is moving the head/nose to move the cursor, eye blinking can not be used to perform actions. alternatively, mouth is used to simulate mouse click (see figure 15). the results are better with this approach, being the system able to detect 9 click of 10 tries, without user training. figure 15 mouth open to determine the maximum distance. iv. testing and analysis to evaluate system accuracy, several tests were carried out. based on usability tests, the approach using eye movement had not enough accuracy. the user can actually move the cursor horizontally, but vertically no movement is practically noticed, since the pupil movement is minimal. figure 16 cursor movement looking at the top left side of the screen. figure 17 – cursor movement looking at the top of the screen. usability tests were also performed with nose tracking approach. the user is able to move the cursor to the desired position, just with a short training (see figure 18). to perform mouse click with the eye blink, a few tries are needed for the user to get the desired result (see figure 19). figure 18 cursor positioning with nose movement. figure 19 – right eye blink to perform click over desktop icon. with mouth open and close, the user is not only able to click over an icon, as he is able to move it to another place, doing drag and drop action (see figure 20). figure 21 mouth open/close click and drag and drop. precision tests were performed with two screen dimensions, 24” and 15”, with two resolutions, fullhd and hd. to have some comparison between tests, a visual pattern is displayed on the screen, like a calibration screen. the user should look to different points from this pattern and the necessary information is acquired. tables 1 and 2 show the errors between pattern and cursor points, for eye gaze and nose tracking, respectively. r.santos et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-1 i-etc: isel academic journal of electronics, telecommunications and compute table 1 cursor positioning error through eye gaze. from table 1, eye gaze presents a mean error of 409 pixels for x ​coordinate (horizontal) and 336 for ​y coordinate (vertical). the same test was performed for nose tracking and cursor positioning, the results are shown on table 2. table 2 cursor positioning error through nose tracking. from table 2, nose tracking represents a much smaller mean error, with 45 pixels for ​x ​coordinate (horizontal) and 20 pixels for ​y ​coordinate (vertical). tests show a big variability in the results. mean error is lower with nose tracking than with eye gaze, due to the information that can be acquired with the nose and with the eyes. v. conclusion this project implements a system that uses non-intrusive technology to develop a more natural human-computer interface. the ambient light is one of the main factors [5] that causes errors in eye gaze tracking. the use of intel realsense camera, that has three sources of information (infrared, depth and color), decreases the impact of the previous problem since the fusion of those information improves the robustness and the accuracy of tracking face features. other aspect is user distance related with the camera. camera’s distance to the eye will influence pupil’s detection since, eye resolution decreases and less information exist to process and acquire the pupil with precision. with those requirements meet the system could get promising results, being the user be able to perform actions. tests for cursor movement with eye show that work needs to be done in order to have better accuracy. one possibility is to develop head movement compensation for eye gaze, this way the eye estimated position can have more accuracy. in another way, nose tracking obtained enough precision to perform actions with good accuracy. actions like drag & drop or mouse clicking to minimize a windows are possible to execute. mouse clicking with mouth open/close showed good results as it complemented the nose tracking and cursor movement. a commercial solution for a human-computer interface might be possible to implement if: (i) a itel realsense camera is available (or other device with the same information); (ii) the cursor movement is controlled with nose movements and (iii) mouse clicking is performed with mouth open/close. references [1] poole, a. e ball, l. j. (2005). eye tracking in human-computer interaction and usability research: current status and future prospects. in c. ghaoui (ed.) encyclopedia of human-computer interaction ​, pennsylvania: idea group, inc. [2] tunhua, b. b. w., changle, l. s. z. e kunhui, l. (2010). real-time non-intrusive eye tracking for human-computer interaction. proceedings of the 5th international conference on computer science and education (iccse).1092-1096. [3] wild, d. j. (2012). gaze tracking using a regular web camera. [4] drewes, h. (2010). eye gaze tracking for human computer interaction. [5] santos, r., santos, n., jorge, p. e abrantes, a. (2013) eye tracking system with common webcam. lisboa, isel. elsevier. [6] jaimes, a. and sebe, n.(2007). multimodal human computer interaction: a survey. computer vision and image understanding, 1-2 (108):116-134. [7] rayner, k. and pollatsek, a. (1989). the psychology of reading. prentice hall, nj. [8] duchowski, a. t. (2003). eye tracking methodology: theory and practice. springer-verlag ltd, london. [9] morimoto, c. h. and mimica, m. r. (2004) eye gaze tracking techniques for interactive applications. elsevier inc. r.santos et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-1 i-etc: isel academic journal of electronics, telecommunications and compute intelligent sensors for real-time hazard detection and visual indication on highways intelligent sensors for real-time hazard detection and visual indication on highways j. oliveiraa, j. soaresa, a. r. lourencoa,b, r. p. duartec,d aadeetc, instituto superior de engenharia de lisboa (isel), 1900-001 lisboa, portugal binstituto de telecomunicações (it), 1049-001 lisboa, portugal cdei, instituto superior técnico (ist), 1049-001 lisboa, portugal dinstituto de engenharia de sistemas e computadores investigação e desenvolvimento (inesc-id), 1000-029 lisboa, portugal joaoliquito@gmail.com joaordsoares@gmail.com arlourenco@deetc.isel.pt rui.duarte@ist.utl.pt abstract— traffic collisions, in particular high speed car accidents often result in huge damages, long traffic queues and loss of human lives. in this work, we present an intelligent modular autonomous system that monitors traffic in highways and alerts drivers of sudden stops, in poor visual conditions. the system is composed of several identical modules, to be placed in the middle of a highway’s lane, that sense the lights and communicate their presence and velocity to their neighbour modules via rf. with such information, the nearby modules estimate the velocity of the passing cars. when the module ahead detects a car passing at a much slower speed than what was previously estimated, it alerts the other modules, so they produce a visual indication for the oncoming drivers, preventing accidents. keywords: intelligent transportation systems, auto traffic monitoring, low-power embedded system, ad-hoc wireless communication, sensor network. i. introduction road safety is a major societal issue. in 2015, more than 26,000 people died on the roads of the european union, i.e. the equivalent of a medium town. additionally, for every traffic death on europe's roads there are an estimated 4 permanently disabling injuries (such as damage to the brain or spinal cord), 8 serious injuries and 50 minor injuries [2]. these statistics motivate the development of intelligent transportation systems (its) [1] as innovative services related to different modes of transport and traffic management, aiming a smarter, and helpfully safer, use of transport networks. this paper presents an autonomous embedded system that tries to improve safety in highways in poor visibility conditions due to the road profile, or adverse meteorological conditions. it acts as an active alert signal to other drivers. previous works [7,10] detect automatically traffic accidents using image processing techniques, based on images acquired by infrastructures deployed on road. other approaches consist of using in-vehicle systems [8], or smartphones [9], that try to automatically detect traffic accidents using accelerometers and acoustic data. these approaches aren't able to automatically warn oncoming traffic of the accident. the proposed system is to be installed on the road, and is composed of several modules that work together to perform real-time traffic monitoring and detection of hazardous situations. each module incorporates very bright leds that are activated when an hazardous situation is identified, thus warning drivers approaching the location. one of the key features, that differentiates the proposed systems from others [6] is the fact that it is pro-active in detecting accidents by exchanging messages between modules. each module measures the time elapsed between the communication of a preceding module and the detection of the vehicle by the sensor. therefore, knowing the distance between each device it is possible to determine a vehicle's speed, and if it reduced drastically the velocity, or even if it stopped. the main contributions of this work are: definition of the architecture of the system and its requirements; definition of hardware and software for the embedded system; communication protocol; characterization and modelling of the light pattern produced by vehicles; simulation of the operation of the modules. the rest of the article is organized as follows: section ii presents the system architecture, giving a general overview of the system; it then describes the module, and the criteria used for the selection of the hardware, and decisions made for the software; section iii give details about the experimental evaluation, and finally in section iv we outline the main conclusions and draw lines for future work. ii. proposed system the proposed system is composed of several identical modules that acquire information about traffic from sensors, detecting vehicles, estimating its velocity and reacting to hazard situations. fig. 1 demonstrates a schematic representation of the system, when installed on a highway. it is composed by several modules positioned equidistant from each other, installed on the middle of the road, and represented by an orange dot figure. on the right-hand lane, two cars are involved in an accident. the system identifies this situation and the preceding modules emit a visual warning (bright red light) for the drivers approaching the location, in order to reduce their velocity, hence avoiding an imminent chain crash. the passage of each car can be detected using several methods [4], that are classified based on their intrusiveness in the road. the intrusive methods require the placing of sensors on or in the road, as pneumatic road tubes, piezoelectric sensors, magnetic loops [5]. the non-intrusive techniques are based on remote observation [7,10]. i-etc: isel academic journal of electronics, telecommunications and computers cetc2016 issue, vol. 3, n. 1 (2017) id-6 http://journals.isel.pt fig. 1. illustration of the deployed system in operation. on the right-hand lane after identification of an accident, the modules warn the drivers of the dangerous situation ahead. on this work we follow an intrusive approach, but based on a light sensor. the detection of the car passage is performed by analysing the patterns of vehicles' head lights, using a photoelectric sensor. the automatic decision about the existence of an accident is based on a message passing algorithm between modules, through a wireless ad-hoc network. each message informs the neighbour modules about each vehicle detection. if a module doesn't detect a passing vehicle during an expected period, either too slow or stopped, then it alerts its neighbourhood until de regular traffic flow is normalized. a. hardware due to the system's operating environment, and its implement restrictions, each module is comprised of a solar cell, a battery, a power management unit, a micro-controller (mcu), an rf transceiver, a light-sensor and leds, as illustrated in fig. 2. the battery is charged via the solar cell. the mcu provides all the control logic of the module and its peripherals, e.g. led, rf module, and light sensor. fig. 3 shows a 3d rendering of the system in its case. the solar cell is on the top, and the leds positioned to guarantee that drivers can see them. the light sensor, is positioned with an angle that better detects the vehicles' head lights. the translucent top panel is to be manufactured of resistant plastic such as plexiglass, while the rest of the case can be manufactured of a less expensive, but equally resisting material. fig. 2. organization of the blocks in the module. b. software intelligence with mechanical and power restrictions, the modules had to be implemented by a very low-power, yet flexible, platform. fig. 3. a 3d rendering of the case for the modules, including its components. on this account the intelligence of the system is modelled by a finite state machine (fsm). the system is meant to react to 2 inputs: the light sensor, and reception of a message from the rf module; and actuate 2 outputs: visual indication via leds, and transmission of message to the rf module. to support such functionality, an fsm was thought to be programmed on the mcu to control the system using 4 states: sleep, detection, decoding, and communication. change of states are described in fig. 4. fig. 4. state machine of the controller. after reset, the system is in the initial state (state 1: sleep), waiting for a message from another module (state 3: decoding) or detection of a vehicle (state 2: vehicle detection). in state 3, the received status can be: a) vehicle detection, b) send a warning, and c) reception acknowledgement. if the message received is a warning, then it goes to state 4 ("communication / alert") and activates the light signalling during a period of time and the message is retransmitted. at the end of transmission, the fsm goes back to the initial state. j. olivira et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-6 i-etc: isel academic journal of electronics, telecommunications and computers in the second state ("vehicle detection") it starts to count the time after receiving the message. the time stops when a vehicle is detected, and since the physical distance between modules is known an estimate of speed can be calculated. after the vehicle is detected, a corresponding message is sent to the neighbour modules. the system waits for acknowledgement messages. if the time waiting is greater than a threshold, it changes to state 4, and turn on the led light and send a warning message to be received by all nearby modules. the device then returns to state 1. c. communication the modules communicate with each other via rf broadcast messages, so that other modules within coverage range can bypass a faulty module in the system and continue its operation. there are 3 types of messages exchanged by the modules are: • detection: after a vehicle’s detection, with is velocity estimate. • warning: when a vehicle doesn't pass on the expected time interval, due to abrupt speed decrease, or did not pass at all, due to an accident. • acknowledgement: indication of the information received by the module. figs. 5-7 illustrate 3 scenarios of operation of the proposed system. in the first one, a car is detected and its information exchanged between modules and not abnormality is detected. the second scenario shows the case where a car as drastically reduced its velocity, and consequently too longer to reach the second module, thus the second module broadcasts a warning message. the third one illustrates the case where the second module failed to send an acknowledgement for the message sent by the first module, and a warning message is then broadcasted by the first module after a timeout period. fig. 5. message sequence chart of normal situation while detecting a car and passing information between modules. fig. 6. message sequence chart of an abnormal situation where the car is not detected in the expected interval by the second module, and a warning is broadcasted. fig. 7. message sequence chart of an abnormal situation where the car is not acknowledged by the second module, and a warning is broadcasted by the first module. iii. experimental validation the first step taken for the implementation of the system consisted on a simulation in matlab environment. it simulated the fsm, the detection algorithm, and the message passing scheme. moreover, it allowed to study the behaviour of the modules to the passage of the light intensity profile of each vehicle, and check the transmitted messages. fig. 8. structure over the road used to acquire the photos during the evening. j. olivira et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-6 i-etc: isel academic journal of electronics, telecommunications and computers the light intensity profile is based on real data, which were obtained by filming part of a highway to collect several light patterns from different vehicles. fig. 8 shows the photograph of the structure over the road used to acquire the photos during the evening. several frames with light footprints of vehicles' headlights were extracted from several video-recordings. they were then cropped and converted to grayscale, as illustrated on fig. 9. to simulate the light sensor analysis, only the intensity values of the footprint were considered, shown in blue in fig. 10, and low frequency signal in red. fig. 9. example of a photograph used to create the models for the different light patterns. fig. 10. example of the data extracted from the light footprints to create the models for different vehicles. fig. 11. output of the simulator created to test the modules. fig. 11 presents the output of the simulation of the system. here, the photo-sensors in the first module detect variation in the intensity of the light sensed, and identify it as a vehicle. this information is then broadcasted to the following modules, and the module waits for the following vehicle before broadcasting new information. the other modules identify the origin of the received message and start counting the time passed between the reception of the message and the detection of the car. the choice of hardware for the implementation involved the selection of the most adequate mcu, which is the unit supporting the intelligence of the module. the parameters used for this selection, were: power consumption, number of io, number of timers, clock frequency, price, and the existence of a development kit. the chosen platform was the texas instruments msp-exp430fr5969 along with the 430boostcc110l kit to provide rf communication using the 868-870mhz band for short range devices (srd). fig. 12 presents the photos of the 2 modules. fig. 12. photographs of the development kits used: mspexp430fr5969 (left) and 430boostcc110l (right). the msp430fr5969 [3] has a very low power consumption, enough i/o ports an uart to connect with peripherals, and an acceptable number of timers needed to count the time that vehicles spend in between modules. the key features of this mcu are listed in table 1. table 1 relevant features of msp430fr5969 parameter value frequency 16 mhz non-volatile memory 64 kb sram 2 kb gpio 40 pins uart 2 units 16-bit timers 5 units active power 101.25 �a / mhz standby power 0.5 �a wake-up time 7 �s price $2,35 (usd) j. olivira et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-6 i-etc: isel academic journal of electronics, telecommunications and computers iii. conclusion we have proposed and demonstrate a novel system to detect hazardous conditions on highways due to sudden stops, or abrupt velocity reductions of vehicles, which are usually associated with accidents. the system is to be used in poor visibility conditions, due to adverse weather conditions, and highway profile. it is composed of many modules present on the highway, that detect passing vehicles by sensing light from their headlights. we have demonstrated the concept by presenting simulation results and an earlier implementation of the embedded system. future work involves exhaustive real-life measurements and tests on different roads and highways with distinctive characteristics, at different times of the day and under different weather conditions. references [1] handbook of transport systems and traffic control, volume 3. emerald, inc., 2010. [2] european commission. statistics on accidents data. http://ec.europa.eu/transport/road_safety/specialist/statist ics/index_en.htm, 2016. [3] texas instruments. msp430fr5969. http://www.ti.com/product/msp430fr5969. [4] guillaume leduc. road traffic data: collection methods and applications. technical report, jrc european commision, 2008. [5] filipe palhinha, duarte carona, antónio serrador, and tomás canas. wireless magnetic based sensor system for vehicles classification. procedia technology, 17(0):632 – 639, 2014. conference on electronics, telecommunications and computers (cetc 2013). [6] radixtraffic. idu600 daytime dimming photocell. http://radixtraffic.co.uk/assets/uploads/ idu600.pdf, 2013. [7] s. sadeky, a. al-hamadiy, b. michaelisy, and u. sayed. real-time automatic traffic accident recognition using hfg. in pattern recognition (icpr), 2010 20th international conference on, pages 3348–3351, aug 2010. [8] h.m. sherif, m.a. shedid, and s.a. senbel. real time traffic accident detection system using wireless sensor network. in soft computing and pattern recognition (socpar), 2014 6th international conference of, pages 59–64, aug 2014. [9] jules white, chris thompson, hamilton turner, brian dougherty, and douglasc. schmidt. wreckwatch: automatic traffic accident detection and notification with smartphones. mobile networks and applications, 16(3):285–303, 2011. [10] kimin yun, hawook jeong, kwang moo yi, soo wan kim, and jin young choi. motion interaction field for accident detection in traffic surveillance video. in pattern recognition (icpr), 2014 22nd international conference on, pages 3062–3067, aug 2014. acknowledgements the authors would like to thank cardio-id technologies for the support acquiring the development kits from texas instruments. j. olivira et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-6 i-etc: isel academic journal of electronics, telecommunications and computers autonomic function evaluation in an intermittent lead exposure animal model autonomic function evaluation in an intermittent lead exposure animal model l. shvachiy1, v. geraldes1,2, m. carvalho1, i. rocha1,2 1cardiovascular autonomic function lab, cardiovascular centre of university of lisbon; 2institute of physiology, faculty of medicine, university of lisbon, av. prof. egas moniz, 1649-028 lisbon, portugal shvachiy.liana@gmail.com vgeraldes@medicina.ulisboa.pt isabelrocha0@gmail.com abstract — lead (pb) is a toxic metal, which widespread use has resulted in environmental contamination, human exposure and significant public health problems. the autonomic nervous system, being a homeostatic controller, is impaired in acute and chronic lead exposure. in fact, sympathoexcitation associated to hypertension and tachypnea has been described together with baroreflex and chemoreflex dysfunction. however, up to date, no studies described the autonomic effects of an intermittent lowlevel lead exposure. in the present work, we addressed in vivo, autonomic behaviour in rats under chronic pb exposure (control) and in rats under intermittent pb exposure. for that, arterial blood pressure (bp) and ecg were recorded in 28 weeks old animal and low frequencies (lf) and high frequencies (hf) were determined (to estimate sympathetic and parasympathetic activities) using fisiosinal software with wavelet module. our preliminary results depict that intermittently lead-exposed rats show a significant decrease in systolic bp, without significant changes on lf, hf and lf/hf bands, when compared to chronically pbexposed rats. our data suggests that the autonomic dysfunction induced by lead exposure is similar in a chronic and intermittent pb exposure. nevertheless, it seems that an intermittent exposure was no effect on systolic bp values. the present study brings new insights on the environmental factors that influence autonomic and cardiovascular systems during development, which can help apprise public policy strategies to prevent and control the adverse effects of pb toxicity. keywords: lead toxicity, autonomic activity, fisiosinal, wavelet analysis, heart rate variability i. introduction lead is the most commonly used heavy metal worldwide for over 8000 years in a large amount of industries like automobiles, ceramics, paint, plastics, among others, in behalf of its unique properties such as high malleability, low melting point, softness, ductility and resistance to corrosion[1], [2]. regarding this wide usage, it is obvious that there is a rise in environmental free lead and its occurrence in biological systems concerning the non-biodegradable nature of this heavy metal. lead toxicity, due to its ingestion, inhalation or, less probably by direct contact, can evoke irreversible effects in a wide amount of body functions affecting mainly the cardiovascular [3], [4] (being one of the causes of hypertension and promoting atherosclerosis, thrombosis, arteriosclerosis and cardiovascular disease) , hematopoietic[1], [5], reproductive[6] (both, in woman and men, leading to misscariage, inferitility and premature delivery in women and reduction of libido, abnormal spermatogenesis and infertility in men), [7] and renal (leading to acute and chronic nephropathy, by molecular and tubular changes) [8] systems. being easily accumulated in bone and soft tissues, and due to the ability of lead to cross the blood brain barrier and to substitute cations like ca2+, mg2+, fe2+ and na2+, lead is also a neurotoxin. these alterations, induce changes in the neurogenesis, especially when the exposure happens in the early stages of life, neurodegeneration and changes on glial cells, that support neurons and play a crucial role in synaptic transmission and plasticity. the changes in the molecular and cellular processes lead to cognitive and behaviour alterations, particularly during developmental phases, persisting during the lifetime. [1], [9], [10] additionally, regarding the autonomic tone, that regulates the cardiorespiratory processes, the exposure to toxic levels of lead produce a significant increase in sympathetic activity associated to high blood pressure, decreased baroreflex function, increased chemoreceptor sensitivity and tachypnoea [11]. two types of lead toxicity could be defined. firstly, acute toxicity, usually in high-levels, is quite uncommon and happens by occupational exposure, while chronic low-level exposure toxicity is more common, happening in a household environment and much severe if not treated in time. however, up to date, no studies were performed to establish the autonomic effects in an intermittent low-level lead exposure. this type of exposure is increasing due to the augmented migration and to the implementation of school exchange programs. nevertheless, people tend to come back to their hometowns after some years abroad, returning to their familiar environment. additionally, in workplaces where the use of lead is recurrent, rotation of workers and moving of these to other workplaces without lead exposure has been performed during the past years as a possible prevention of i-etc: isel academic journal of electronics, telecommunications and computers cetc2016 issue, vol. 3, n. 1 (2017) id-11 http://journals.isel.pt mailto:vgeraldes@medicina.ulisboa.pt adverse health effects caused by occupational permanent lowlevel lead exposure. therefore, our work is focused on studying the autonomic changes provoked by a newintermittent low-level lead exposure profile.the section ii (methods) describe the design of the present study, including the development of two animal models and provide a description of the procedures that were performed.for the autonomic evaluation. the preliminary results obtained in our study are presented in the section iii and then discussed, comparing our results with other published evidence in section iv. some final, summative statements, based upon our findings are reported in section v: ii. methods a. development of the animal model of chronic and intermittent lead exposure seven-day pregnant wistar rats (n = 2) were exposed to lead during pregnancy, via water containing lead acetate (0.2% w/v) that replaced tap water. after birth and weaning at 21 days, the rats of both sexes were exposed to the same lead solution as that of their mother and were divided into 2 groups (n = 6/group; pb1 – used as control group – and pb2). at 12 weeks, the intake of the solution ended in both groups; and this was followed by a period without exposure that lasted up to 20 weeks. the pb2 group underwent a second exposure which lasted 8 weeks (named intermittent lead exposure group) while pb1 wasn’t exposed to lead for the second time (named chronic lead exposure group). b. surgical protocol at 28 weeks, the acute experiment was carried out in both groups of animals (n=12). animals of both sexes were anesthetized with sodium pentobarbital (60mg/kg, ip). the femoral artery and vein were cannulated for blood pressure monitoring and injection of saline and drugs, respectively. rectal temperature was maintained between 36.5-39ºc. the electrocardiogram (ecg) was recorded with subcutaneous electrodes inserted in three of the four members and heart rate was also obtained through this registration method (neurolog, digitimer[12]). c. data acquisition signals were amplified with the alternative current pre amplifier gain of 5k and a high-pass filter of 0.1hz, filtered by band-pass filter of 500-5000hz and a 20hz wave width (neurolog, digitimer) and acquired at 1 khz (powerlab, adinstruments[13]). all signals were converted to digital form and stored for further analysis. d. analysis of heart rate and systolic blood pressure variability using fisiosinal [14], an in-house developed software with wavelet module (db12)[15], [16], the lf, hf and lf/hf indexes were determined in basal conditions. low frequencies (lf; 0.15-0.6 hz) that were obtained from systolic bp are a marker of sympathetic activity, high frequencies (hf; 0.6-2.0 hz) attained from r-r interval represent both parasympathetic and respiratory variations, and lf/hf is the ratio between sympathetic and parasympathetic systems. in this study and to the fact that the respiratory rate was high in some animals, we proceeded to an adjustment of the high frequencies band while maintaining its lower end and extending to its upper limit to 2.4 hz. e. statistical analysis variables are expressed as mean and sem. statistical data was obtained by comparisons between groups of animals – pb1 and pb2 – using student’s t-test. significance was considered to p<0.05. data was analysed using graphpad prism 6 software[17] (graphpad software, inc., usa). iii. results a. cardiovascular variables and autonomic tone baseline levels of all measured physiological parameters were similar between the 2 groups, except for systolic blood pressure with significantly higher values in pb1-rats compared with pb2-rats as shown in table 1. table 1 baseline levels of all measured physiological parameters. bp, blood pressure; hr, heart rate; rf, respiratory frequency values presented as mean ± sem (n=6), ap<0.05 group pb1 pb2 mean bp (mmhg2) 125 ± 4 116 ± 3 systolic bp (mmhg2) 144 ± 3 a 126 ± 4 diastolic bp (mmhg2) 110 ± 5 106 ± 3 basal hr (bpm) 371 ± 8 404 ± 17 basal rf (cpm) 56 ± 4 58 ± 5 l. shvachiyet al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-11 i-etc: isel academic journal of electronics, telecommunications and computers the sympathetic and parasympathetic tone evaluated through lf and hf band, respectively, as well as lf/hf band presenting the ratio between sympathetic and parasympathetic systems (1.5 ± 0.3 vs 1.7 ± 0.5 mmhg2, 1.9 ± 0.7 vs 2.8 ± 1.2 bpm2 and 1.2 ± 0.4 vs 1.1 ± 0.3 mmhg2/bpm2, respectively) didn´t change significantly between the 2 groups, as shown in figure 1, 2, and 3, respectively, for each band. m m h g 2 p b 1 p b 2 0 1 2 3 4 5 figure 1. sympathetic activity presented by lf (low frequencies) band for each group. no significant difference in lf band values was observed in the pb2 group – intermittent lead exposure (1.7 ± 0.5 mmhg2) when compared to the pb1 group – chronic lead exposure (1.5 ± 0.3 mmhg2). values presented as mean ± sem (n=6). b p m 2 p b 1 p b 2 0 1 2 3 4 5 figure 2. parasympathetic activity presented by hf (high frequencies) band for each group. no significant difference was detected in hf band of the pb2 group intermittent lead exposure (2.8 ± 1.2 bpm2) when correlated to the pb1 group chronic lead exposure (1.9 ± 0.7 bpm2). values presented as mean ± sem (n=6). figure 3. ratio between sympathetic and parasympathetic systems presented by lf/hf band values. no significant difference between lf/hf band of the pb2 group – intermittent lead exposure (1.1 ± 0.3 mmhg2/bpm2) and the pb1 group chronic lead exposure (1.2 ± 0.4 mmhg2/bpm2) was distinguished. values presented as mean ± sem (n=6). iv. discussion the current study is the first to address a new profile of lead exposure, the intermittent low-level lead exposure. our reported results demonstrate that the autonomic effects of this type of exposure are similar to the chronic lead exposure, since we didn´t find differences in sympathetic and parasympathetic tones between the two types of exposure, providing an association between the autonomic function and lead intoxication, regardless of the type of lead exposure. in relation to the other physiological parameters evaluated, i.e. blood pressure, heart rate and respiratory frequency, all are similar except for systolic blood pressure that is significantly decreased in the intermittent lead exposure group (pb2) (126 ± 4 mmhg2) when compared to the chronic lead exposure group (144 ± 3 mmhg2). the decrease on arterial compliance, which is used as an indication of arterial stiffness, provoked by a chronic lead toxicity may be the reason for the increased systolic blood pressure recorded in the chronic lead exposure group (pb1). even though autonomic function is not known to be primarily affected by chronic lead exposure, the autonomic nervous system has been found compromised by pb-poisoning, causing an overall autonomic dysfunction[11]. additionally, it was shown that lead reduces baroreflex sensitivity, vagal parasympathetic tone, and increases sympathetic activity, by impairment of dopamine and acetylcholine transmission, as well as oxidative stress induced by lead poisoning [18]–[20]. l. shvachiyet al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-11 i-etc: isel academic journal of electronics, telecommunications and computers these reported effects baroreflex hyposensitivity, sympathetic overexcitation and decreased parasympathetic tone are associated with several pathologies such as hypertension, acute cardiac ischemia or even heart failure. the mechanisms underlying those pathological effects haven’t been clarified but it is thought that there is an initial protective reaction which in time, turns into deleterious sympathoexcitation [11]. here we present our preliminary results regarding the autonomic nervous system evaluation. these results indicate that the intermittent low-level lead exposure is associated with sympathetic hyperactivity, as already described for a chronic exposure, but with no effect on blood pressure values, since in the intermittent exposure the animals remain normotensive. regarding the fact that our results are preliminary, more evaluations should be carried on, for a full characterization of the new animal model that was developed, using other electronic equipment, machinery and software’s (such as blood pressure and ecg telemetry) or other molecular and physiological evaluations. these additional evaluations would be of extreme importance to better understand the underlying mechanisms of lead toxicity. v. conclusion in conclusion, the present study brings new insights into how different lead poisoning profiles may influence autonomic and cardiovascular systems during developmental phases. thus, helping to upraise public policy strategies to prevent and control the adverse effects of pb toxicity. references [1] g. flora, d. gupta, and a. tiwari, “toxicity of lead: a review with recent updates,” interdiscip. toxicol., vol. 5, no. 2, pp. 47–58, 2012. [2] world health organization, “exposure to lead: a major public health concern,” world heal. organ., p. 6, 2010. [3] a. navas-acien, e. guallar, e. k. silbergeld, and s. j. rothenberg, “lead exposure and cardiovascular disease a systematic review,” environ. health perspect., vol. 115, no. 3, pp. 472–482, 2007. [4] n. d. vaziri, “mechanisms of lead-induced hypertension and cardiovascular disease.,” am. j. physiol. heart circ. physiol., vol. 295, no. 2, pp. h454–h465, 2008. [5] d. c. basha, s. s. basha, and g. r. reddy, “leadinduced cardiac and hematological alterations in aging wistar male rats: alleviating effects of nutrient metal mixture,” biogerontology, vol. 13, no. 4, pp. 359–368, 2012. [6] n. a. brown, “reproductive and developmental toxicity of styrene,” reprod. toxicol., vol. 5, no. july 1985, pp. 3–29, 1991. [7] m. ahamed and m. k. j. siddiqui, “low level lead exposure and oxidative stress: current opinions,” clin. chim. acta, vol. 383, no. 1–2, pp. 57–64, 2007. [8] m. loghman-adham, “renal effects of environmental and occupational lead exposure.,” environ. health perspect., vol. 105, no. 3, pp. 103–106, 2008. [9] c. d. toscano and t. r. guilarte, “lead neurotoxicity: from exposure to molecular effects,” brain res. rev., vol. 49, no. 3, pp. 529–554, 2005. [10] y. finkelstein, m. e. markowitz, and j. f. rosen, “low-level lead-induced neurotoxicity in children: an update on central nervous system effects,” brain res. rev., vol. 27, no. 2, pp. 168–176, 1998. [11] v. geraldes, m. carvalho, n. goncalves-rosa, c. tavares, s. laranjo, and i. rocha, “lead toxicity promotes autonomic dysfunction with increased chemoreceptor sensitivity,” neurotoxicology, vol. 54, pp. 170–177, 2016. [12] “neurolog system, signal processing &amp; conditioning electrical stimulation.” [online]. available: https://digitimer.com/products/neurologsystem/. [accessed: 27-nov-2017]. [13] “https://www.adinstruments.com/products/powerlab.” . [14] c. tavares, r. c. martins, s. laranjo, and i. rocha, “computational tools for assessing cardiovascular variability,” 1st port. meet. biomed. eng. enbeng 2011, 2011. [15] i. daubechies, “ten lectures of wavelets,” 1992. [16] i. daubechies, “the wavelet transform, timefrequency localization and signal analysis,” inf. theory, ieee trans., vol. 36, no. 5, pp. 961–1005, 1990. [17] “home graphpad.com.” [online]. available: https://www.graphpad.com/. [accessed: 27-nov2017]. [18] h. bielarczyk, x. tian, and j. b. suszkiw, “cholinergic denervation-like changes in rat l. shvachiyet al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-11 i-etc: isel academic journal of electronics, telecommunications and computers hippocampus following developmental lead exposure,” brain res., vol. 708, no. 1, pp. 108–115, 1996. [19] n. bourjeily and j. b. suszkiw, “developmental cholinotoxicity of lead: loss of septal cholinergic neurons and long-term changes in cholinergic innervation of the hippocampus in perinatally leadexposed rats,” brain res., vol. 771, no. 2, pp. 319– 328, 1997. [20] b. j. brockel and d. a. cory-slechta, “lead-induced decrements in waiting behavior : involvement of d 2 -like dopamine receptors,” science (80-. )., vol. 63, no. 3, pp. 423–434, 1999. [21] “https://www.adinstruments.com/products/labchart.” . l. shvachiyet al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-11 i-etc: isel academic journal of electronics, telecommunications and computers optics in data processing and data transmission optics in data processing and data transmission joão m. a. frazão área departamental de engenharia electrónica e telecomunicações e de computadores (adeetc), instituto superior de engenharia de lisboa (isel), lisboa, portugal e-mail: jfrazao@deetc.isel.ipl.pt abstracttoday optical systems are more and more important in data communications (optical fibers) and are also becoming important in data processing (optical and quantum computing) allowing for a fully optical communication network where all signals will be processed and transmitted in the optical domain. this paper gives an overview of optical fiber communications and analyses some optical devices and applications such as optical computing, holographic memory and optical pattern recognition. keywords: optical fibers, nlo, soliton, optical computing. 1 introduction optical technology is capable of providing the required capacity for the rapidly increasing demand in data transmission and processing. optics is of greatest importance in telecommunications due to the high bandwidth and lower attenuation obtained in optical fibers. in addition it begins to be implemented in real information processing as pattern recognition using optical computing. in future is desirable that all processes involved in data networks, such as amplification, multiplexing, demultiplexing, switching and signal processing take place in the optical domain which can be more efficient than electrical signal processing and avoid bottlenecks of electrical to optical and optical to electrical conversions [1-5]. 2 optical fiber communications in figure 1 is shown a state of the art wavelength division multiplexing (wdm) optical fiber system used for long-distance, high-bandwidth telecommunication. in the present work, the performance and limitations of the different elements that are part of this system are analysed. 2.1 optical fiber characterisation and elements in this optical wdm fiber system the emitter consists of n independent optical beams coming from n laser sources with proper i wavelength individually modulated by n electrical signals. the external modulation employing electro-optic materials is much faster than direct modulation of laser output power. the different modulated i laser beams are coupled (mixer coupler) in the same optical fiber. in long distance fibers the optical amplifier allows signals to be regenerated without the use of electro-optical converters. erbium-doped fiber amplifiers (edfa) pumped usually by diode lasers are used. in wdm or dense wavelength division multiplexing (dwdm) systems fiber bragg gratings are used to separate closely spaced wavelengths (< 0.8 nm). the elementary fiber bragg grating comprises a short section of single-mode optical fiber in which the core refractive index is modulated periodically. for optical detection the most commonly used devices are the pin or avalanche photodiodes (apd). 2.2 optical fiber limitations the most important limitations in single mode fibers are the attenuation due to material absorption, linear dispersion due to the variation of linear refractive index nl as a function of wavelength causing the pulses to broaden (limiting the overall bandwidth) and rayleigh scattering (or elastic scattering) due to random fluctuations of the refractive index on a scale smaller than the optical wavelength. all previous processes described are linear or intensity-independent, but in single mode fibers with high light intensity, due to the small cross section inside the fiber, another type of intensity-dependent processes occur. these nonlinear effects are described by nonlinear optics (nlo). in optical fibers the nlo effects can be divided in nonlinear refractive processes and inelastic scattering phenomena. i-etc: isel academic journal of electronics, telecommunications and computers cetc2016 issue, vol. 3, n. 1 (2017) id-5 http://journals.isel.pt mailto:jfrazao@deetc.isel.ipl.pt fig 1. typical wavelength-division multiplexing-fiber optic communication system. nonlinear refractive index change includes: selfphase modulation (spm) related to changes of refractive index caused by variation in signal intensity and resulting in a temporarily varying phase change that leads to additional dispersion; cross-phase modulation (cpm) related to change of refractive index of an optical beam produced by the intensity of that beam and the intensity of other beams co-propagating in the same optical fiber; four-wave mixing (fwm) process originated from 3rd order susceptibility (3)) resulting in a fourth frequency 4 related to 1, 2 and 3 frequencies which copropagate simultaneously inside a fiber by 4 = 1 ± 2 ± 3. if the light intensity in the optical fiber exceeds a certain threshold value the inelastic scattering light grows exponentially. contrary to elastic scattering, the frequency of scattered light is red-shifted during inelastic scattering and can induce stimulated effects such as stimulated brillouin-scattering (sbs) and stimulated raman-scattering (srs). fig 2. an input pulse of intensity i(z=0,t) and central angular frequency 0 travelling in the z direction in a linear anomalous dispersive, nonlinear (with nnl > 0) nondispersive and nonlinear dispersive medium, respectively. when the input pulse travels a distance z in the three different transmission mediums the output pulse exhibits different shapes. at the top output pulse spreading can be observed with higher frequencies travelling faster than lower frequencies. in the middle the output pulse is chirped with the same shape and with a negative frequency shift in the leading half of the pulse, and a positive frequency shifted in the trailing half. at the bottom, the output pulse is identical to the input pulse (optical solitons), depending on the amplitude and sign of the linear dispersion and nonlinear effects. i(t) di(t)/dt linear dispersive transmission medium nonlinear nondispersive transmission medium nonlinear dispersive transmission medium t t t t t input pulse output pulse lasern pump laser mixer coupler electrical modulation signal laser1 laser2 1+ 2+ … +n amplifier wavelength filter optical detector electrical output signal 1 2 n fiber optical detector optical detector modulator modulator modulator j. frazão et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-5 i-etc: isel academic journal of electronics, telecommunications and computers all these linear and nonlinear processes in general result in degrading the overall performance of an optical fiber telecommunication system but in certain situations can interact positively. an example is the effects of linear and spm dispersions that can be compensated mutually by proper choice of light pulse shape and amplitude as illustrated in figure 2. for a linear dispersion medium a chirp pulse with initial value c (see appendix a12) at a distance z the chirp changes to: 𝐶(𝑧) = 𝐶 + (1 + 𝐶2)𝛽2𝑧/𝑇0 2 (1) the chirp value and sign depends of c and 2 values, but even for initial unchirped pulse (c=0) the pulse will be chirped as a function of z. for a linear anomalous dispersion medium (which usually occurs in fiber optics for wavelengths in vacuum 0 > 1,312 m) the dispersion coefficient d is positive and 2 is negative the instantaneous frequency decreases linearly as function of z (appendix a13). even for an unchirped pulse (c=0) broadening is observed. thus in a pulse the higher frequencies travel faster than the lower frequencies (see top output pulse in figure 2). during propagation the pulse width t1 is a function of z given by [3]: 𝑇1 𝑇0 = [(1 + 𝑐𝛽2𝑧 𝑇0 2 ) 2 + ( 𝛽2𝑧 𝑇0 2 ) 2 ] 1/2 (2) the chirped pulse may broaden or compress depending on the sign of the product 2c. for 2c > 0 the chirped gaussian pulse broadens monotonically. for 2c < 0, the pulse width initially decreases and becomes minimum at a distance zmin after which it increases monotonically. in figure 2 consider the top output pulse as 2c > 0. for a nonlinear medium the dispersion related to spm may be understood by examining a pulse of intensity i(z,t) of carrier angular frequency 0 traveling a distance z in a nonlinear medium with refractive index 𝑛𝑒𝑓𝑓 = 𝑛𝑙 + 𝑛𝑛𝑙𝐼 (see appendix equation a9). for such pulse the argument of the electric field or instantaneous phase (see appendix equation a2) is 𝜑(𝑡) = 𝜔0𝑡 − 𝐾𝑧 = 𝜔0𝑡 − 𝑛𝑒𝑓𝑓𝐾0𝑧 = 𝜔0𝑡 − 2𝜋 𝜆0 [𝑛𝑙 + 𝑛𝑛𝑙𝐼(𝑧, 𝑡)]𝑧 (3) so that the instantaneous angular frequency is 𝜔 = 𝑑𝜑 𝑑𝑡 = 𝜔0 − 2𝜋 𝜆0 𝑛𝑛𝑙 𝜕𝐼(𝑧,𝑡) 𝜕𝑡 𝑧 (4) if 𝑛𝑛𝑙 is positive, the frequency of the trailing half of the pulse (the right half) is increased since 𝜕𝐼(𝑧,𝑡) 𝜕𝑡 < 0, whereas the frequency of the leading half (the left half) is reduced since 𝜕𝐼(𝑧,𝑡) 𝜕𝑡 > 0 as illustrated in middle output pulse of figure 2. at a certain level of intensity and for certain pulse profiles, the effects of self-phase modulation and groupvelocity dispersion are balanced so that a stable pulse travels without spread, as illustrated in bottom output pulse of figure 2. in such situation the pulse would propagate undistorted and is called optical soliton, with applications in high bandwidth optical communication systems [6]. 3 optical data processing we can divide optical computing in digital and analogue processes. digital optical computing employs optical gates and switches. the main technical difficulty remains in the creation of large high-density arrays of fast optical gates. the principle of analogue optical computing [7-10] is based in the property of the lens which perform in their back focal plane the fourier transform of a 2d image located in their front focal plane as illustrated in figure 3. an object consisting of a fine wire mesh is illuminated by a parallel coherent light beam. in the back focal plane of the imaging lens appears the fourier spectrum of the periodic mesh. by placing a narrow horizontal slit in the focal plane to pass only a single row of spectral components (horizontal pass filter) vertical frequencies are blocked and horizontal frequencies are transmitted. the corresponding image, (seen in image plane of figure 3), contains only the vertical structure of the mesh. the suppression of the horizontal structures is quite complete. j. frazão et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-5 i-etc: isel academic journal of electronics, telecommunications and computers fig 3. two dimensional optical frequency processor the inherent parallel processing is one of the key advantages of optical processing compared to electronic processing that is mostly serial. optical analogue processing is useful when the information is optical and no electronics to optical transducers are needed. in a parallel optical computer, a parallel access optical memory is required as for example 3d optical holographic memories using different materials such as photorefractive crystals. to date the optical computers were not able to compete with the electronic computers essentially due to the lack of appropriate optical components, but in the future the employment of nanotechnologies can change this situation [11-14]. references [1] f. idachaba, d. u. ike, and o. hope, proceedings of the world congress on engineering, vol i (2014). [2] s. p. singh and n. singh, progress in electromagnetics research, pier, 249-275 (2007). [3] g. p. agrawal, “fiber-optic communication systems” 4th edition, willey & sons (2010). [4] b. e. a. saleh and m. c. teich, “fundamentals of photonics” 2nd edition, wiley & sons (2007). [5] g. d. baldwin, “an introduction to nonlinear optics” plenum publishing corporation (1974). [6] s. oda and a. murata, optics express, vol. 14, 78957902 (2006). [7] j. w. goodman, “introduction to fourier optics” 3ndedition, roberts § company (2005). [8] k. preston, “coherent optical computers” mcgrawhill (1972). [9] e. n. leith, ieee journal on selected topics in quantum electronics, vol. 6, no.6, 1297–1304 (2000). [10] s. h. lee, “optical information processing fundamentals” springer (1981). [11] d. r. smith, j. b. pendry and m. c. k. wiltshire, science, 305, 788-792 (2004). [12] c. m. soukoulis and m. wegener, nature photonics 5, 523‐530 (2011). [13] j. k. gansel, et al. science 325, 1513‐1515 (2009). [14] t. utikal, m. i. stockman, a. p. heberle, m. lippitz and h. giessen, phys rev lett, 104, 113903 (2010). appendix for a nonlinear nondispersive dielectric medium the vector polarization p induced by electric dipoles satisfies the general nonlinear relation: 𝑷 = 𝜀0𝜒 (1)𝑬 + 𝜀0𝜒 (2)𝑬2 + 𝜀0𝜒 (3)𝑬3 + ⋯ (a1) where 0 is the permittivity of vacuum and  (i) ( i = 1, 2, ...) is ith order tensor susceptibility. for isotropic medium, as in optical fibers, we can use scalar notation instead of vector notations because the polarization vector p has the same direction of the electrical field e. for an electrical field associated to a plane wave propagated in z direction 𝐸 = 𝐸0cos (𝛽𝑧 − 𝜔𝑡), (a2) where  is the phase constant, the polarization p becomes 𝑃 = 1 2 𝜀0𝜒 (2)𝐸0 2 + 𝜀0𝐸0 [𝜒 (1) + 3 4 𝜒(3)𝐸0 2] cos(𝛽𝑧 − 𝜔𝑡) + 1 2 𝜀0𝜒 (2)𝐸0 2cos [2(𝛽𝑧 − 𝜔𝑡)] + 1 4 𝜀0𝜒 (3)𝐸0 3cos [3(𝛽𝑧 − 𝜔𝑡)] + ⋯ (a3) f(x,y) f(x2,y2) f(u,v) f f f f ima ge pla neobject pla ne spa tia l frequency pla ne spa tia l frequency filter y2 x2 y1 x1 v u j. frazão et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-5 i-etc: isel academic journal of electronics, telecommunications and computers in the above equation the first term is constant and gives a constant field in the medium. the second third and fourth terms correspond to oscillating frequencies , 2 and 3 respectively known as fundamental, second, third harmonics of polarization. as silica used in optical fibers consists of symmetric molecules, (2) vanishes, and neglecting 3 term due to variation in refractive index of the fiber inducing a phase mismatch between frequencies  and 3 equation (a3) becomes 𝑃 = 𝜀𝑜𝜒 (1)𝐸0 cos(𝛽𝑧 − 𝜔𝑡) + 3 4 𝜀𝑜𝜒 (3)𝐸0 3 cos(𝛽𝑧 − 𝜔𝑡) (a4) where (i) terms are neglected for i > 3. for a plane wave propagating in a dielectric linear isotropic and homogeneous the intensity (i) is 𝐼 = 1 2 𝑐𝜖0𝑛𝑙𝐸0 2 (a5) where c is velocity of light in vacuum and nl is the linear refractive index of the medium. therefore, 𝑃 = 𝜀0 [𝜒 (1) + 3 2 𝜒(3) 𝑐𝜀0𝑛𝑙 𝐼] 𝐸0cos (𝛽𝑧 − 𝜔𝑡) (a6) defining the effective susceptibility (eff) of the medium as 𝜒𝑒𝑓𝑓 = 𝑃 𝜀0𝐸 = 𝜒(1) + 3 2 𝜒(3) 𝑐𝜀0𝑛𝑙 𝐼 . (a7) hence, effective refractive index (neff) can be written as 𝑛𝑒𝑓𝑓 = (1 + 𝜒𝑒𝑓𝑓) 1/2 = [1 + 𝜒(1) + 3 2 𝜒(3) 𝑐𝜀0𝑛𝑙 𝐼] 1/2 = 𝑛𝑙 [1 + 3 2 𝜒(3) 𝑐𝜀0𝑛𝑙 3 𝐼] 1/2 , (a8) where 𝑛𝑙 = (1 + 𝜒 (1)) 1/2 is the linear refractive index. in equation (a8) the last term in parenthesis is usually very small compared to unity, so neff can be approximated by the first term of the taylor’s series expansion as 𝑛𝑒𝑓𝑓 = 𝑛𝑙 + 3 4 𝜒(3) 𝑐𝜀0𝑛𝑙 2 𝐼 = 𝑛𝑙 + 𝑛𝑛𝑙𝐼, (a9) where 𝑛𝑛𝑙 = 3 4 𝜒(3) 𝑐𝜀0𝑛𝑙 2 is the nonlinear refractive index. for a linear dispersive medium nl (for simplicity we replace nl by n) is a function of angular frequency . for pulses with spectral width  much smaller than the carrier frequency 0 ( << 0) the propagation constant () = n()/c can be expanded in a taylor series around the carrier frequency: 𝛽(𝜔) ≅ 𝛽0 + 𝛽1(∆𝜔) + 𝛽2 2 (∆𝜔)2 (a10) in the above equation ∆𝜔 = 𝜔 − 𝜔0 , 𝛽1 = ( 𝑑𝛽 𝑑𝜔 ) 𝜔=𝜔0 = 1 𝑣𝑔 , 𝛽2 = ( 𝑑2𝛽 𝑑𝜔2 ) 𝜔=𝜔0 = − 𝐷𝜆0 2 2𝜋𝑐 (a11) where vg is the group velocity and d is the dispersion parameter. let us consider the propagation in z direction in a linear dispersive medium of a frequency modulated gaussian pulse (chirped pulse) with an initial electric field (at z = 0) 𝐸(0, 𝑡) = 𝐸0𝑒𝑥𝑝 [− 1 2 ( 𝑡 𝑇0 ) 2 ] × 𝑒𝑥𝑝 [−𝑖 𝐶 2 ( 𝑡 𝑇0 ) 2 ] 𝑒𝑥𝑝[−𝑖𝜔0𝑡] (a12) where e0 is the amplitude, t0 the half width at 1/e intensity point, c parameter that control the frequency chirp and 0 the carrier frequency. the instantaneous frequency is the derivative of the phase, 𝜔 = 𝑑 𝑑𝑡 (𝜔𝑜𝑡 + 𝐶 2𝑇0 2 𝑡 2) = 𝜔0 + 𝐶 𝑇0 2 𝑡 (a13) and is a linear function of time. the electric field at z position e(z,t) is from [3] 𝐸(𝑧, 𝑡) = 𝐸0 √𝑄(𝑧) 𝑒𝑥𝑝 [ − (1 + 𝑖𝐶) (𝑡 − 𝑧 𝑣𝑔 ) 2 2𝑇0 2𝑄(𝑧) ] × 𝑒𝑥𝑝[𝑖(𝛽0𝑧 − 𝜔0𝑡)] (a14) where 𝑄(𝑧) = 1 + (𝐶 − 𝑖)𝛽2𝑧/𝑇0 2. this equation shows that a gaussian pulse remains gaussian on propagation but its width, chirp, and amplitude changes continuously. j. frazão et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-5 i-etc: isel academic journal of electronics, telecommunications and computers room temperature photoluminescence and photoconductivity of wet chemical deposited zno nanowires used for solar cells room temperature photoluminescence and photoconductivity of wet chemical deposited zno nanowires used for solar cells s.r. bhattacharyya, r. ayouchi, j. pereira and r. schwarz department of physics & icems, instituto superior técnico, lisbon, portugal soumya.bhattacharyya@ist.utl.pt keywords: zno nanowires, wet chemical process, transport properties, tpc, pl. abstract: zno 1-d nanostructures (nanowires) were deposited by a two-step wet chemical process. the dimensions of wires were about 100 nm 1100 nm in length and about 20 120 nm in diameter. scanning electron microscopy (sem) and x-ray diffraction (xrd) technique was used to obtain the microstructural information from the films. the nanowire films were also characterized optically by transmittance measurement and room temperature photoluminescence (pl) measurements. the transport properties of the samples were characterized by performing transient photoconductivity (tpc) experiments. 1 introduction in the attempts to tap the unconventional energy resources by utilization of solar emission as a clean source of energy, during the last 30 years, there has been a steady progress from silicon-based solar cells to quarternary chalcogenide based compound semiconductor solar cells [1]. but the main bottleneck in this regard has been its efficiency [1], which has seldom crossed the double-digit figure and has thus proved detrimental for adaptation by industry for wider marketing. the latest attempt to break this throttle in regards to solar cell deficiencies has been the utilization of organics and dyes for fabrication of cheap and efficient organic solar cells and dye sensitized solar cells (dssc). along with the organic solar cell technology [2], the dssc research also seems promising [3]. one of the main components of a dssc is nanoparticles of a wide direct band gap semiconductor, like titanium oxide (tio2) or zinc oxide (zno) for enhanced light capture and scattering within the cell to increase its efficiency. in conventional nanoparticle-based dsscs, the electrons move through the electrolyte to reach the electrode passing within the nanoparticles by percolation or hopping mechanism. with each hop, the electron can recombine with the electrolyte. under this condition, the diffusion is slow and thus the efficiency of the cell is limited by the recombination. however, zno nanowire dsscs provide a direct path to the anode, which increases the diffusion rate, hence its mobility, without increasing the recombination rate; the other advantages of using nanowires being lower reflective losses and intensive light trapping within the solar cell structure [4]. this could prove important for increasing the solar cell efficiency. but in order to put it into optoelectronic device application, it is imperative to have clear idea about the optical and transport properties of the nanowires. apart from its application in dssc, zno nanowires has demonstrated potentiality in the field of energy harvesting through harnessing its piezoelectric propertyto convert mechanical energy to electrical energy. this constitutes an alternative application of semiconducting material for energy production and is currently a very hot topic of research [5,6]. in this communication we present our results on the microstructural, optical and transport i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-10 http://journals.isel.pt/index.php/iajetc properties of the zno nanowires grown by two-step wet chemical technique. 2 experimental procedure the zno nanowire films were deposited by a simple two-step wet chemical technique on corning glass substrates. the initial seeding layer with a thickness of about 50 nm was deposited onto the glass substrate by pulsed laser deposition (pld) technique using solid zno (purity 99.999%) target. for the deposition, the 2 nd harmonic line at 532 nm of a q-switched nd:yag pulsed laser was used. the 5 ns duration pulses were shot at the target at the rate of 5 hz. the plasma plume formed by the laser hitting the target produced a homogeneous and consistent film with the particle size of about 10 nm, at a deposition rate of about 2 nm/min. prior to deposition; the system was evacuated to a base pressure of 10 -7 mbar. the deposition was carried out in oxygen atmosphere with a system pressure of 10 -2 mbar. the substrate temperature was kept at 673 k and the target to substrate distance was fixed at 5 cm. the as-deposited films were used as seeding layer for the wet chemical deposition of nanowires. the chemical deposition of nanowires on top of the seed layer was carried out in aqueous medium using varying ratio r (1:10, 1:20, 1:30, 1:35 and 1:40) of zinc nitrate (zn(no3)2.6h2o) as zinc containing precursor and sodium hydroxide (naoh). all the chemicals were reagent grade. the depositions were carried out for 90 minutes at 333 k in a magnetic stirrer at a stirring speed of 500 rpm to obtain films with nanowires of different aspect ratio. for the depositions, the samples were immersed vertically into a 100 ml beaker containing 40 ml of aqueous solution with the two precursors mixed proportionately. after the deposition, the substrates were removed and deionized distilled water was trickled from the back side to wash away any excess residues and the films were subsequently dried in air at 350 k for 15 minutes before being utilized for further characterizations. the microstructure of these nanowires were studied by scanning electron microscopy (sem), while the crystal orientation and structure was verified from x-ray diffraction (xrd) scans in the θ-2θ mode. the optical transmittance and reflectance for the films was measured at room temperature using a jobin-yvon uv-vis-nir spectrophotometer. additionally, steady-state pl was performed on these samples at room temperature (300 k) to obtain defect related information. the pl measurement was carried out using a 325 nm hecd laser (10 mw power) to excite the zno samples. photoluminescence from the sample was directed into a monochromator and collected by a photomultiplier for the preset wavelength range. the signal was then fed into a dual-phase stanford research sr830 lock-in amplifier, which was interfaced with the computer by labview 8 for automatic data acquisition. a chopper was placed in the beam path and set at 19 hz frequency to produce a chopped (a.c.) signal. a 475 nm cut-off filter was used to suppress the first and the second order laser line from the spectra. the transport properties of the films were investigated by transient photocurrent measurements. for the transient photocurrent-response (tpc), a 266 nm uv line of q-switched nd:yag laser, shooting 5 ns pulses at ~ 2 mj/cm 2 , with a repetition rate of 5 pulses/second and a spot size of 1 mm diameter was used. the measurements were carried out with coplanar aluminium contacts deposited on the films. the current generated was dropped along a variable shunt resistance ranging from 50 ω to 1 mω, as selected for different time windows, and the corresponding signal was measured with a tektronix 200 mhz bandwidth digital oscilloscope. it was later converted into current for further calculations, with proper zero corrections. the applied voltage on the samples was 15 v and the photocurrent decay was measured in several time windows ranging from 500 ns to 100 µ s. below the 500 ns time scale, the decay signal was too weak to capture. 3 results and discussions the nanostructure of the zno layers was studied using a jeol 7001f sem. figs. 1 (a-e) show the representative sem micrographs of the zno nanostructures for the films z1, z2, z3, z4 and z5 deposited with r as 1:10, 1:20, 1:30, 1:35 and 1:40, respectively. the sample z1 deposited with r = 1:10 showed very nascent tendency for 1-d growth, with the morphology representing dwarf stubs, with arrested vertical growth. also for the sample z2, the nanostructures formed aggregated clusters with nanoflower like feature, with each of the nanostructures constituting the “petals” which were broader at the centre and pointy towards the tips. they showed arrested vertical elongation. but for samples z3, z4 and z5, definite 1-d growth in the vertical direction was observed with their lengths (in these cases) being much larger than their diameters. since our aim was growth of vertically s.r. bhattacharyya et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-10 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc aligned nanowires only the samples z3, z4 and z5 was utilized for pl and tpc experiments. the lengths of the nanowires were measured laterally, across the crack lines by tilting the sample and using the built-in system of the sem unit to obtain the data. z4 had the highest nanowire length (l) of ~ 1100 nm, followed by z5 with 750 nm and finally z3 had the nanowire length of ~ 600 nm. z5 had the diameter, d ~ 40 nm, followed by z4 with ~ 33 nm and by z3 with ~ 24 nm. for the sample z1, the length of the nanostructures were around l ~ 100 nm, with a diameter of d ~ 60 nm; whereas the nanoflower-like structures in z2 consisted of petals of length l ~ 300 nm, with a diameter at the centre of d ~ 120 nm. all the lengths and diameters stated here are representative average lengths and diameters. it was observed that the nanowires were more aligned towards the vertical in the case of z4. the tips of the nanowires were found to cluster together, as compared to the other samples where they were more oriented randomly towards the vertical. sem was also performed on the seeding layer. it exhibited a smoothly grained, compact surface with an average particle size of ~ 10 nm, with the exception of few larger particulates (30 40 nm) sitting on top of the surface. this might have been some larger particles sputtered from the target due to laser pulsing and also agglomerates of the smaller particles at the surface. the xrd for a representative nanowire film (fig. 2) indicated a sharp, intense peak due to (002) plane of hexagonal zno along with other very small peaks corresponding to (100), (101) and (103) planes of hexagonal zno (icdd card no. 361451). the xrd thus confirmed highly c-axis oriented growth. the same c-axis orientation was followed for all the other samples. fig. 1 representative sem images of asdeposited zno nanostructures: (a) z1, (b) z2, (c) z3, (d) z4 and (e) z5 film. s.r. bhattacharyya et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-10 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc the band gap of the films were determined from the transmittance spectra (not shown here) by calculating the optical absorption coefficient [7], and was found to be between 3.2 ev to 3.3 ev for all the samples z1 to z5. they did not show too much variation for the five samples. the pl measurements were performed at 300 k (room temperature). our experimental setup was not sensitive enough to detect possible band edge luminescence (when measured without the 475 nm filter), whose intensity for our samples should be highly quenched due to their very small diameter. fig. 3 shows the pl spectra for the samples z3, z4 and z5 at 300 k. all the spectra were normalized to the highest intensity value. the spectra yielded broad yellow-orange luminescence peaks centred at ~ 2.1 ev. the sample z4 having highest surface to volume ratio (aspect ratio), thereby highest density of surface defects, as compared to z3 and z5, had the yellow-orange band of highest intensity, followed by z3 and z5, in the order of their aspect ratios. the spectra for z3 and z5 seemed to be blue shifted with high asymmetry with a clear indication of a second peak at lower wavelengths. when deconvoluted, the spectrum for z3 at 300 k revealed a peak at ~ 2.49 ev along with the peak ~ 2.08 ev, and the spectrum for z5 also revealed a second peak at ~ 2.44 ev, along with a ~ 2.10 ev emission which is comparable to z3. the blue shift in emission peaks (as a whole) for the z3 and z5 may thus be due to the appearance of these additional higher energy peaks (2.4 2.5 ev), which also pulled the centre towards higher energy. the green band (around 2.4 2.5 ev) may be ascribed to the recombination of electrons with holes trapped in singly ionized oxygen vacancies (vo + ) whereas the orange emission (around 2.1 ev) might be due to interstitial oxygen ions (oi ) [8]. tpc measurements were performed on the three samples z3, z4 and z5, with vertically aligned nanowire structures. the sample z4 was extremely highly resistive and did not yield any response at 15 v applied between the contacts. for the samples z3 and z5 decays could be measured in several time windows. it was found for both films, the decay had slow non-exponential power-law behaviour over several magnitudes of time. fig. 4 shows a log-log plot of transient photocurrent decay vs. time for sample z3. typically, for room temperature tpc measurements in wide bandgap semiconductors, having high band tailing or a large density of defects within the band gap, a sheet of charge is created on one side of the sample by strongly absorbed light. charges of a given polarity drift through the sample to the opposite or adjacent collecting contact due to an external electric field. during the transit, carriers are frequently trapped in localized states below the band edges where they are immobile until reemitted to the bands. some carriers, however, might disappear via recombination. under the condition that carrier trapping is dominant, the photocurrent decay for these samples follow slow power law behaviour over several scales of time. this is the underlying theory of multiple trapping (mt) model [9], in disordered materials. in our case, the application of power-law equation for the photocurrent decay, i = i0t α yield α ~ -0.40 for z3 and α ~ -0.63 for z5, respectively, from the log-log plot of the photocurrent decay vs. time. in other words, the decay corresponding to z3 is slower than z5, indicating larger density of surface states and defects within the forbidden gap, which contributes to the slower carrier mobility decay and hence a slower photo-response. the decay from sample z5 contained a stronger contribution from recombination reflected by a faster decay process. fig. 3 representative pl spectra at 300 k. fig. 2 xrd of representative zno nanowire film. s.r. bhattacharyya et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-10 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 4 conclusion zno nanowire thin films were deposited by a seeding layer mediated wet chemical process. the initial seeding layer with grain size ~ 10 nm was grown on glass substrate by pld technique, upon which the zno nanostructures with diameters varying from 20 120 nm and lengths from 100 1100 nm were grown in aqueous medium using a magnetic stirrer at 333 k. the pl of the samples indicated broad yellow-orange and green emission bands. the green band may have originated from oxygen vacancy related transitions, while the orange band arises from transitions due to interstitial oxygen ions. the large visible pl peaks and the absence of band-edge luminescence, which might be due to the size effect of the nanowires, indicate that the samples are highly defective with large density of surface states forming broad band tails within the forbidden gap, which governs the pl process for these samples. transient photocurrent decay measurements for the samples indicated slow, non-exponential, power law type decays over several magnitudes of time, characteristic of multiple trapping of carriers by the deep defects within the bandgap. acknowledgement one of the authors srb wishes to thank fct for granting a postdoctoral fellowship. he also wishes to express his gratitude towards s.v. college, west bengal, india for their encouragement and support. jp and ra wish to thank fct for grant of their fellowships. the work benefitted from funding of projects ptdc/fis/108025/2008 and ptdc/ctmcer/115085/2009. references [1] green m.a., 2007. thin film solar cells: review of materials, technologies and commercial status, j. mater. sci.: mater. electron., vol. 18, no. 1, p. 15-19. [2] yeh n., yeh p., 2013. organic solar cells: their developments and potentials, renew. sust. energ. rev., vol. 21, p. 421-431. [3] arjunan t.v., senthil t.s., 2013. review: dye sensitized solar cells, mater. technol: adv. performance mater., vol. 28, no. 1-2, p. 9-14. [4] zhang q., yodyingyong s., xi j., myers d., cao g., 2012. oxide nanowires for solar cell applications, nanoscale, vol. 4, no. 5, p. 1436-1445. [5] kumar b., kim s.-w., 2012. energy harvesting based on semiconducting piezoelectric zno nanostructures, nano energy, vol. 1, no. 3, p. 342-355. [6] briscoe j., bilotti e., dunn s., 2012. measured efficiency of a zno nanostructured diode piezoelectric energy harvesting device, appl. phys. lett., vol. 101, no. 9, p. 093902. [7] bhattacharyya s.r., majumder s., 2010, synthesis of al doped zno films by sol-gel technique, funct. mater. lett. vol. 3, no. 2, p. 111-114. [8] greene l.e., law m., goldberger j., kim f., johnson j.c., zhang y., saykally r.j., yang p., 2003. lowtemperature wafer-scale production of zno nanowire arrays, angew. chem. int. ed., vol. 42, no. 26, p. 3031-3034. [9] tiedje t., rose a., 1981. a physical interpretation of dispersive transport in disordered semiconductors, solid state commun., vol. 37, no. 1, p. 49-52. fig. 4 representative log-log plot of tpc for z3. s.r. bhattacharyya et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-10 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc social mobility advisor social mobility advisor andré costa, david antunes and porfírio filipe área departamental de engenharia de electrónica e telecomunicações e de computadores instituto superior de engenharia de lisboa, lisbon, portugal {31250, 31212}.alunos.isel.pt, pfilipe@deetc.isel.ipl.pt key words: sustainable m obility , social network, rep utation sy stems, recommender sy stems, incremental collaborative filtering, scalability . abstract: seamless access to up dated information about mobility op tions is a key factor for sustainable economic develop ment. in this context, the adop tion of information sy stems, with innovative features, meet these needs by stimulating the interest and curiosity of end users. there are well known difficulties, even for regular travelers, to get the most suitable mobility op tions combining available transp ort modes. also, the sustainable mobility has become the new imp erative for transp ort p olicy . in this p ap er we p resent a collaborative p latform that sup p orts a social network, in which users can search and share sustainable mobility exp eriences. this p latform includes a rep utation and a recommender sy stems sp ecially designed to deal with mobility op tions. 1 introduction we live in a society of information. sharing and accessing that information has become an acquired right. the internet has become a great ally in the dissemination of information process, partly due to the popularity of social networks and its collaborative based interaction. currently, there are successful examples of websites that operate either under the social network basis or by taking advantage of other social networks (e.g., authentication, friend list invitations, etc.). trip advisor (www.tripadvisor.com), rail europe (http://www raileurope.com) and momondo (http://www momondo.pt/) are some examples in the mobility context, which try to provide mobility options. however, these services are unable to indicate alternative means of transportation; also, the results do not reflect the personal tastes of those who use the service. therefore, we identify the need for a system that is able to provide multimodal mobility options , and is built under each user’s tastes and interests in order to generate relevant search results and recommendations. in this paper we present a collaborative solution that aims to assist travelers to share their experiences and find the best suitable mobility alternatives, bearing in mind the best sustainable options. those actions are supported by the collaborative community interactions , in other words , under the social network concept. this approach leads to a credibility problem which we had into account by assuring both users and information are cataloged by confidence levels. we also propose an automated moderation mechanism that tries to reduce the number of required staff to administer the system. the social mobility advisor (sma) project aims to contribute for a more accessible, social and sustainable mobility across the europe, dealing with reputation and recommender systems in order to build a social network. 2 concept social networks (sn) have evolved over time as a way of information exchange, even before the existence of the internet. in fact, already in ancient egypt there were groups of people with common interests who came together to share knowledge. with the emerging of internet, information share and access has become more simple, fast and effective. therefore, sn naturally shifted to this environment, grouping a greater number of people, breaking down physical, geographical and cultural barriers. sma takes advantage of collaborative approach of social networks , by supplying a way for users to share their mobility experiences where the participating entities collaborate actively to improve the quality of information, updating information items, for example, targets, gates, and operators / transport authorities. thus, we can incrementally gather and provide mobility experiences, decreasing information disparity problem on this matter, that daily harms regular travelers. i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-3 http://journals.isel.pt/index.php/iajetc due to the system’s collaborative way, on which information deeply depends on users’ interactions, we identify the lack of confidence as a major problem. this problem is commonly addressed by commercial companies through the use of reputation systems. we managed to combine existing reputation systems tendencies (e.g., amazon, ebay, slashdot) in order to minimize this problem. with the aim of minimize the need for dedicated human resources , especially to maintenance, the sma social network tends to be self-managed by the participants which are dynamically assigned, with more or less privileges, depending on their current reputation. so far we have described a process that can be used to populate the system with filtered information. despite that, a user may ask: “which information is relevant for me?”. we consider that this question cannot be left unanswered. our goal here is to provide relevant information for each user, rather than dump random (and possible undesirable) information for every users. recommender systems help to increase the effectiveness and capability of recommendation generation, which will also highly increase the computation weight. the underlying techniques we present try to overcome the scalability problem, by computing heavy calculations into simpler incremental summations. the information available in the sma social network is the result of the integration of information from internal sources (internal repository with mobility options maintained by the participants) as well as information provided by external sources, such as wik ipedia (http://www.wikipedia.org/), google maps, google places (http://code.google.com), facebook (http://www facebook.com/) and brighter planet (http://www.brighterplanet.com/). 3 reputation system the traditional cues of trust and reputation that we are used to observe and depend on in physical world are missing in online environments, so that electronic substitutes are needed. a reputation system collects, distributes, and aggregates feedback about participants’ past behavior, helping users to decide whom to trust and encouraging trustworthy behavior [1, 3]. we assumed that such mechanism can be adopted in our context in order to overco me the confidence problem. the proposed reputation system is based on the combination of existing ones, such as amazon (http://www.amazon.com/), stack overflow (http://stackoverflow.com/) and ebay (http://www.ebay.com/). also, we grouped users by credibility, assigning them reputation levels. 3.1 reputation level the reputation levels system is inspired in videogames like world of warcraft (http://eu.battle net), which needs exponential requirements to level up. equation 1 illustrates the adopted formula to generate reputation level requirements for the next level, supplying the current level. ( ) ( ( ) equation 1: rep utation level requirements. the reputation level is an important issue in our reputation system, since it will define the weight of the user’s ratings, described below, as well as their role/privileges in social network. 3.2 ratings our system provides an input for users to express their opinion related to item’s information quality by rating it as being helpful or not helpful (+1, -1), as showed in illustration 1. this range permits a less ambiguous evaluation, since it will influence the reputation of those who submitted the item. illustration 1: binary rating rep resentation. using this schema, is possible to filter and organize information by its quality and, at the same time, indirectly rate the user who submitted that information. this approach aims to decrease public reputation of those who submit poor information and, inversely, reward who submit relevant information. item’s information reputation is determined using the equation 2, which consists on a summation of all positive and negative ratings weighted by rater’s reputation level. rater’s reputation level is used to avoid unfair ratings [1], one of reputation systems known problem, in assumption that users with better reputation provide more reliable rat ings. ( ) equation 2: binary rating calculation. – item current rating (in cache); a. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc – summation of item raters reputation (in cache); – new rating for item of active user; – item active user reputation level (in cache); reevaluation factors: – old rating for item of active user; – reputation level of active user when submitted the old evaluation (in cache); in order to calculate the mobility item’s quality itself (e.g., how good is a gate, operator, transport, etc.), our system also provides an input for users to express their mobility experience quality by a quantitative rating from 1 to 5 stars, as illustrated in illustration 2, plus comment. illustration 2: star rating rep resentation. item’s quality reputation is calculated by the equation 3 and it consists in a weighted average between evaluations and raters reputation. rater’s reputation level is used to avoid unfair ratings as mentioned before. ( ) equation 3: start rating calculation. – item current rating (in cache); – summation of item raters reputation (in cache); – new rating for item of active user; – item active user reputation level (in cache); reevaluation factors: – old rating for item of active user; – reputation level of active user when submitted the old evaluation (in cache); 3.3 user’s reputation as mentioned before, users indirectly rate others by rating their submitted items as being helpful or not helpful, according to the equation 4. ( ) equation 4: user’s rep utation adjustment . – reputation increment; – rater reputation level; – new rating (+1, -1); reevaluation factors : – old rater reputation level; – old rating; for example, when a user a submits an item and a second user b on level 2 rates that item as useful (+1), the user a gets two points of reputation as presented in illustration 3. illustration 3: evaluation examp le. 3.3 interaction incentives due to the lack of incentive for users to provide their mobility experiences and ratings, we included incentive mechanisms in our reputation system. users are rewarded for data submission, where the importance of the submitted item defines the assigned reputation points. table 1 shows the possible configuration values of rewarded points. action reputation binary rating +2 star rating +1 mobility item submission, except route +4 route submission +6 comment submission +1 collaboration action +3 revaluation 0 table 1: possible configuration for incentives. for example, someone that submits a new gate will be rewarded with 4 reputation points . we also adopted a badge reward system, used by systems like stack overflow (http://www.stackoverflow.com/) and foursquare (https://www foursquare.com/) with great acceptance by the community, where users are rewarded by reaching certain objectives . in our context, we reward the users when they reach, for example, a pre-defined number of done routes or rated items . a. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 3.4 moderation in our approach, the information is strongly dependent on user’s collaborative interactions consequently, moderation actions are needed . to reduce the amount of necessary staff to keep the system, we included a role based moderation mechanism, inspired on slashdot, which takes advantage of user’s reputation to delegate them moderation actions . this way, most reliable users are assigned to perform some moderation actions. our mechanism has two moderation roles, where the first, called “collaborator”, are assigned to manage regular users by editing/hiding comments and item’s information. the second role, called “moderator”, was introduced to reduce the number of unfair collaborators, as described by slashdot [1]. 4 recommender system in everyday life, we are confronted with recommendations from other people, television, ads, news reports from news media, and so on. in a web environment, recommender systems can help us sift through all the available information to find which is most valuable for us, accordingly with our tastes and interests. e-commerce companies rely on recommender systems to ensure that customers are recommended the right products, improving sales. [4] it is natural to assume that these mechanisms can be introduced in order to assist travelers to find relevant mobility alternatives. in our study, collaborative filtering (cf) is described as one of the most successfully technique to generate relevant recommendations. the fundamental assumption of cf is that if users x and y rate n items similarly, or have similar behaviors (e.g., buying, watching, listening), and hence will rate or act on other items similarly. the cf algorithms can be divided in three main parts, listed below:  representation of input data: deals with the submitted ratings data, which is represented by a user-item matrix.  neighborhood formation: is the most important step. it focuses on the problem of how to identify the most similar users (neighbors) with active user . the most commonly used technique is the pearson’s correlation, that measures the extent to which two variables linearly relate with each other [4], described in equation 5 ∑ ( ̅̅̅̅ ) ( ̅̅ ̅̅ ) √∑ ( ̅̅̅̅ ) √∑ ( ̅̅ ̅̅ ) equation 5: pearson’s correlation calculation. where the ∈ summations are over the items that both the users and have rated and ̅ is the average rating of the co-rated items of the th user.  recommendation generation: deals with the problem of finding the top-n recommended products from the neighborhood users. a cf algorithm should be both accurate (the recommended objects should subsequently receive high ratings) and efficient in terms of computational complexity. therefore, accomplishing that is quite an engineering challenge. cf fails to scale up its computation with the growth of both the number of users and items in the database. [2] this is mainly caused by the neighborhood formation task, which performs heavy computations in order to calculate the similarity for each pair of users. we studied “classic” collaborative filtering techniques and realized that real-time recommendations, as expected in a web application, would tend to generate unacceptable performance values. in order to address the scalability challenge, we included in our project an incremental collaborative filtering (icf) based algorithm which allows generating recommendations in a reasonable time. 4.1 incremental cf to address scalability problem, we adopted the proposal presented by papagelis et al [2]. it consists on an incremental user-based cf method, based on incremental updates of the user-to-user similarity matrix, defined in step two (neighborhood formation). the adopted mechanism to calculate the user similarity is the pearson’s correlation, described in equation 5. in order to update users’ similarity, based on previous values , we adopted the notation described on equation 6. a. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc √ √ ∑ ( ̅̅ ̅̅ ) ( ̅̅ ̅̅ ) ∑( ̅̅̅̅ ) ∑ ( ̅̅ ̅̅ ) equation 6: , , and definition factors. , and are updated every time a user submits a new rating or updates an existing one. the new similarity calculation is based on previous factor’s values. √ √ √ √ equation 7: new similarity calculation formula. the new similarity value is calculated based on , and factors generated by adding and increments on , and values , as described in equation 7. the formulas to calculate these increments are described in table 2. table 2: e, f and g calculation formulas. we managed to build a recommendation algorithm which is able to operate under an online environment. the incremental calculation used in neighbourhood formation returns the same values as the “classical” technique. 4 conclusion this paper describes a collaborative platform for sharing and getting sustainable mobility options by taking advantage of emerging concepts, like social network. we developed a platform that offers to its users the possibility of information sharing focused on user’s needs. in this perspective, users can interact and collaborate within a virtual community supported by the world wide web. therefore, users are involved in a media dialogue, as both producers and consumers of information, which contrasts with the access through a portal (website) where they are limited to being passive consumers of content specifically created for them. the implemented system manages to organize information by its credibility and also assign to most reputable users moderation actions , through the adopted reputation system. the submitted information is weighted in order to organize it by its relevance and the users are cataloged by reputation levels. therefore, we can decrease information disparity. currently, the sma social network has exceeded the testing phase, which counted with more than one hundred users. the present database includes more than ten thousand items (gates, transports, operators), covering mobility information in a worldwide context, and about one thousand rates and comments. as future work, we consider relevant to embrace mechanisms that allow transportation authorities to interact directly with the system in order to keep updated information about mobility items provided by them, ticket pricing and departure/arrival schedules, converging our solution in a door-to-door scenario. also, these authorities could be part of incentive mechanism, offering real prizes (discounts, vouchers) for those who best contribute to system’s information quality. everyone can obtain more information and join the social network accessing http://start.isel.pt. references [1] audun jøsang, roslan ismail and colin boy d. a survey of trust and reputation systems for online service provision. (the netherlands 2007), elsevier science publishers b. v. amsterdam. [2] m . pap agelis, i. rousidis, d. plexousakis, e. theoharop oulos, incremental collaborative filtering for highly-scalable recommendation algorithms, m .-s. hacid, n. v. m urray , z. w. ras, and s. tsumoto, editors, ism is, volume 3488, lecture notes in comp uter science, 2005. [3] p. resnick, r. zeckhauser, e. friedman and ko kuwabara, reputation systems: facilitating trust in internet interactions, communications of the acm , 43, 2000. [4] b. m . sarwar, g. kary p is, j. a. konstan and j. riedl. analysis of recommendation algorithms for ecommerce, new york, usa, 2000, acm . a. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc benchtop ss-oct – layout and performance evaluation benchtop ss-oct – layout and performance evaluation josé p. dominguesa,b, susana f. silvab, rui bernardesb and a.m. morgadoa,b a department of physics, university of coimbra, coimbra, portugal bibili, faculty of medicine, university of coimbra, coimbra, portugal jpd@uc.pt susana.f.s.89@gmail.com rmbernardes@fmed.uc.pt miguel@fis.uc.pt abstract — optical coherence tomography (oct) is a noninvasive biomedical imaging technique that provides high speed and high resolution three dimensional and cross sectional images of biological samples, in vivo and in situ. oct applications targeting small animals is believed to bring developments in medical techniques, instruments, diagnosis and therapies for a number of human diseases as always have been the case of animal experimentation. with the swept source oct (ss-oct) system presented in this work, we were able to achieve performance parameters that meet the requirements to image the retina of small animals. performance characteristics include 105 db for system sensitivity, a roll-off below 1 db/mm over 3 mm depth and an axial resolution of 8 µm. we describe the layout and acquisition/processing solutions towards fast imaging of in vivo samples. keywords— swept source oct, imaging, instrumentation. i. introduction since its first appearance in the early 1990’s, optical coherence tomography (oct) was recognized as a very helpful tool for ophthalmology. based on the optical interference phenomenon, the oct is capable of producing high-resolution cross-sectional images of non-homogeneous samples, among which the biological tissue, being particularly tailored to image ocular structures such as the retina. over the last twenty years, oct has experienced a fast and steady growth. it now offers a wide range of applications, is clinically very well accepted and is the subject of active research worldwide. the technique and theory behind it – in all possible configurations, being it time-domain oct (td-oct), spectral-domain oct (sd-oct) or swept-source oct (ssoct) – is extensively described in the literature [1][2][3][4]. ss-oct has been appointed as the most promising technology for oct imaging by offering higher scanning speed, reduced sensitivity roll-off and better overall performance, when associated with balanced detection,[3] in comparison to time and fourier/spectral domain approaches. oct systems specifically developed to image small animals, used as physiological models of disease, are fundamental to test and develop new medical therapies. our group is engaged in the development of a dedicated ss-oct platform for small animals based on the most recent technological advances. it should allow testing different configurations and explore new concepts. the basic layout has already been presented [5] along with simulations and preliminary performance evaluation. in this paper, we report recent developments regarding in vitro performance and discuss key details with respect to acquisition/scanning synchronism. the motivation for this development is definitely the possibility of building a valuable tool for research using animal models of disease. in the field of biomedical research, small animals are very often used to develop, validate and test new techniques and therapies. oct imaging can provide researchers with the means to understand physiology, pathology and phenotypes in health and disease. furthermore, this benchtop oct system can be used as a platform for the continuing development and test of new oct components, instrumentation and methods. ii. materials and methods in its simpler formulation, the oct can be considered as an open-air michelson interferometer (fig 1) with a beamsplitter, reference and sample paths and a gaussian-shaped broadband source. fig. 1 michelson interferometer schematics based oct. light coming from the source can be represented by its electric field wave component ein expressed as a complex exponential [1][2]: 𝐸𝑖𝑛 = 𝑠(𝑤)𝑒 𝑖(𝑤𝑡−𝑘𝑧) (1) i-etc: isel academic journal of electronics, telecommunications and computers cetc2016 issue, vol. 3, n. 1 (2017) id-8 http://journals.isel.pt mailto:jpd@uc.pt mailto:susana.f.s.89@gmail.com file:///c:/users/jpd/documents/rmbernardes@fmed.uc.pt file:///c:/users/jpd/documents/miguel@fis.uc.pt where s(w) is the source amplitude spectrum, w the angular frequency, k the wavenumber and z the distance along the propagation direction. phase appears due to different interactions throughout the interferometer both for the reference and the sample arms, respectively: 𝐸𝑅 = (𝑇𝑟 𝑇𝑠) 1 2⁄ 𝐸𝑖𝑛 𝑒 −𝑖2𝑘𝑛𝑧𝑅 (2) 𝐸𝑆 = (𝑇𝑟 𝑇𝑠 ) 1 2⁄ 𝐸𝑖𝑛 𝑒 −𝑖2𝑘𝑛𝑧𝑆 (3) where n is the refraction index, zr and zs are the path lengths in the reference and sample arms and the factor of 2 arises from the double pass of light. tr and ts are the beam-splitter reference and sample arm intensity transmission coefficients. the sample is not a simple reflecting surface and has a multilayered structure with each layer interface having a reflectance, r, given by the fresnel’s equations. the internal sample refraction index structure is revealed by the eout field that results from superposition (interference) of es and er. optical detectors measure irradiance, which is proportional to the square of the amplitude of the e field. the detector output photocurrent, i(w,k), is proportional to the time average of the irradiance and can be written as: 𝐼(𝑤, 𝑘) = 𝑙𝑖𝑚 𝑇→∞ 1 2𝑇 𝜌 ∫ 𝐸𝑜𝑢𝑡 𝐸𝑜𝑢𝑡 ∗𝑇 −𝑇 𝑑𝑡 (4) where ρ represents detector responsivity. photocurrent appears as a sum of a self-interference term (arising from source self interference and autocorrelation from the multiple reflectors of the sample) and a cross-interference term. 𝐼(𝑘) = 𝐼0 + 𝜌𝑆(𝑘)√𝑅𝑟 𝑅𝑆𝑛 𝑐𝑜𝑠2𝑘(𝑧𝑟 − 𝑧𝑆𝑛 ) (5) i0, also called the dc term, accounts for all self-interference contributions. the second term, obtained by using the euler’s formula, is the cross-interference term, which contains the information on the sample structure. r is the power reflectance and subscripts sn and r refer to sample layer n and reference mirror respectively. following time-domain oct, the technique evolved to the fourier-domain oct where the interference information is spatially encoded (sd-oct), requiring a detector array, or time encoded (ss-oct). in this case, a single-element detector is enough but the broadband source must be wavelength-swept and the data acquisition system must handle high acquisition rates. in the fourier domain, the reference plan is fixed, an important advantage over time-domain oct, which requires mechanical sweeping of the reference mirror position. the layout presented here corresponds to fiber optic based ssoct. iii. experimental apparatus a. oct layout the basic ss-oct layout, composed of interferometer, scanning and acquisition system, is schematically depicted in fig 2. the fundamental blocks of the system are: an axsun laser swept-source axp50125-3 1060 nm (axsun technologies, billerica, ma, usa), with central 1060 nm wavelength, 110 nm bandwidth and sweep frequency of 100 khz; a 400 msps multi i/o acquisition board (x5-400m – innovative integration, california, usa) with two a/d and two d/a channels; an ingaas balanced amplified photodetector (pdb471c – thorlabs) and a x-y positioning/scanning system (gsv002 galvo scanning system, thorlabs gmbh, germany). optical components such as fiber cables, attenuators, couplers, polarization controllers, collimators and focusing objectives complete the system. fig. 2 – overall ss-oct layout the scanning, for cross-sectional (b-scan) and volume scans, is ensured by a x and y galvanometer/mirror system which is controlled, respectively, through a precision waveform generator (afg3101 arbitrary function generator, tektronix) and a dac board (ni 6010, national instruments, austin, texas). digitized fringe data undergoes fourier analysis to produce individual a-scans (along the beam direction), which can be composed in final b-scan images (xz) and/or volumes (xyz). these tasks are carried out by dedicated acquisition/control and processing software. configuration parameters in the software can be tuned according to the needs: the scanning angles, the number of ascans per b-scan and number of b-scans per volume. after acquisition, the whole volume can be stored for posterior analysis or discarded. overall timing parameters that determine the a-scan rate, a/d conversion sampling time and scanning positioning are obtained from two output signals provided by the laser source: sweep trigger and k-clock. fig 3 shows the characteristics of the trigger and clock signals as well as the optical power at the useful sweep range. b. software a customizable software was developed using objectoriented programming (microsoft visual c++/ide) for the 64bit microsoft windows 7 operating system resorting to libraries from innovative technologies to deal with data acquisition and hardware control. j.p. domingues et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-8 i-etc: isel academic journal of electronics, telecommunications and computers fig. 3 – typical spectral power, trigger and clock data for the axsun source (from axsun oct source manual rev. 09) with the current hardware implementation, 1510 data points are digitalized in less than 5 µs, which correspond to the number of sample wavelengths outputted by the laser source. a full frequency sweep corresponds to the entire a-scan that can be computed owing to the fourier transform. the number of useful data points is, nevertheless, restricted to 1376, with the remaining ones produced by the dummy clock outputted by the laser source. table 1 typical scan parameters (from axsun oct source manual rev. 09) parameter value wavelength range 985.0-1095.0 nm sweep frequency 100 khz maximum samples 1510 selected number of samples 1376 duty cycle 45% synchronization between the acquisition and laser emission is paramount for ss-oct. as such, the laser source has an embedded mach-zehnder interferometer (mzi) to provide an uniform clock signal in the wavenumber space (k-space). this allows direct a/d sampling of the optical detector signal event though the wavelength sweep of the laser output is non-linear. the period of the k-clock signal is defined by the free spectral range of mzi. linearized fringe signals with equal k-spacing can be achieved by clocking the high-speed a/d channel of the acquisition board with the clock provided by the source. the fourier transform analysis can then be directly applied to the acquired data. moreover, the laser source also provides a trigger signal which is connected to the sync port of the acquisition board. this signal is responsible for starting the i/o module. fig 4 shows a typical result of an a-scan obtained over a 1 mm thickness glass. fig. 4 interferogram and fft spectrum for 1mm thick coverglass. peaks represent air-glass and glass-air interfaces. iv. performance analysis the performance of the system was established based on commonly parameters found on the literature [6][7]: axial resolution (ar), sensitivity (s), dynamic range (dr) and depth sensitivity roll-off. preliminary results have been obtained with current setup which compares and surpass the required values for biomedical applications which are listed in table 2. table 2 minimum values for the main oct parameters in biomedical applications.[6] [7] parameter value axial resolution < 10 𝜇𝑚 sensitivity fall-off 20 𝑑𝐵 over 2 𝑚𝑚 depth dynamic range 40 − 50 𝑑𝐵 sensitivity > 95𝑑𝐵 the point spread function (psf), the system’s response to a point object (or impulse), corresponds, for the described oct system, to the fourier transform of the interference signal due to a single, perfect reflector. this is accomplished using a gold mirror as sample. the psf of the oct system is shown in fig 5. j.p. domingues et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-8 i-etc: isel academic journal of electronics, telecommunications and computers fig. 5 – point spread function using a gold mirror reflector a. axial resolution axial or depth resolution is one of the key parameters to evaluate an oct system performance. it is defined as the shortest interface separation that can be measured along the light beam´s propagation direction. the theoretical value of the axial resolution, usually specified in the air and assuming a gaussian laser beam power spectrum, can be calculated as follows. being 𝑆(𝑘) = 1 𝜎𝑘 √2𝜋 exp [− (𝑘−𝑘0) 2 2𝜎𝑘 2 ] (6) the gaussian power spectrum and 𝛥𝑘 = (2√2 ln(2))𝜎𝑘 (7) its full width at half maximum (fwhm), its fourier transform is given by 𝛾(𝑧) = 1 √𝜋 exp [− 𝛥𝑘2(𝑧−𝑧0) 2 16ln(2) ] (8) this is the coherence function whose fwhm corresponds to the coherence length of the laser beam. if we used the well established criterium of defining the axial resolution as half of the source coherence length, we obtain for our system (λ0=1060 nm and δλ=100 nm). 𝑧 − 𝑧0 = 𝛥𝑧 = 2𝑙𝑛 (2)𝜆0 2 𝜋𝛥𝜆 = 4.51 µ𝑚 (9) actually, the power spectrum of our laser source does not show a gaussian profile. as such, the psf is experimentally determined by fitting a lorentzian function to the measured psf profile (a-scan) when using a gold mirror sample (fig 6). for our system we obtained an axial resolution of 8.1 µm, in the air, which corresponds to 6.2 µm in tissue. b. sensitivity sensitivity is related with the smallest sample reflectivity, rs,min, that can be detected, defined as the signal level when snr=1. in db units: 𝑆𝑑𝐵 = 10𝑙𝑜𝑔10 𝑅𝑠 𝑅𝑠,𝑚𝑖𝑛 (10) fig. 6 – experimental axial resolution measurement using curve fitting to the gold mirror psf one way of experimentally measure s is to use a gold mirror (rs=1) on the sample side. thus, (6) becomes: 𝑆𝑑𝐵 = 10𝑙𝑜𝑔10 1 𝑅𝑠,𝑚𝑖𝑛 (11) in terms of the measured quantities, sensitivity of the system can be estimated from the ratio between the output from the optimal reflector (the system point spread function, psf) and that from the noise (no sample), the latter being defined as the standard deviation of the readings, leading to: 𝑆𝑑𝐵 = 20𝑙𝑜𝑔10 𝑃𝑆𝐹𝑝𝑒𝑎𝑘 𝜎𝑛𝑜𝑖𝑠𝑒 (12) notice that the factor 2 arises from the square correlation between the photodetector current and the reflectance, r, that is: 𝑃𝑆𝐹 ∝ 𝐼𝐷 (𝑧) ∝ √𝑅 . one way to experimentally determine the sensitivity is to place a neutral density filter of a known optical density (od) before the gold mirror. light will then be doubly attenuated by the filter, resulting in a total attenuation of 20×od (db). added to (12) results in: 𝑆𝑑𝐵 = 20𝑙𝑜𝑔10 𝑃𝑆𝐹′𝑝𝑒𝑎𝑘 𝜎𝑛𝑜𝑖𝑠𝑒 + 20×𝑂𝐷 (13) in these conditions, a sensitivity of 105 db was experimentally determined, allowing to calculate rs,min from (11). c. dynamic range dynamic range (dr) regards the ratio between the maximum and the minimum reflectivity signal that can be measured within the same a-scan. in oct, the smallest reflectivity is taken as the standard deviation of the noise floor when we measure the psf using a gold coated mirror as sample. this way the expression is similar to (8) but all measurements refer to the same a-scan: 𝑆𝑑𝐵 = 20𝑙𝑜𝑔10 𝑃𝑆𝐹𝑝𝑒𝑎𝑘 𝜎𝑛𝑜𝑖𝑠𝑒 (14) j.p. domingues et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-8 i-etc: isel academic journal of electronics, telecommunications and computers experimentally we measured the dynamic range always above 55 db, a value better than the minimum acceptable for biomedical tissues (40-50). better results were obtained with the inclusion of an electronic 100 mhz low-pass filter at the output of the balanced amplifier to remove high frequency noise. although the presence of the filter limits the maximum measurement depth, this is not a problem for the experimental setup as the mirror position is far from the filter’s influence. in the case depicted in fig 7 for the current setup we obtained: 𝐷 = 20 log ( 1627.9 1.1820 ) = 62.78𝑑𝐵 (15) fig. 7 ratio of maximum to the minimum reflectivity that can be measured simultaneously (same a-scan). d. sensitivity roll-off in oct, a depth dependent sensitivity fall-off occurs independently of absorption or scattering from the sample. this fall-off, which places a limit to the maximum tissue depth that can be assessed, is primarily due to the washout of spectral fringes caused by the limited spectral sampling resolution of the spectrometer (in case of sd-oct) or by the limited sweeping resolution and detector bandwidth (in case of ssoct). the sensitivity roll-off of our system was evaluated by doing the sensitivity measurement procedure for different axial positions of the mirror, used as sample. the sensitivity roll-off with depth was assessed over an axial range of 4 mm to find it to be less than 1 db/mm in the first 2.8 mm and about 3 db/mm thereafter as shown in fig 8. v. conclusions the presented layout and setup allowed to achieve the required performance for biomedical field and opens the possibility to future developments in the instrumentation and application sides. fig. 8 – sensitivity measurement over 4 mm depth range. the corner for a depth near 2.8 mm is due to the 100 mhz lp electronic filter. higher electronic frequencies on the detector signal correspond to higher penetration depths. acknowledgment this study was supported by the portuguese foundation for science and technology (pest-uid/neu/04539/2013) and by feder-compete (poci-01-0145feder-007440). references [1] tomlins p h and wang r k (2005) theory, developments and applications of optical coherence tomography. journal of physics: appl. phys. 38, pp. 2519-2535 doi 10.1088/0022-3727/38/15/002 [2] targowski, p wojtkowski m et al (2004) complex spectral oct in human eye imaging in vivo. optics communications 229 pp. 79-84. [3] marschall s, sander b et al (2011) optical coherence tomography – current technology and applications in clinical and biomedical research. anal bioanal chem 400 pp. 2699-2720 doi 10.1007/s00216-011-5008-1 [4] huang d, swanson c et al (1991) optical coherence tomography. science, 254(5035) pp. 1178-1181 [5] silva, s., domingues, j., agnelo j, morgado a, bernardes r (2014), development of an optical coherence tomograph for small animal retinal imaging, ifmbe proceedings volume 41, , pp 419-422. doi: 10.1007/978-3-319-00846-2_104 [6] w. wieser, b. r. biedermann, t. klein, c. m. eigenwillig, and r. huber, “multi-megahertz oct: high quality 3d imaging at 20 million a-scans and 4.5 gvoxels per second.,” opt. express, vol. 18, no. 14, pp. 14685–14704, 2010. [7] m. d bayleyegn, h. makhlouf et al (2012) ultrahigh resolution spectral-domain optical coherence tomography at 1.3 µm using a broadband superluminescent diode light source, opt commun., vol 285, nº 24, pp 5564-5569 j.p. domingues et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-8 i-etc: isel academic journal of electronics, telecommunications and computers mobilewheel a mobile driving station mobilewheel a mobile driving station vitor cunha faculdade de engenharia da universidade do porto (feup), porto, portugal pro10003@fe.up.pt carlos campos faculdade de engenharia da universidade do porto (feup), porto, portugal pro10005@fe.up.pt keywords: human-computer interaction, driving simulator, mobile sensing. abstract: current mobile devices (e.g., smartphones) are equipped with several sensors that allow different forms of user interaction. these devices also offer several connectivity options and a growing computing power which supports its use in new human computer interaction (hci) scenarios. this paper presents the mobilewheel, a system that exploits the capabilities of current mobile devices as a means of interaction with a real-time graphical driving simulation running on a desktop computer. the main aim of this work is to study the possibility of deploying several simplified driving stations to carry out driving simulation trials with scientific purposes. the developed application on the mobile device performs data acquisition from various sensors (focusing on the 3d accelerometer) and also provides different types of feedback to the user. this system represents a ubiquitous, simple and affordable alternative approach to the traditional control of virtual vehicles in driving simulators and could also be applied in other similar architectures. to evaluate and validate this approach several tests were conducted with volunteer users. in these experimental tests different driving operation modes were evaluated: combinations of throttle and gearbox controls. based on the results gathered, we can conclude that the best combination of driving interface, the one that allows better user control over the driving task, is the use of the accelerometer to control the steering wheel coupled with the use of automatic transmission. 1 introduction scientific studies related to driving simulation must conform to specific requirements so they don’t undermine the conducted experimental work. traditional driving stations, such as the one used in the dris [1] driving simulator (fig. 1), require an immersive environment with the presence of an actual vehicle properly equipped with sensors so the participant can control the simulation and to allow data collection on the experimental simulation to pos-processing. in the simulated road environment, in addition to the vehicle driven by the participant, there may be other agents, including other vehicles. the aim of described work in this paper is to study the practicability of a simplified driving station by studying different control interfaces and also determine the most suited for implementing multiple driving stations that are equally interactive. this approach allows that other vehifigure 1: dris driving simulator. cles involved in the simulation could also be controlled by actual participants, thus improving the system scalability. in the previously described context, the increasing number and variety of sensors available on current mobile devices (e.g., smartphones, pdas, tablets) makes them attractive to use in several interaction contexts. in order to exploit the capabilities i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-9 http://journals.isel.pt/index.php/iajetc of these devices and provide an alternative interaction in the particular situation of driving simulations, this article presents the system mobilewheel. the main elements are two software applications, one running on a pc and designed to provide a real-time graphical simulation of the virtual environment and the other one in a smartphone that acts as the controller and dashboard of the driven vehicle. one of the sensors with the highest potential to control the virtual vehicle is the multiaxial accelerometer, which is used in a similar way in current game consoles and by vajk et al. in [2]. in the proposed system, the smartphone application acquires data from this sensor, processes it, and then sends it to the application on the pc via the wireless network infrastructure. additionally, several commands and settings are input through the touchscreen. the driving simulation application logic then uses a dynamic vehicle model that, based on the data received from the mobile device, updates the graphical simulation. within this iteration, the vehicle’s dashboard displayed on the smartphone’s screen is also updated with the new computed values. in addition to visual feedback, sounds and vibration are used to signal events in the simulation (e.g., collision). the remainder of this paper is structured as follows: section 2 presents and discusses previous works that use the accelerometer of mobile devices. section 3 describes in detail the mobilewheel system. section 4 presents the system’s evaluation process and results discussion. finally, in section 5 the conclusions and proposals for possible future developments are presented. 2 related work several already published works demonstrate the potential of the embedded acceleration sensor present in current mobile devices. some examples include the use of this type of sensor as a means to recognize gestures to control software applications (e.g., [3]), integrate continuous sensing approaches (e.g., [4]), among many others. lane’s et al. [5] survey discusses the major possibilities offered by the variety of sensors present in most current mobile devices. one approach that uses the data from the accelerometer as a way of controlling a game that runs on the device and that does not use its keyboard is presented by gilbertson et al. [6]. when compared with the traditional control through the keyboard buttons, this method allowed users to achieve better scores. vajk et al. [2] used mobile phones as controllers in a multiplayer game projected in a large screen. the authors stress that the success of this approach was due to the fact that the users didn’t need to look at the device’s screen, resembling a games console controller. on the other hand, wang and ganjineh [7] used an iphone application as a way to control a real instrumented vehicle. in this case the accelerometer is only used to steer the vehicle while the throttle is controlled through onscreen buttons. the iphone (and other platforms) is also used as the mobile interface to control the lowcost multicopter parrot ar.drone [8], in this case the device’s accelerometer is used together with the touchscreen. the proposed mobilewheel system allows the use of the data gathered from the 3d accelerometer, to control the direction and throttle of a virtual vehicle, while using the pc and smartphone screens to display graphical information from the simulation to the user. 3 the mobilewheel system to achieve the proposed features, fig. 2 presents the mobilewheel adopted client-server system architecture. the applications for the pc and smartphone are the main blocks of this approach, which interact through the network infrastructure. a detailed description of each block is presented in the following sections. 3.1 server application the real-time driving simulation application was developed using mainly the openscenegraph (osg) api [9]. this api consists of a set of libraries used for the development of 3d applications: osg (core), osgutil , osgdb (reading and writing 3d database). additionally it also provides a set of libraries for the manipulation and visualization of 3d environments, specifically the osgga (gui abstraction library) and the osgviewer (viewer library) libraries. the osgga library contains a set of classes and methods to help the easy creation of handlers to control the simulation. in the scope of this server application the drivemanipulatorpda handler was developed to control the simulation, which was derived from the class osgga::matrixmanipulator (osg v2.8.2). linked to the handler was developed a dynamic model of a vehicle, where the engine and gearbox parameters are simulated. two types of gearbox were implemented: manual and automatic. the selection between the different models in the pc server application is made by the operation settings of the application installed in the smartphone. the engine speed can vary between a minimum of 900 rpm and a maximum of 7000 rpm. the automatic transmission model v. cunha et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-9 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc tcp server dynamic models tcp client smartphone cliente software main loop 3d models inputs accelerometer touchscreen outputs screen vibration sound options parameters pc server software main loop wireless figure 2: mobilewheel system architecture. figure 3: driving (left) and settings (right) screens on the smartphone’s application. has three modes: drive, neutral and reverse. in neutral mode the vehicle is disengaged, where a change in the engine rotation does not produce any variation in the vehicle’s speed. in drive mode there is a linear variation between the engine speed and vehicle speed ranging from 0 km/h (900 rpm) to 200 km/h (7000 rpm). the reverse mode implies a negative linear variation between the engine rotation and the speed of the vehicle corresponding to 0 km/h at 900 rpm up to -50 km/h at 7000 rpm. with the manual gearbox model, in addition to the neutral mode, 4 gears (reverse + 1st , 2nd and 3rd ) can be used. in the reverse gear, the relationship between engine rotation and vehicle speed is the same as in the automatic gearbox reverse mode. in 1st gear the vehicle can vary from 0 km/h to 50 km/h (7000 rpm). with the 2nd gear the vehicle speed goes up to 110 km/h and in 3rd gear reaches the maximum speed of 200 km/h. the communication between the pc and smartphone applications is done through a tcp/ip socket connection. at startup the pc application creates an independent thread for the tcp server, which performs all data reading and writing operations with the smartphone client application. this approach also allows the use of several mobile client controllers for more complex driving simulation scenarios. 3.2 client application the smartphone client application was developed in c# (.net cf 3.5) for the windows mobile platform. given the absence of an abstraction layer for the accelerometer (and other sensors) in the windows mobile 6.5 operating system, the wmus api [10] was used. this api allows that the necessary data can be retrieved from the accelerometer in the mobile device used, an htc hd2. fig. 3 illustrates the graphical interface that consists of two main forms. the settings screen lets the user set the various parameters and options of the application: communication settings for connecting to the pc (server); throttle control, through onscreen buttons or the accelerometer; manual or automatic gearbox; smoothing applied to the accelerometer data; use of sounds and vibration. all settings are stored in a local xml file to avoid losing the settings between runs. the two options for gearbox type and throttle control result in four possible combinations to control the v. cunha et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-9 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc steering wheel throttle gearbox data type unsigned char unsigned char unsigned char data range 0 180 0, 1, 2, 3 0, 1, 2 table 1: client to server data structure. speed rpm current gear collision sound data type unsigned char unsigned char unsigned char unsigned char unsigned char data range 0 255 0 255 0, 1, 2, 3, 4 0, 1 0, 1, 2 table 2: server to client data structure. +x -y -z +y -x +z throttle direction figure 4: htc hd2 accelerometer axes. z x � 40° 20° 30° accelerate brakereverse idle figure 5: throttle control action assigned to each angular range. the featured symbols are shown in the speedometer (white bottom area of the dial) to inform the user of the currently active action. virtual vehicle. common to all control modes is the use of the accelerometer to control the direction of the vehicle by rotating the device around the z axis (fig. 4). the throttle (acceleration) control based on this sensor is performed by rotating the device around the y axis. in this case angular intervals are used (fig. 5), where which one is related to a throttle action: accelerate, idle, brake and reverse (to gear reverse when in automatic gearbox). the rotation of the device is converted into angular values using inverse trigonometric functions that use the values obtained from the accelerometer regarding the xz (throttle) and xy (direction) axes. regarding the main driving screen, this form shows the vehicle’s dashboard for visual feedback (e.g., speed) and also capture commands through the touchscreen to send to the server. when the options for manual gearbox and/or throttle control through onscreen buttons are selected, graphical elements in the gui are activated to capture those commands. in addition to visual feedback through the graphical interface, the mobilewheel system includes the use of vibration and sounds. both mechanisms are triggered when the respective command is received from the simulation on the (server) pc. a set of predefined sound files is stored in the smartphone so the simulation application only signals which should be used. two different situations were implemented: collision (sound and vibration) and speeding over a predefined limit (sound). 3.3 communications protocol the data exchange between the pc server application and the smartphone client application uses a purpose developed protocol. data is transmitted via a tcp/ip socket connection at a rate of 10 hz imposed by the mobile device. table 1 shows the data structure sent to the driving simulation application on the pc. the client application sends three variables to the server: steering wheel angle, throttle command (0 reverse, 1 brake, 2 idle 3 accelerate), gearbox (0 automatic gearbox selected, 1 manual gearbox selected, 2 shift one gear up, 3 shift one gear down). table 2 shows the data structure sent in the opposite direction. the software application on the pc sends a structure with five variables in response to the data sent by the smartphone, in this case: speed for the speedometer (0 210 km/h), rpms (0 7000 rpm), current gear to display in the dashboard (0 reverse, 1 neutral, 2 drive/1st gear, 3 2nd gear, 4 3rd gear), collision (on/off), sound file to play (0 none, 1 collision, 2 over speed limit). v. cunha et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-9 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 6: volunteer candidate performing the driving task. figure 7: view of the 3d virtual road environment. 4 evaluation in order to assess the hci capabilities of the system, several tests oriented to the task of driving were performed. the analysis of the results obtained from these tests will allow determining which driving interface is best suited for controlling a simulated vehicle. the main driving task, performed by the volunteer candidates, consisted in driving in a virtual 3d environment until given the indication to stop. candidates were given the instruction: ”drive normally as in a two-way street, without taking any turns at intersections or junctions, until given the order to stop.”. the candidates, before performing the experimental task, drove freely in another virtual environment to become familiar with the system controls. during the driving task several experimental data related to the candidate’s behavior and performance were collected. each candidate performed the driving task using the four different control combinations (modes) in random order: throttle through the accelerometer with automatic gearbox; throttle through the accelerometer with manual gearbox; throttle through onscreen buttons with automatic gearbox and throttle through onscreen buttons with manual gearbox. the parameters collected during the experiments were: • time, measured in seconds since the beginning of the task until the order to stop was given; • diverted attention from the road to the dashboard displayed in the smartphone, which quantifies the number of times the candidate devalues the main driving task to operate the smartphone’s screen; • misalignments, measuring the level of misalignment relative to the expected driving path (0 none, 1 mild, 2 severe, 3 very severe); • evaluation of the number of times the user goes into the wrong (incoming traffic) side of the road (0 never, 1 occasionally, 2 often, 3 always); • off-road/collision, number of times the candidate crashed or gone totally off-road. fig. 6 shows the lab environment where the experiments were carried out, while fig. 7 pictures the 3d virtual road environment used. the sample of volunteer candidates was characterized by considering the following parameters: • driver’s license; • usual driver; • usual computer user; • computer games player; • games console user; • car racing games player. the data collected shows that all candidates have license and use a computer in a daily basis but only 50% of the candidates play driving/racing games. fig. 8 sums up the volunteer users profile. the data analysis methodology was based on applying weights to the different parameters gathered for each interaction mode and candidate. the scores shown in the chart of fig 9 result from the following applied weights: time = 1; deviated attention from the road = 1; misalignment = 10; wrong lane driving = 10; off-road/collision = 60. with this methodology the interaction mode with the best result is the one with the highest numerical value. as illustrated in fig. 9, the interaction mode with the best result is the one using automatic gearbox and the accelerometer to control the throttle (and direction) of the vehicle. in a first stage the tests were performed with ten candidates. after that, in order to sustain the preliminary obtained results a series of additional tests was performed for a total of 21 participants. these additional tests support the earlier results and are included in the previously presented analysis. v. cunha et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-9 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc driver's license usual driver pc user games console user pc games player racing games player 0,00% 20,00% 40,00% 60,00% 80,00% 100,00% sample figure 8: sample profile. accelerometer+auto buttons+auto accelerometer+manual buttons+manual throttle + gearbox: figure 9: total scores by control mode. 5 conclusions and future work the presented work demonstrates the potential of interaction offered by current mobile devices, including the use of the 3d accelerometer for control. the results obtained point in the same direction of the ones reported in [2]. the sample shows that the results are consistent and provide a valid data analysis. data was collected based on the direct observation of the driving task, and each analyst focused solely on collecting data regarding one parameter. by analyzing only the variable time, it can be observed that the average time required by the candidates to perform the main task is lower with the configuration accelerometer/manual gearbox. this situation can be attributed to the fact that the candidates after placing the vehicle in motion, do not often use the gearbox. in all experiments, the data exchanged between the client and server applications was stored in disk files, allowing further analysis. to conclude, the main contribution of this work was the evaluation of the presented control approach in the particular scenario of real-time driving simulation. from the several interfaces evaluated to control a virtual vehicle, the one that produced the best results according to the metrics used is the combination that uses the 3d accelerometer to control the steering and speed of the vehicle (automatic gearbox). based on the presented results it can be stated that this approach effectively enables the use of simplified driving stations in the context of driving simulators with scientific purposes. in the future, is intended to develop the support for several simultaneous users and also to demonstrate the applicability of the mobilewheel system as a mobile driving station alternative in the dris driving simulator. references [1] j.m. leit ao, a. coelho, and f.n. ferreira. dris a virtual driving simulator. in second international seminar on the human factors in road traffic, braga, portugal, 1997. [2] t. vajk, p. coulton, w. bamford, and r. edwards. using a mobile phone as a “wii-like” controller for playing games on a large public display. international journal of computer games technology, 2008(2):1– 7, 2008. [3] s. strachan, r. murray-smith, and s. o’modhrain. bodyspace: inferring body pose for natural control of a music player. in chi ’07 extended abstracts on human factors in computing systems, chi ea ’07, pages 2001–2006, ny, usa, 2007. [4] h. lu, j. yang, z. liu, n.d. lane, t. choudhury, and a.t. campbell. the jigsaw continuous sensing engine for mobile phone applications. in proceedings of the 8th acm conference on embedded networked sensor systems, sensys ’10, pages 71–84, ny, usa, 2010. [5] n.d. lane, e. miluzzo, hong lu, d. peebles, t. choudhury, and a.t. campbell. a survey of mobile phone sensing. ieee communications magazine, 48(9):140 –150, sept. 2010. [6] p. gilbertson, p. coulton, f. chehimi, and t. vajk. using “tilt” as an interface to control “no-button” 3-d mobile games. computers in entertainment, 6(3):38:1–38:13, nov. 2008. [7] m. wang and t. ganjineh. remote controlling an autonomous car with an iphone. technical report, free university of berlin, mar. 2010. [8] ar.drone. http://ardrone.parrot.com. [9] openscenegraph. http://www.openscenegraph.org. [10] windows mobile unified sensor api. http://sensorapi.codeplex.com. v. cunha et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-9 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc optimal control of a multi-field irrigation problem: validation of a numerical solution by the optimality conditions optimal control of a multi-field irrigation problem: validation of a numerical solution by the optimality conditions s.o. lopesa, f.a.c.c. fontesb acmat and departamento de matemática e aplicações, universidade do minho, portugal bsystec-isr, faculdade de engenharia, universidade do porto, portugal sofialopes@math.uminho.pt, faf@fe.up.pt abstract—in this paper, we address the problem of minimizing the total water consumption over the period of a year used to supply different fields with different types of crops. we start by recalling a previous study, where the authors developed an optimal control model for this problem by minimizing the water flowing into a reservoir and where the water from the precipitation can be collected. the numerical solution obtained using such model is analyzed. the main results in this paper are the theoretical validation of the numerical solution via a verifications that such solution satisfies the necessary conditions of optimality in the form of a maximum principle. this way, we are giving further evidence of the optimality of the numerical solution found. keywords: irrigation, optimality conditions, optimal control, maximum principle. i. introduction climate change is a reality. recent findings from nasa [17] report ”globally-averaged temperatures in 2016 were 1.78 degrees fahrenheit (0.99 degrees celsius) warmer than the mid-20th century mean. this makes 2016 the third year in a row to set a new record for global average surface temperatures. (...) noaa scientists concur with the finding that 2016 was the warmest year on record based on separate, independent analyses of the data.”, on the other hand, although our planet is blue, only 2.5% of it is freshwater. from this 2.5% of freshwater, 70% is used in agriculture. according to [9], in 165 countries and 2 territories: • the irrigation water requirement is 1500, 464 km3/yr • the irrigation water total withdrawal is 2672, 640 km3/yr. so, in this huge volume of a scarce resource, there is a waste of 43.8% of water in the current irrigation systems. consequently, there is much that can be done to save water, and it is of utmost important to our planet. naturally, the problem of minimizing water consumption in irrigation systems has been the subject of several researcher works, namely [1], [4] to mention a few using optimization models. using specifically optimal control techniques, that take advantage of managing the storage of rainwater to save water consunption, we mention the works [5], [20]. fig. 1. the planet’s long-term warming trend is seen in this chart of every year’s annual temperature cycle from 1880 to the present, compared to the average temperature from 1880 to 2015. record warm years are listed in the column on the right. credit: nasa/earth observatory.joshua stevens. the authors have been studying the irrigation problem as an optimal control problem. in a previous work [11], the authors develop a model to minimize the water flowing into a reservoir that supplies fields with different crops. in such work, only the numerical solution was obtained, . here, it is proved that the numerical solution satisfies the necessary conditions of optimality in the form of a maximum principle. therefore, in this paper the authors validate the numerical solution, adding evidence of optimality to such solution. the paper is organized as follows. section 2 describes formally the problem and the model used. section 3 presents the numerical results obtained for the problem. in section 4 the numerical solution is validated. concluding remarks are drawn in section 5. ii. problem setting in this section, we describe the problem of minimizing the total water consumed in irrigation over a long period of time (one year is considered here) to supply various cultivation fields with different types of crops. the problem is modeled as an optimal control problem, based on the recent work reported in [13], being able to capture adequately not only the dynamics of the evolution along time of the variables used, but also i-etc: isel academic journal of electronics, telecommunications and computers cetc2016 issue, vol. 3, n. 1 (2017) id-9 http://journals.isel.pt the constraints that represent physical restrictions as well as constraints that guarantee the needs of heathy crops. we start by defining the variables used in the model. the controls are: v total water flow coming from the tap and uj water flow introduced in field j via its irrigation system. the states are: xj water in the soil of field j and y total of amount of water stored in reservoir. the aim is to minimize the total of water flow coming from the tap to a reservoir, min ∫ t 0 v(t)dt, where v(t) denotes the total water flow coming from the tap at time t. the reservoir supplies the multiple cultivated fields where each field can have different areas and different cultures. the water from the precipitation also can be collected into the reservoir. therefore, the variation of water in reservoir is given by ẏ(t) = v(t) − p∑ j=1 ajuj(t) + cg(t), where y represents total of amount of water stored in the reservoir, aj represents the area of each field j, cg(t) represents collected water in a certain area c coming from the precipitation g in the time t, and uj is the water flow introduced in field j via its irrigation system. the variation of water in the soil is given by the hydrological balance equation, that is, the variation of water in the soil is equal to what enters (water via irrigation systems and precipitation) minus the loss (evapotranspiration of each crop hj and loss by deep percolation βxj(t), a percentage of water that is in the soil). so, ẋj(t) = uj + g(t) −hj(t) −βxj(t), ∀j = 1, . . . ,p where xj water in the soil of field j. we assume that each field has only one crop. in order to ensure that the crop is in good state of conservation, the water in the each field has to be sufficient to satisfy the hydric needs of each crop (xmin), that is: xj(t) ≥ xminj. the physical limitations of the amount of water that comes from a tap, the amount of water that comes from the irrigation systems, and the reservoir are given, respectively, by: y(t) ∈ [0,ymax] uj(t) ∈ [0,mj] v(t) ∈ [0, ∑ j ajmj] where ymax is the maximum quantity of water in the reservoir and mj is the maximum water flow coming from the tap in each field. we assume that at the initial time the humidity in the soil of each field and the water in the reservoir are given. also, the water in the reservoir at the initial time and to the final time are imposed to be equal. so, we assume that xj(0) = x0j y(0) = y(t) = y0 a. the model in summary, the optimal control formulation for our problem (ocp) with p fields and final time t is min ∫ t 0 v(t)dt subject to: ẋj(t) = −βxj(t) + uj + g(t) −hj(t) a.e. t ∈ [0,t], ∀j = 1, . . . ,p ẏ(t) = v(t) − ∑p j=1 ajuj(t) + cg(t), a.e. t ∈ [0,t], xj(t) ≥ xminj, ∀t ∈ [0,t], ∀j = 1, . . . ,p y(t) ∈ [0,ymax], ∀t ∈ [0,t] uj(t) ∈ [0,mj], a.e. , ∀j = 1, . . . ,p v(t) ∈ [0, ∑ j ajmj], a.e. xj(0) = x0j, ∀j = 1, . . . ,p y(0) = y(t) = y0. iii. numerical results we use a direct method discretization to obtain the numerical results for the optimal control problem by transcribing it as a mathematical programming problem. for that, we start by considering a finite time grid i = 1, . . . ,n + 1 xi = x(ti) ui = u(ti) where ti = (i− 1)h and h = t/n. using the euler-type discretization, the differential equation ẋ = f(t,x,u) is approximated by xi = xi−1 + hf(ti−1,xi−1,ui−1). to implement this optimization problem, the fmincon function of matlab is used with the algorithm “active set”, by default and the parameter “tolfun” is considered 1e − 6. the rainfall is estimated using an average of 10 years data of rainfall for each month of the year, the dates are collected from instituto português do mar e da atmosfera ([10]). the pennman monteith methodology [22] is used to calculate evapotranspiration of culture along the year, according the following formulation: et(ti) = kcet0(ti), where kc is the culture coefficient for the evapotranspiration and et0 is the tabulated reference value of evapotranspiration from [19] for the lisbon region. to simulate the problem, we assume that: p = 3 t = 12 mj = 10 m3/month for each fieldj y0 = 0.01 ×at ymax = 0.05 ×at β = 15% s.o. lopes et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-9 i-etc: isel academic journal of electronics, telecommunications and computers where at is the total area, (for more details see [12]). here, we also consider three crops: wheat in 1000 m2, sugar cane in 750 m2 and olive in 1250 m2. the precipitation water is collected in an area of 200 m2. for these three crops, we consider: xmin = [0.033 0.021 0.032] kc = [0.825 0.95 0.5] x0 = 5xmin the results obtained are reported in fig. 2 and 3, (see also [12]): as expected, the crops need incoming water between may and september. the months when the water consumption is higher are june, july. the crop that needs less water is olive with the largest area 1250 m2. note that we are imposing the constraint that the water volume in the reservoir at the initial time is equal to the water at the final time. it can be seen that until april the water from the precipitation is saved in the reservoir and until this month the tap is not opened. the tap is closed in september. june and july are the months when the irrigation takes the highest value, and also when the water in the reservoir decreases. iv. validation of the solution since the problem has inequality constraints that are active for some period of time, it is not easy to get a complete analytical solution. however, we can verify that the numerical solution for the mathematical programming problem satisfies the necessary condition of optimality in the normal form of maximum principle (mp). by normal form, we mean that the scalar multiplier associated with the objective function — here called λ — is nonzero. the normal forms of the mp are guaranteed to supply non-trivial information, in the sense that they guarantee that the objective function is taken into account when selecting candidates to optimal processes. there has been a growing interest and literature on strengthened forms of nco for optimal control problems (ocp). the normality results reported in literature require different degrees of regularity on the problem data [6], [18], [16], [3], [2], [7], [8], [14], [15]. in rampazzo and vinter [18], the mp can be written with λ = 1, if there exists a continuous feedback u = η(t,ξ) such that dh(ξ(t)) dt = ht(t,ξ) + hx(t,ξ) ·f(t,ξ,η(t,ξ)) < −γ′ (1) for some positive γ′, whenever (t,ξ) is close to the graph of the optimal trajectory, x̄(·), and ξ is near to the state constraint boundary. there should exist a control pulling the state variable away from the state constraint boundary. in the problem, we can find three inequality constraints, where the respective function are: h1(x) = xmin −x h2(y) = −y h3(y) = y −ymax. observing the numerical solution in fig. 3, we note that the trajectory y never touches the upper limit. so, the inequality constraint y(t) ≤ ymax is not active for any time in [0,t]. in this case, it is not necessary verify the constraint qualification to ensure normality. from (1), we write h1xj(ξ1j(t)) ·f(t,ξ1j,η1j(t,ξ1j)) = −(η1j(t,ξ1j) + 4j(t,ξ1j)) ≤−γ1, (2) where 4j(t,ξ1j) = g(t)−βξ1j . for a ξ1j on a neighbourhood of x̄j , we can always choose η1j sufficiently large so that the equation (2) is satisfied, as long as mj > βx̄j(t) − g(t), a condition we can impose with loss of generality. again, from (1), we have h2y(ξ2(t)) ·g(t,ξ2,η2(t,ξ2)) = −(η2(t,ξ2) − ∑p j=1 ajuj(t) + cg(t)) ≤−γ2, (3) as long as condition ymax > ∑p j=1 ajūj(t)−cg(t) holds, equation (3) is satisfied. this condition can be made to hold but adequately dimensioning the reservoir capacity. thus, the inward pointing condition (1) is satisfied and normality follows. the problem data is sufficiently regular and the standard hypotheses under which the maximum principle holds are satisfied. since the normality is ensured, we can apply the maximum principle with λ = 1, (for example from [21]). define hamiltonian function: h(t, (x,y), (p,r), (u,v),λ) = (p,r) · (−βx + u + g(t) − h(t),v + a · u + cg(t)) −λv assuming that ((x̄, ȳ), (ū, v̄)) is a minimizer for (ocp), then ∃(p,r) ∈ w1,1([0, 1] : rnp × r) and (µj,νl,νu ) ∈ c∗(0, 1): s.o. lopes et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-9 i-etc: isel academic journal of electronics, telecommunications and computers fig. 2. crops need. fig. 3. reservoir −(ṗ(t), ṙ(t)) = h(x,y)(t, (x̄, ȳ), (q(t),w(t)), (ū, v̄), 1) h(t, (x̄, ȳ), (q(t),w(t)), (ū, v̄), 1) = max(u,v)∈[0,m]×[0,r] h(t, (x̄, ȳ), (q(t),w(t)), (u,v), 1) a.e.; supp{µj}⊂{t ∈ [0,t] : xj(t) = xminj} supp{νl}⊂{t ∈ [0,t] : y(t) = 0} supp{νu}⊂{t ∈ [0,t] : y(t) = ymax} (q(0),−w(0),−q(t),−w(t)) ∈ + nc (x̄(0), ȳ(0), x̄(t), ȳ(t)) (4) where: q(t) =   p(t) − ∫ [0,t) µ(ds), t ∈ [0,t) p(t) − ∫ [0,t ] µ(ds), t = t, w(t) =   r(t) − ∫ [0,t) νl(ds) + ∫ [0,t) νu (ds), t ∈ [0,t) r(t) − ∫ [0,t ] νl(ds) + ∫ [0,t ] νu (ds), t = t. and c = {x0}×{y0}×rnp ×{y0} from the adjoint equation, we obtain  ṗ(t) = βq(t) ṙ(t) = 0 (5) from the weierstrass condition, we get: (q(t) − a) · (u − ū) + (w(t) − 1)(v(t) − v̄(t)) ≤ 0 (6) since the endpoints are fixed, except for the final endpoint of x that is free, the transversality condition is simplified to q(t) = 0. therefore, in this problem, the maximum principle can be written as follows. if ((x̄, ȳ), (ū, v̄)) is a minimizer for (ocp), s.o. lopes et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-9 i-etc: isel academic journal of electronics, telecommunications and computers then ∃(p,r) ∈ w1,1([0, 1] : rnp × r) and (µj,νl,νu ) ∈ c∗(0, 1):  ṗ(t) = βq(t) ṙ(t) = 0 (q(t) − a) · (u − ū) + (w(t) − 1)(v(t) − v̄(t)) ≤ 0 q(t) = 0 furthermore, in the case where (ū, v̄) ∈]0, m[×]0,r[, the weierstrass condition can be written as:{ q(t) = a(t) w(t) = 1. (7) from fig. 4, we can conclude that (ū, v̄) ∈]0, m[×]0,r[ in june, july and august, and therefore we can apply the weierstrass condition equation in the form of (7) in these months. we can observe the multiplier associated with the crops state are equal to the area of the fields of the corresponding crops in the months mentioned above: wheat multiplier := 1000 sugar canne multiplier := 750 olive multiplier := 1250. we also observe in fig, 5, that the tranversality condition is satisfied: q(t) = 0. we also can see that in june, july and august the multiplier associated with the reservoir state is equal to 1. we conclude that for (ū, v̄) ∈]0, m[×]0,r[ the numerical solution fulfills the weierstrass and remaining conditions of the maximum principle in the normal form. in summary, we conclude that: • the constraint qualification is verified, consequently we can ensure normality; • for (ū, v̄) ∈]0, m[×]0,r[ the numerical solution satisfies the weierstrass condition of the maximum principle in the normal form. • the numerical solution satisfies the transversality condition. • the numerical solution satisfies the adjoint equation. v. conclusion we have addressed an irrigation planning problem of minimizing the water used to supply different fields with different types of crops, sharing a common reservoir, which not only serves as storage, but also is from where the rain fall water is collected. we have used an optimal control model, which is capable of adequately representing the dynamics of the humidity of the soil taking into account irrigation, evapotranspirations and infiltration, as well as taking into account the physical and problem requirement constraints. we have carried out a detailed analysis of the results obtained numerically, namely by establishing for the problem the necessary conditions of optimality in the form of a maximum principle and verifying that the numerical solution fulfills all the necessary conditions. in the absence of additional candidates for optimality (extremals satisfying the necessary conditions) in the neighborhood of the solution analysed, we may conclude that such solution is in fact locally optimal. acknowledgments. this work was partially supported by norte regional operational program through project norte–45–2015–02 stride – smart cyberphysical, mathematical, computational and power engineering research for disruptive innovation in production, mobility, health, and ocean systems and technologies, as well as projects poci–010145-feder-006933 -systec–research center for systems and technologies institute of systems and robotics, universidade do porto, uid/mat/00013/2013 cmat centro de matemática da universidade do minho, and ptdc/eei-aut/2933/2014, poci-01-0145-feder-016858 toccatta, funded by feder funds through compete2020 programa operacional competitividade e internacionalização (poci) and by national funds through fct fundação para a ciência e a tecnologia. references [1] surface irrigation optimization models. journal of irrigation and drainage engineering, 112(1), 1061. [2] p. bettiol and h. frankowska. normality of the maximum principle for nonconvex constrained bolza problems. journal of differential equations, 243:256–269, 2007. [3] a. cernea and h. frankowska. a connection between the maximum principle and dynamic programming for constrained control problems. siam journal of control and optimization, 44:673–703, 2005. [4] d. j. bernardo, n. k. whittlesey, k. e. saxton, and d. l. bassett. irrigation optimization under limited water supply. transactions of the asae, 31(3):0712–0719, 1988. [5] he fawal, d georges, and g bornard. optimal control of complex irrigation systems via decomposition-coordination and the use of augmented lagrangian. in systems, man, and cybernetics, 1998. 1998 ieee international conference on, volume 4, pages 3874–3879. ieee, 1998. [6] m. m. a. ferreira and r. b. vinter. when is the maximum principle for state constrained problems nondegenerate? journal of mathematical analysis and applications, 187(2):438–467, 1994. [7] f. a. c. c. fontes and s. o. lopes. normal forms of necessary conditions for dynamic optimization problems with pathwise inequality constraints. journal of mathematical analysis and applications, 399:27– 37, 2013. [8] fernando a.c.c. fontes and helene frankowska. normality and nondegeneracy for optimal control problems with state constraints. journal of optimization theory and application, 2015. [9] karen frenken and virginie gillete. irrigation water requirement and water withdrawal by country. fao aquastat, 2012. [10] ipma. https://www.ipma.pt/pt/. accessed: 2017-01-20. [11] s.o. lopes, f. a.c.c. fontes, rui m.s. pereira, mdr de pinho, and c. ribeiro. optimal control for an irrigation planning problem: characterization of solution and validation of the numerical results. lecture notes in electrical engineering, 321:157–167, 2015. [12] s.o. lopes and f.a.c.c. fontes. optimal control for an irrigation problem with several fields and common reservoir. lecture notes in electrical engineering, pages 179–188, 2016. [13] sofia o. lopes, f. a.c.c. fontes, rui m.s. pereira, mdr de pinho, and m. gonçalves. optimal control applied to an irrigation planning problem. mathematical problems in engineering, 17:10, 2016. s.o. lopes et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-9 i-etc: isel academic journal of electronics, telecommunications and computers fig. 4. controls an their boundaries fig. 5. multipliers associated to the crops fig. 6. multipliers associated to the water in the reservoir [14] sofia o lopes, facc fontes, and mdr de pinho. on constraint qualifications for nondegenerate necessary conditions of optimality applied to optimal control problems. discrete and continuous dynamical systema, 29(2), 2011. [15] sofia o lopes, fernando acc fontes, and mdr de pinho. an integral-type constraint qualification to guarantee nondegeneracy of the maximum principle for optimal control problems with state constraints. systems and control letters, 62(8):686–692, 2013. [16] k. malanowski. on normality of lagrange multipliers for state constrained optimal control problems. optimization, 52(1):75–91, 2003. [17] nasa. http://climate.nasa.gov/news/2537/nasa-noaa-data-show-2016warmest-year-on-record-globally. accessed: 2017-01-20. [18] f. rampazzo and r. b. vinter. a theorem on the existence of neighbouring feasible trajectories with aplication to optimal control. ima journal of mathematical control and information, 16:335–351, 1999. [19] j. r. raposo. a rega — dos primitivos regadios as modernas técnicas de rega. fundação calouste gulbenkian, 1996. [20] j mohan reddy. local optimal control of irrigation canals. journal of irrigation and drainage engineering, 116(5):616–631, 1990. [21] r. vinter. optimal control. birkhauser, boston, 2000. [22] i. a. walter, r. g. allen, r. elliott, d. itenfisu, and et.al. the asce standardized reference evapotranspiration equation. rep. task com. on standardized reference evapotranspiration, 2002. s.o. lopes et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-9 i-etc: isel academic journal of electronics, telecommunications and computers bluelab iot, a universal software platform for iot data acquisition devices bluelab iot, a universal software platform for iot data acquisition devices vitor vaz da silvaab, adepartment of engineering of electronics, telecommunication and computers, isel/ipl instituto superior de engenharia de lisboa, portugal bcts-uninova, portugal vsilva @ deetc.isel.ipl.pt abstract1— physical devices with different sensors and sampling rates distributed over several unrelated locations need to store their values over time. applications that need the result of a set of sensors must access their data. a common and simple interface within the physical devices, monitoring stations, to store data on a database is needed; as also a simple and common retrieval interface for any application that only shows the data as it is or processes it into higher levels of significance. bluelab iot is a platform with libraries and an interface application to aid that development; a working implementation is provided. keywords: iot, database, station, device, storage, monitoring i. introduction the expected number of iot (internet of things) connected and accessed through the internet is increasing day by day [1]. the applications covered by these devices are very diverse: simple fixed environmental sensors, home gadgets and appliances, industry modules, and mobile devices like wearables and transportation systems’ data. the ease with which these devices can be bought and installed also contributes to the increasing number of available iots including the young developers, students, hackers and curious self-learning individuals. the number of devices which are independent of any enterprise or government will increase likewise, and contribute to the smart cities as shrubs in a forest. hardware and software on the devices and software on the server need a bridge that has to be built from both sides. this may be a difficult task for developers that are only focused on one of the sides of that bridge. to overcome that gap, this work provides such a bridge with a library on each end. this platform aims for data acquisition, monitoring, thus does not allow, at this stage, the possibility of acting upon the devices, even if there is an actuator on them. the main purpose of this work is to build a universal platform that is simple, for immediate usage with minimal configuration; a plug and play type, which can also be used as a tool to develop other more complex approaches [2]. ii. architecture the bluelab iot architecture for data acquisition devices is presented in figure 1. each physical device can have several sensors, and devices may differ among themselves. these devices communicate either directly with the server or they may communicate with a data gateway where after some processing the result is forwarded, relayed, to the server where it is stored. each gateway can also be seen as a high level sensor and the data that it sends to the server is stored as if it had been collected by the gateway’s sensors. to use the platform each device or gateway that communicates with the server needs to create a session, which is done by a login with password and, after a successful authentication, receives a session id that must be used in all subsequent communications. each session can be limited in time and or data volume, and when it expires a new connection must be established. there are two databases: user and data; one for user authentication and session validation and another where the data is stored. physical devices are called stations and each data is stored as an entry. the database is implemented with postgresql [3]. different users have different schemas, i.e. each has its own database, this means that there is independency of data; two different users can have stations with the same station number. figure 1 – bluelab iot structure. a. user a user must register itself into the platform by email and password, where after a successful registration process can use the system. the user must create an environment, which is composed by all stations, and is the owner of that environment. i-etc: isel academic journal of electronics, telecommunications and computers iot-2018 issue, vol. 4, n. 1 (2018) id-4 http://journals.isel.pt b. station a station belongs to the user that created it and has a unique sequential number that is associated with it during the creation process where other information about the station is also stored as for example the station’s unique number (e.g.: mac address, imei, …), its latitude, longitude and height, a name and a description. a set of privileges define a station, delete protecting it, so that a specific station may or not be deleted along with all its entries, or if a range of entries can be deleted from that station. c. entry the iot data database’s entity-relation model is presented in figure 2. it is the schema of a single user. figure 2 – bluelab iot data database entity relationship model regarding the entry entity shown in figure 2, a data entry is a combination of a key-value pair, a key (kkey) and a value (vvalue) that is associated with it (e.g.: “temp”, 12.3) and, two date stamps, one indicating the time the sample was done (t_stamp) and, the other the time the value was stored in the database (db t stamp). these timestamps allow synchronization of data between stations, and their retrieval. some stations, devices, may not have a real time clock (rtc) and their timestamp might be measured from the last reset. different sensors may also have different sampling frequencies, digital filtering and eventually a temporal reference associated with it and stored as the t_stamp. an entry is identified by its increasing sequential number seq_num within a specific station. the sequence number is used for several reasons: to reference all key value pairs to the same sampling window time; as an acknowledge, as it is used as a result of storing the data value pairs and avoid repeating values if the device recovers from an error condition. iii.station each device used in this project is composed by an esp8266-e12 embedded system with wifi [4], connected to all or eventually some of the following set of sensors: humidity, pressure and temperature sensor, bme280 [5], air quality sensor mq135 [18, 19], and real time clock ds3231. the sensors can also be switched off in each module during configuration right after the reset of the device. the software base is that of arduino’s system [8]. the arduino system has two top level functions, the setup() and the loop(). the setup is meant to be executed only once, and the loop is called repeatedly by the system. shown in figure 3 is a detail of the setup code related to the bluelab iot library. the code shown does not handle error conditions, to make it easy to understand, and the access to the wifi is also not shown. a real implementation should handle error conditions, have a retry loop and eventually make reset to the module if needed, as it is in the complete code available in links shown further down. the setup code example shown in figure 3 involves the creation of an object, dbconn, that handles all calls to the iot data database. a login is then necessary, with the user email and password, which corresponds to the registration process, which had to be done previously, and is described further down in the text. a successful login returns a session id which is then used on subsequent calls using the dbconn object. the actual sequence number is retrieved from the database and incremented for the next frame to be sent. figure 3 – detail of the setup code of a device although not necessary it is a good idea to store in the database that the device has had a reset. a frame with the reset condition, along with the time it occurred, is stored in the iot data database. any value can be associated with the “reset” key, for example indicating the reason or error code. there is no difference between a “reset” key and any other key (e.g. “abc213”) in the way it is processed and stored. after the setup process where all the device’s configuration is made, the arduino calls the loop() function repeatedly and the associated code is where the data is acquired and sent to the database. a detail of that process is shown in the code of figure 4, where all sensors are acquired with the same rate. it is a very simplistic approach and a real situation would preferably use // create an object to access bluelab dbconn = new databaseconnection( login_host, login_url, data_host, data_url); // login in dbconn->login(usr_email, usr_password); // get current sequence number int seqnum=dbconn->getseqnum(deviceid); // prepare next frame’s sequence number seqnum++; // running the setup code means that // a reset has occurred in the device // store that information! // ... get the time the reset occured long long tstamp=util.epochmicroseconds(); // ... build the frame dbconn->newframe(deviceid, seqnum, tstamp); // ... add the data to be stored dbconn->addkeyvalue("reset",0); // ... send the frame dbconn->sendframe(); // prepare next frame’s sequence number seqnum++; v.silva | i-etc iot 2018, vol. 4, n. 1 (2018) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt interrupts. in either case, only the data that is available with the same timestamp (tstamp) should be on the same frame because all key values will have the same sequence number. on the other hand, if two different frames are sent with different timestamps and with the same sequence number they will be stored as long as the keys are different. this is helpful when it is necessary to relate and synchronize sampling events from sensors with different sampling times. figure 4 – detail of the acquisition and storing code of a device each data value is identified by a key string and a value. to store data with the same sequence number but different timestamps it is necessary to build two different frames as shown in figure 5. figure 5 – key-value pairs with same sequence number but different timestamps. devices need not be identical. the key-value pairs stored in the database can be specific to each device. if for example the same type of readings has to be made on a large area, it makes sense that the key strings be identical, even if the devices are not physically identical: i.e. two devices with different hardware both reading temperatures, can use the same key string “temp”. all calls are made through the https protocol by json format posts. a station is in fact a set of sensors and a communication link to a database. so any other equivalent configuration can also be considered as a station. this project also uses the smartphone as a mobile station. a. mobile station a smartphone has sensors that can be sampled and its values sent to the bluelab iot, this is achieved by using the same https calls as any other station. a simple application is shown in figure 6. figure 6 – smartphone bluelab iot application showing a) login b) station description c) light sensor ad d) gps sensor figure 6 shows several sequences of the smartphone as a bluelab iot station. after a successful login a) the station is created automatically if it doesn’t exist and the user provides some information that is shown b), right after logging in. then choosing the light sensor c) all light changes within a certain threshold are sent to the database. by choosing the gps sensor d) a set of geospatial and speed information changing within certain thresholds are also sent to the database. // new frame dbconn->newframe(deviceid, seqnum, tstamp); dbconn->addkeyvalue("pressure", press); dbconn->addkeyvalue("humi", humi); // send the frame dbconn->sendframe()); // new timestamp other =util.epochmicroseconds(); // new frame – same seqnum different tstamp dbconn->newframe(deviceid, seqnum, other); dbconn->addkeyvalue("temp", temp); dbconn->addkeyvalue("air", airqual); // send the frame dbconn->sendframe()); seqnum++; // get new values float temp=bme.readtemperature(); int press=bme.readpressure(); float humi=bme.readhumidity(); int airqual=analogread(sensorgas); tstamp=util.epochmicroseconds(); // build a new frame dbconn->newframe(deviceid, seqnum, tstamp); // add the values to the frame dbconn->addkeyvalue("pressure", press); dbconn->addkeyvalue("humi", humi); dbconn->addkeyvalue("temp", temp); dbconn->addkeyvalue("air", airqual); // ... send the frame dbconn->sendframe(); // prepare next frame’s sequence number seqnum++; v.silva | i-etc iot 2018, vol. 4, n. 1 (2018) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt iv.functions available functions for the hardware module can be downloaded from https://github.com/tektonia/bluelab iot where a working example can be uploaded to a compatible esp8266-e12 hardware, or adapted for another microcontroller module with equivalent functionalities. the bluelab iot functions available for the device are shown in figure 7. most of these functions have been used in the code shown in figure 3, namely: databaseconnection(), that creates a new object, and it is only needed to be done once in the program; login(), that establishes a connection with the user data system. it has to be called every time a session id is needed to access the iot data database system; newframe(), creates a new frame header with an empty data content header; addkeyvalue(), adds key-value pairs to the data content of a frame; sendframe(), sends a frame with the pre filled headers and content data header; getseqnum(), returns the highest sequence number stored in the database for a specific station id. figure 7 – functions available for the device of the functions listed in figure 7 that have not been used in the examples shown are: setdebug(), that fills a debug header for all frames allowing a more verbose response from the calls made to the bluelab iot data database system; longlongtostring(), which converts a long long data type value to a string; and sendlastframe(), that resends the last frame with the most recent session id. this function is called specifically when the sendframe() call fails due to an invalid or outdated session id. in that situation a new login has to be issued and after its success the sendlastframe() function is called to send the outstanding frame. functions available for the api (application programming interface) software module can be found in https://github.com/tektonia/bluelab iot under the api folder. these functions are shown in figure 8. all functions have the same $data parameter which corresponds to the contents of the post call made through the https protocol to https://bluelab.pt/iot/calliot.php; the only entry point interface to access the bluelab iot data database. it is expected that all the https calls have a valid e-mail and corresponding session id before the functions of figure 8 end up being called. these functions allow the creation, destruction of the database, and the insertion and retrieval of data related to the iot database represented in figure 2. functions to change or remove a station definition, or delete data within a time range are available. though, it is not possible to change the data acquired by a device. if a device changes place (latitude or longitude), then it is preferable to create a new station with the new position and an observation stating the changes and data is stored with the new station id. figure 8 – functions available for the api a working api that uses the functions shown in figure 8 is presented in section v. any user can build its own api by calling the required functions as shown in several examples on figure 9, where javascript functions are called as they are in the html files on the bluelab site. /**** https post function calls ****/ function getallstationids($data); /* information of a specific station_id */ function getstation($data) function getallstations($data); /* all entries from requested station_id and sequence number */ function getentry($data) /* all entries from requested station_id */ function getallentriesfromstation($data) /* entry with the lowest sequence number of the requested station_id */ function getfirstentry($data); /* highest sequence number entry of the requested station_id */ function getseqnumber($data); /* stores a single key-value pair and its timestamp for the specified station_id and sequence number */ function storekeyvalue($data) /* stores an array of key-value pairs with the same timestamp for the specified station_id and sequence number */ function storearraykeyvalue($data) /* creates a new station with information supplied by the user */ function createstation($data) function createdatabaseifnotexists($data) function createdatabase($data); function destroydatabase($data); /**** databaseconnection class ****/ databaseconnection(string l_host, string l_url, string d_host, string d_url); bool login(string mail, string pass); string longlongtostring(long long ll); void newframe(int stationid, int seq, long long timestamp); void addkeyvalue(string key, <type> value); void setdebug(bool deb); string sendframe(); string sendlastframe(); int getseqnum(int stationid); v.silva | i-etc iot 2018, vol. 4, n. 1 (2018) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt https://github.com/tektonia/bluelab_iot https://github.com/tektonia/bluelab_iot https://bluelab.pt/iot/calliot.php figure 9 – javascript calls to the bluelab iot database functions examples in figure 9 show the interaction with the https calls and processing its result. function seqnum() sends a request to the database and when it is processed, and its result is received through the function seqnumreceived(data) defined as a callback function in seqnum(); where data is a structure with the result of the call and data.seq_num holds the highest sequence number in the entries that belong to the station_id parameter in the getseqnumber function call. the main difference between newentry() and multientry() is that the first only sends a key-value pair and the second an array of key-value pairs. v. validation and use of the proposed platform at the bluelab iot site https://bluelab.pt choose the iot link on the iot icon and a page will be displayed as shown in figure 10, where it is possible to make a new registration, recover a forgotten one, or login in with the correct credentials. a) b) figure 10 – bluelab iot a) entry page and links the register process is to acquire a unique entry and identity into the system, with the e-mail as the user and a chosen password. if a new registration is not made, a demonstration login can be used instead (user: a@a.a password: a); the state of that login depends only on those that use it so no relevant information may be present on its associated database. nevertheless some collected data is available to be used on the working examples as those below. after a successful login a page similar to that of figure 11 shows up. figure 11 – example showing the user environment and data to be requested there are some clickable buttons, shown in figure 11, that allow: to show all user stations (the environment), insert data into any station (as a debugging procedure, simulating the entry of device data, which is supposed to be added through the physical device interaction), and show data (download it as well) that can be visualized either as a scatter graph or on a map if the data holds geospatial information. the selected selection shown in figure 11 is to get data from 10/08/2018 20:00h till 11/08/2018 08:59h from stations 1,17,18, and 19 that hold the key “temp”. a dataset with 4 variables was received and shown as figure 12. function seqnum(sid, statid){ $.post( "https://bluelab.pt/iot/calliot.php", {func: "getseqnumber", email: $('#email').val(), sessionid: sid, station_id: statid }, seqnumreceived, "json" ); } function seqnumreceived(data){ seqnum = parseint(data.seq_num)+1; } function newentry(sid,st_id,seq,tm,ky,vl){ $.post( https://bluelab.pt/iot/calliot.php", { func: "storekeyvalue", sessionid: sid, email: $('#email').val(), station_id: st_id, seq_num: seq, t_stamp: tm, key: ky, value: vl }, newentrycreated, "json" ); } function multientry(sid,st,seq,t,mx,mn,ac){ $.post( "https://bluelab.pt/iot/calliot.php", { func: "storearraykeyvalue", sessionid: sid, email: $('#email').val(), station_id: st, seq_num: seq, t_stamp: t, dados: { tempmax: mx, tempmin: mn, tempact: ac } }, multientrycreated, "json" ); } v.silva | i-etc iot 2018, vol. 4, n. 1 (2018) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt https://bluelab.pt/ figure 12 – temperature data in c from 4 different devices the scatter graph of figure 12 represents the temperature that similar stations acquired during the selected period. stations 1,18 and 19 are similar as they were measuring room temperature in different places of a house, while station 17 was positioned outside the house in a shade so that it would not receive direct sunlight. data in station 19 has a rhythmic modulation which is correlated to a continuous charging of a battery that is inside the physical device; the battery heats up while it charges. this station can be seen on figure 13 a). figure 13 – two different stations a) rechargeable battery, temperature, humidity and pressure sensor and b) rtc, temperature, humidity, pressure and air quality sensor. presented in figure 13 are two different stations with the same microcontroller; station number 1 is on image b), it is continuously connected to a power source and has a real time clock, temperature, pressure, humidity and air quality sensors. it has been continuously sending data to the database since 25.07.2018, every 90s interval. to try with a real example, upload the code supplied in https://github.com/tektonia/bluelab iot to an esp8266-e12 module. the ssid and ssid_password for the wifi network must be set (and also the user e-mail and password) before compilation and uploading to the physical device; station. vi.conclusions this system is easy to setup and made to work with the code supplied through the links available on previous sections. the current database [3] can hold up to 4 tera byte of data. if this is an issue, data can be withdrawn from this database and stored somewhere else, eventually not storing the raw data but statistical outputs. a developer user of the bluelab iot can build its own api with the functions provided above, and install any quantity of different stations (maximum set as 25). using a mobile station with geolocation for example a smartphone, which stores the path to bluelab iot, and another station, carried along with the smartphone, that sends temperature values, allows the temperature to be geosynchronized. future developments like a sophisticated representation for the visualization of the data, including statistic information will be available in the bluelab iot site; one of the basic requirements of the iot elements [9]. references [1] t. kramp, r. van kranenburg, and s. lange, “introduction to the internet of things,” in enabling things to talk: designing iot solutions with the iot architectural reference model, 2013. [2] m. bauer et al., “iot reference architecture,” in enabling things to talk: designing iot solutions with the iot architectural reference model, 2013. [3] postgresql, “postgresql: the world’s most advanced open source database,” http://www.postgresql.org/, 2014. . [4] espressif, “esp8266ex overview | espressif systems,” esp8266, 2017. . [5] bosch sensortec, “bme280 combined humidity, pressure and temperature sensor,” datasheet. 2015. [6] k. vandana, c. baweja, simmarpreet, and s. chopra, “influence of temperature and humidity on the output resistance ratio of the mq-135 sensor,” int. j. adv. res. comput. sci. softw. eng., 2016. [7] x. liu, s. cheng, h. liu, s. hu, d. zhang, and h. ning, “a survey on gas sensing technology,” sensors, 2012. [8] arduino, “software arduino (ide),” arduino, 2017. [online]. available: https://www.arduino.cc/en/guide/environment. [9] a. knud and l. lueth, “iot basics : getting started with the internet of things,” iot anal., 2015. https://github.com/tektonia/bluelab_iot compressed learning for text categorization compressed learning for text categorization artur ferreira1,2 mário figueiredo2,3 1instituto superior de engenharia de lisboa 2instituto de telecomunicações 3instituto superior técnico, lisboa, portugal arturj@isel.pt mario.figueiredo@lx.it.pt keywords: random projections, random subspaces, compressed learning, text classification, support vector machines. abstract: in text classification based on the bag-of-words (bow) or similar representations, we usually have a large number of features, many of which are irrelevant (or even detrimental) for classification tasks. recent results show that compressed learning (cl), i.e., learning in a domain of reduced dimensionality obtained by random projections (rp), is possible, and theoretical bounds on the test set error rate have been shown. in this work, we assess the performance of cl, based on rp of bow representations for text classification. our experimental results show that cl significantly reduces the number of features and the training time, while simultaneously improving the classification accuracy. rather than the mild decrease in accuracy upper bounded by the theory, we actually find an increase of accuracy. our approach is further compared against two techniques, namely the unsupervised random subspaces method and the supervised fisher index. the cl approach is suited for unsupervised or semi-supervised learning, without any modification, since it does not use the class labels. 1 introduction the need for feature selection (fs) and/or feature reduction (fr) arises in many machine learning and pattern recognition problems [1]. on large datasets (in terms of dimension and/or number of samples), the use of search-based or wrapper techniques can be computationally prohibitive. for instance, in text classification based on the bag-of-words (bow) or similar representations (where texts are represented by high dimensional vectors with the frequencies of a set of terms in each text) we usually have a large number of features, many of which are irrelevant (or even harmful) for the classification task in hand. in this context, fs and fr play important roles in reducing the number of features. the use of fs or fr techniques may improve the accuracy of a classifier (avoiding the “curse of dimensionality”) and speeds up the training process [1]. the literature on fs and fr is too vast to be reviewed here. comprehensive coverage of these techniques, and pointers to a vast literature, can be found in several books, namely [1], [2], [3], and [4]. 1.1 compressed learning in the past decade, there has been some interest in random projections (rp, see [5, 6, 7] and references therein) for fr. recently, theoretic support for rpbased fr (termed compressed learning – cl) was provided in [8]. cl is inspired by the compressed sensing (cs) [9, 10] framework, in which an rp matrix is used to map from the data domain to the measurement domain. the theory of cs provides conditions (on the projection matrix and the level of sparseness of the data vectors) under which this (noninjective) mapping can be inverted. some cs-based techniques for classification, based on rp, have been proposed [11]. recently, it was shown that compressed learning (cl) is possible [8]; specifically, it was proved that learning in the compressed domain is guaranteed to be, in the worst case, only slightly worse than learning on the original data domain, if the rp matrix satisfies some conditions and the feature vectors are sparse (possibly on some unknown basis). since bow representations are usually very sparse, text classification using this type of representation seems like an obvious candidate for cl. i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-1 http://journals.isel.pt/index.php/iajetc 1.2 our contribution in this work, we assess cl for bow-based text classification using linear support vector machines (svm) and several types of rp matrices. svm have been found very effective for bow-based text classification [12]. as shown in our experimental results, the classifiers obtained via cl, on significantly reduced dimensions, exhibit improved classification accuracy with respect to the classifiers trained on the original features. these results suggest that (for this type of data) the bound for cl given in [8] is pessimistic: instead of the mild decrease in accuracy upper bounded by the theory in [8], we actually find an increase of accuracy. the remaining text is organized as follows. section 2 briefly reviews the main results of cl theory and rp-based fr techniques. some fr and fs techniques, used for benchmark purposes are described in section 3. the experimental setup and the experimental results are described in section 4. section 5 ends the paper with some concluding remarks. 2 compressed learning 2.1 projection-based dimension reduction let d = {(x1,c1),...,(xp,cp)} be a labeled dataset, where xi ∈ rn denotes the i-th feature vector and ci ∈ {−1,+1} is its class label. letting a be an m × n matrix, with m < n, we obtain a reduced/compressed training dataset da = {(y1,c1),...,(yp,cp)} via yi = axi. (1) each new feature (component of y = ax) is a linear combination of the original features. many techniques have been proposed to obtain “good” (in some sense) projection matrices. in the case of random projections (rp), the entries of a are randomly generated. for reasons explained below, the following distributions yield good rp matrices: • (i) gaussian n (0,1/ √ m); • (ii) bernoulli over ±1/ √ m with equal probability; • (iii) probability mass function {1/6, 2/3, 1/6} over {− √ 3/m,0, √ 3/m}, proposed by achlioptas [5]; • (iv) probability mass function {1/(2s),1 − 1/s,1/(2s)} over {− √ s/m,0, √ s/m}, proposed by li et al [7]. notice that the bernoulli and achlioptas matrices are particular cases of (iv), with s = 1 and s = 3, respectively. choices (iii) and (iv) lead to sparse a, which may be interesting from a computational point of view. 2.2 restricted isometry properties the use of rp is inspired by the johnsonlindenstrauss lemma [13, 14], which states that, under some conditions, a set of points in a highdimensional space can be mapped down to a much lower dimensional space, such that the euclidean distances between these points are approximately preserved. a closely related concept is that of restricted isometry property (rip) [9, 10, 13]: a m × n matrix is said to satisfy the (k,ε)-rip if for any k-sparse vector x (up to k non-zeros), (1 − ε)‖x‖2 ≤ ‖ax‖2 ≤ (1 + ε)‖x‖2, (2) where ‖ · ‖ is the `2 norm. the random generation procedures described in subsection 2.1 are known to yield matrices satisfying the rip with small ε, with overwhelming probability, if m = ω(k log(n/k)), (3) that is, m is a (small) factor of ω and needs to grow only logarithmically with the dimensionality of the input patterns n. as is well-known, linear svm classifiers use only inner products between training patterns. it is thus clear that a good rp matrix should preserve these inner products, a stronger requirement than the rip. the generalized rip (grip) gives conditions under which the inner products are approximately preserved [8] (see also [15]). lemma 1 ([8]) : let a ∈ rm×n be a matrix satisfying the (2k,ε)-rip and x and x′ be two k-sparse vectors such that ‖x‖,‖x′‖ ≤ r. then, letting y = ax and y′ = ax′, (1 + ε)xt x′ − 2r2ε ≤ yt y′ ≤ (1 − ε)xt x′ + 2r2ε this lemma suggests that, if the training patterns are k-sparse and a satisfies the (2k,ε)-rip, a linear svm learnt from the compressed patterns y will be very similar to one obtained from the original patterns x, as formalized in the compressed learning bound shown in [8]. the following subsection details the compressed learning bound which is the main motivation for this work. a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 2.3 compressed learning bound theorem 1 ([8]) : let w0 be the best linear classifier in the original data domain, with low expected hinge loss h(w0) = ed [ 1 − c wt0 x ] , where the expectation is with respect to an unknown distribution d from which the training set d is assumed to have been generated. let za be the soft-margin svm trained on the compressed dataset da. then, with probability 1 − 2η, h(za) ≤ h(w0)+ o   √ √ √ √‖w0‖2 ( r2ε + log 1η p )  , (4) where r is as defined in lemma 1. this theorem shows that the svm obtained from the compressed data is never much worse than the best linear classifier in the original high dimensional space. 3 feature reduction and selection this section presents some fr and fs techniques, used as benchmark for comparison purposes with the rp-based techniques. 3.1 random subspaces the (unsupervised) random subspaces method (rsm) [16] acts by (pseudo) randomly selecting subsets of components of the feature vector, that is, it choosing axes-aligned random subspaces. a multivariate approach for fs based on rsm has been proposed [17]. the authors apply a multivariate search technique on a randomly selected subspace from the original feature space, to better handle the noise in the data on the reduced dimensionality domain. this procedure is repeated many times and the q chosen feature subsets are combined into a final list of selected features, used to train some classifier. 3.2 fisher ratio the well-known (supervised) fisher ratio (fir) for binary problems (e.g., where ci ∈ {0,1}) of each feature is defined as firi = ∣ ∣ ∣ x (0) i − x (1) i ∣ ∣ ∣ √ var(xi)(0) + var(xi)(1) , (5) where x (0) i , x (1) i , var(xi) (0), and var(xi) (1), are the sample means and variances of feature i, for the patterns of each class. the fir measures how well each feature alone separates the two classes [18]. 4 experiments in this paper, we report a set of experiments on cl for text classification based on bow representations. as mentioned above, bow representations are usually very sparse, thus being in favorable conditions for the applicability of cl via (1). 4.1 experimental setup we consider the four rp matrices described in subsection 2.1, which we will refer to as: gaussian, bernoulli, achlioptas, and li et al (with s = n). we use linear svm classifiers, provided by the entool1 toolbox, trained up to 20000 iterations. each input pattern is normalized to unitary `2 norm (original domain). we have used the following four (publicly available) bow datasets: example12, example22, dexter3, and spambase4. these datasets have undergone the standard preprocessing (stop-word removal, stemming). table 1 shows the main characteristics of these datasets,as discussed in sub-sections 2.2 and 2.3 [8]: • k̄ is the average `0 norm of each pattern; • m̂r is an estimate of m to satisfy the (k,ε)-rip condition, given by m̂r = ω(k̄ log(n/k̄)); • m̂g is an estimation of m to satisfy the (2k,ε)-rip condition, given by m̂g = ω(2k̄ log(n/(2k̄))). in the case of example1, each pattern is a 9947dimensional bow vector. the classifier is trained on a random subset of 1000 patterns (500 per class) and tested on 600 patterns (300 per class). on example2 dataset, we have 9930-dimensional bow vectors, with only 10 training patterns. the dexter dataset has the same data as example1 with 10053 additional distractor features with no predictive power at random locations, and was created for the nips 2003 feature selection challenge5. we train with a random subset of 200 patterns (100 per class) and evaluate on the validation set, since the 1zti.if.uj.edu.pl/˜merkwirth/entool.htm 2download.joachims.org/svm_light/examples 3archive.ics.uci.edu/ml/datasets/dexter 4archive.ics.uci.edu/ml/datasets/spambasebase 5www.nipsfsc.ecs.soton.ac.uk a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc zti.if.uj.edu.pl/~merkwirth/entool.htm download.joachims.org/svm_light/examples archive.ics.uci.edu/ml/datasets/dexter archive.ics.uci.edu/ml/datasets/spambasebase www.nipsfsc.ecs.soton.ac.uk table 1: example1, example2, dexter, and spambase datasets main characteristics. k̄ is the average `0 norm of each pattern. m̂r and m̂g are the estimates of m to satisfy the (k,ε)-rip and the (2k,ε)-rip conditions, respectively. dataset, n subset patterns (+1,-1) k̄,m̂r,m̂g example1, 9947 train 2000 (1000,1000) 47.1 , 253, 436 test 600 (300,300) 39.5 , 219, 383 example2, 9930 train 10 (5,5) 46.6 , 247, 429 test 600 (300,300) 39.4 , 216, 376 dexter, 20000 train 300 (150,150) 94.1 , 505, 878 test 2000 (1000,1000) 96.2 , 514, 894 valid. 300 (150,150) 93.1 , 501, 872 spambase, 54 – 4601 (1813,2788) 9.8 , 17, 20 labels for the test set are not publicly available; the results on the validation set correlate well with the results on the test set. the task of both example1 and dexter is learn to classify reuters articles as being about “corporate acquisitions” or not. in the spambase dataset, we have used the first 54 features, which constitute a bow. we have randomly selected 1000 patterns for training (500 per class) and 1000 (500 per class) for testing. the spambase task is to classify email messages as spam or non-spam. a collection of bow documents is usually represented by the term-document (td) [19] matrix whose columns hold the bow representation for each document whereas its rows correspond to the terms in the collection. the reported results are averages over 10 replications of different training/testing partition and random matrices, except on the example2 dataset in which we make no partition (the dataset has only 10 patterns). to serve as a benchmark, we compare with both the unsupervised rsm and supervised fisher index procedures, described in sub-section 3. 4.2 test set error rate figure 1 and figure 2 show the average test set error rates (over 10 replications) for the example1 and example2 datasets, as functions of the number of features m. the horizontal dashed blue line corresponds to the classifiers trained on the original data. figure 3 shows the error rate on the validation set, for the dexter dataset. finally, in order to assess the performance on lower dimensional sparse datasets, we compute the test set error rate on the spambase dataset; figure 4 plots these results. on the first three datasets we have an improvement on the error rate, after fr with any of the four probability distributions except for the li et al. distribution on example2. typically achlioptas distribution leads to lower test set error rate than gaussian and rademacher matrices, with about 1/3 non-zero 1000 2000 3000 4000 5000 6000 7000 8000 9000 4 5 6 7 8 9 10 #features (m) t es t s et e rr or r at e [% ] test set error rate on example 1 with linear svm original gauss bernoulli achlioptas li et al. fisher figure 1: average test set error rates (ten runs with different train/test partitions) for the example1 dataset of the linear svm classifier for fr based on rp and fs with fisher index. 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 20 22 24 26 28 30 32 34 36 #features (m) t es t s et e rr or r at e [% ] test set error rate on example 2 with linear svm original gauss bernoulli achlioptas li et al. fisher figure 2: average test set error rates (ten runs with different train/test partitions) for the example2 dataset of the linear svm classifier for fr based on rp and fs with fisher index. a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 8 10 12 14 16 18 20 22 24 26 #features (m) v al id at io n s et e rr or r at e [% ] validation set error rate on dexter with linear svm original gauss bernoulli achlioptas li et al. fisher figure 3: average validation set error rates (ten runs with different train/test partitions) for the dexter dataset of the linear svm classifier for fr based on rp and fs with fisher index. 15 20 25 30 35 40 45 50 55 9 10 11 12 13 14 15 16 17 18 19 #features (m) t es t s et e rr or r at e [% ] test set error rate on spam with linear svm original gauss bernoulli achlioptas li et al. fisher figure 4: average test set error rates (ten runs with different train/test partitions) for the spambase dataset of the linear svm classifier for fr based on rp and fs with fisher index. entries; the use of this distribution is efficient because the input patterns are also sparse and most of the point by point products can be avoided. the upper bound for cl defined in (4) is conservative when applied to text classification problems. the test set error rate on the reduced domain is below the test set error rate on the original data domain. moreover, an adequate value of m to achieve lower test set error than on the original domain is about 2m̂g to 5m̂g, as shown in table 1. on the spambase dataset, we get improvement with about m ≥ 40. on figure 5 we compare the performance of rpbased methods with the rsm method, combining q = 20 different subspaces of features. we have the average test set error rates for the dexter dataset, as functions of the number of features m. the horizontal dashed blue line corresponds to the linear svm classifier trained on the original data with n features, which we call the baseline error. the vertical line corresponds to the mg estimate, that is, the smallest value of m that satisfies the grip condition. the rp meth0 2000 4000 6000 8000 10000 0 5 10 15 20 25 m g # features (m) v al id at io n s et e rr or r at e [% ] dexter with linear svm original td gaussian bernoulli achlioptas li et al. rssm q=20 figure 5: validation set error rates (average over 10 random training/test partitions) for the dexter dataset, with a linear svm classifier, as functions of the number of features, using random projections and random subspaces (with q=20 subspaces). ods lead to features that are slightly better than those obtained by rsm. as compared to the rsm method, the rp method has the advantage to be faster, since it involves solely a matrix multiplication. 4.3 training time analysis the reduction in the number of features leads to a reduction of the training time. table 2 shows how the training time of the linear svm varies with the number of features for the high-dimensional dexter dataset. the decrease in the dimensionality of the data leads to reasonable improvements on the training time. table 2: analysis of the training time (in seconds) for the dexter dataset as a function of the number of features m, using rp (average of ten runs with different training/test partitions). m time [sec] 20000 3.16 5000 2.41 4000 1.99 3000 1.85 2000 1.81 a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 5 conclusions in this paper, we have reported a set of experiments on compressed learning for text classification based on (sparse) bag-of-words representations. the compressed features are obtained by (sparse) random projections. our experimental results show that four probability distributions (two of which are sparse) for the random projection matrix significantly reduce the number of features as well as the training time, while also improving the classification accuracy. we have found that the recently proved theoretical bound for the test error of compressed learning is conservative; the test set error on the compressed domain is below the test set error on the original data domain. the number of reduced dimensions can be computed by a simple sparsity analysis of the training data, regardless of the label of each pattern. the random projection technique performs slightly better than the random subspaces method and the fisher index technique, on high-dimensional datasets. in future work, we will apply this technique to semi-supervised text classification. references [1] t. hastie, r. tibshirani, and j. friedman. the elements of statistical learning. springer, 2nd edition, 2009. [2] i. guyon and a. elisseeff. an introduction to variable and feature selection. journal of machine learning research, 3:1157–1182, 2003. [3] i. guyon, s. gunn, m. nikravesh, and l. zadeh (editors). feature extraction, foundations and applications. springer, 2006. [4] f. escolano, p. suau, and b. bonev. information theory in computer vision and pattern recognition. springer, 2009. [5] d. achlioptas. database-friendly random projections. in acm symposium on principles of database systems, pages 274–281, santa barbara, california, 2001. [6] ella bingham and heikki mannila. random projection in dimensionality reduction: applications to image and text data. in proc. 7th acm sigkdd int. conf. on knowledge discovery and data mining – kdd’01, pages 245– 250, new york, 2001. [7] p. li, t. hastie, and k. church. very sparse random projections. in kdd ’06: proc. of the 12th acm sigkdd int. conf. on knowledge discovery and data mining, pages 287–296, new york, 2006. [8] r. calderbank, s. jafarpour, and r. schapire. compressed learning: universal sparse dimensionality reduction and learning in the measurement domain. [online] http://dsp.rice.edu/files/cs/cl.pdf, 2009. [9] e. candes, j. romberg, and t. tao. compressed sensing. ieee trans. inf. theory, 52(2):489– 509, 2006. [10] d. donoho. compressed sensing. ieee transactions on information theory, 52(4):1289– 1306, 2006. [11] m. duarte, m. davenport, m. wakin, j. laska, d. takhar, k. kelly, and r. baraniuk. multiscale random projections for compressive classification. in ieee int. conf. on image processing (icip), 2007. [12] t. joachims. learning to classify text using support vector machines. kluwer academic publishers, 2001. [13] r. baraniuk, m. davenport, r. devore, and m. wakin. a simple proof of the restricted isometry property for random matrices. constructive approximation, 2007. [14] w. johnson and j. lindenstrauss. extensions of lipschitz mappings into a hilbert space. in conf. in modern analysis and probability, pages 189–206, 1984. [15] j. haupt and r. nowak. a generalized restricted isometry property. technical report, university of wisconsin, madison, 2007. [16] t. ho. the random subspace method for constructing decision forests. ieee trans. pattern analysis machine intelligence, 20(8):832–844, 1998. [17] carmen lai, marcel reinders, and lodewyk wessels. random subspace method for multivariate feature selection. pattern recognition letters, 27(10):1067–1076, 2006. [18] t. furey, n. cristianini, n. duffy, d. bednarski, m. schummer, and d. haussler. support vector machine classification and validation of cancer tissue samples using microarray expression data. bioinformatics, 16(10), 2000. [19] c. manning, p. raghavan, and h. schütze. introduction to information retrieval. cambridge university press, 2008. a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc automatic acoustic scene classification automatic acoustic scene classification gonçalo marquesa, thibault langloisb aisel, electronic telecommunications and computer department, lisbon, portugal bfcul, informatics department, lisbon, portugal gmarques@deetc.isel.pt tl@di.fc.ul.pt abstract— this paper presents a baseline system for automatic acoustic scene classification based on the audio signals alone. the proposed method is derived from classic, content-based, music classification approaches, and consists in a feature extraction phase followed by two dimensionality reduction steps (principal component analysis and linear discriminant analysis) and a classification phase done using a k nearest-neighbors algorithm. this paper also reports on how our system performed in the context of the dcase 2016 challenge, for the acoustic scene classification task. our method was ranked fifteenth amongst forty nine contest entries, and although it is below the top performing algorithms, in our perspective it is still interesting to see a low-complexity system such as ours obtain fairly good performances. keywords: machine learning, signal processing, music information retrieval, bag of frames. i. introduction automatic identification of sound sources in an urban environment has a huge potential in several applications related to the current panorama of intelligent cities. these applications include monitoring systems able to recognize activities, sound environments, and create city sound maps to provide to the general public information about environmental noise, or other acoustic factors. however, a lot of research is still needed to reliably detect and recognize sound events and scenes in realistic environments where multiple sources, often distorted, are simultaneously present. this work focuses on one particular aspect of urban sound analysis: acoustic scene classification. the system we propose is a classical classification system in the sense that it uses typical machine-learning data transformation and classification algorithms in the decision making process. first, each audio excerpt is converted into a single feature vector which is the representation of choice for standard machine-learning methods. then, the whole dataset is transformed via principal component analysis (pca) [10], an unsupervised dimensionality reduction technique, followed by a linear discriminant analysis (lda) projection [9]. lda is a supervised process, and the projection tries to maximize the ratio between intra and inter class scatter, but it is not a classification method since no decision is involved. for classification, a k-nearest neighbors (k-nn) algorithm was used [7]. the experimental configuration used in our tests is common in many audio classification works (or at least parts of it – see for instance [6], [11], [14], [16], [18]) and therefore it does not bring any original contribution in terms of the algorithmic setup. in fact, our system falls under the standard “bag of frames” classifiers commonly used in music information retrieval applications (see [3], [4], [12], [17] and references therein). our main objective was not to bring forth a new audio classification or feature extraction method, but rather see how a simple, non parametric algorithm performed in the acoustic scene classification challenge. we used the same data partitioning and cross-validation setup provided with the database and our results are a bit better than the ones reported in [13]. the method in [13] was the baseline system provided with the challenge and ranked twenty eighth while our system ranked fifteenth (amongst a total of forty nine entries). however, the experiments we conducted also revealed some unexpected variations in terms of accuracy when the whole dataset or just part of it was used to estimate the pca and lda projections. this is an indication that there may be differences between feature class-dependent distributions among folds. the structure of the remainder of this paper is as follows: section ii describes the data and the feature extraction process used in our experiments, section iii describes our approach to acoustic scene classification, followed by section iv where we present our results. in this section, we also report on the challenge results and how we faired against other contestants. section v concludes this paper. ii. data and feature extraction the dataset used in this work was created in the context of the dcase2016 challenge [1] for the acoustic scene classification task. the dataset contains 1170, 30-seconds audio excerpts from the following acoustic scenes: beach, bus, café/restaurant, car, city center, forest path, grocery store, home, library, metro station, office, urban park, residential area, train, and tram. the dataset is divided into four folds for cross-validation testing. we used the same data partition in our experiments and our results are averaged over the four test folds. the features used are the all-purpose mel frequency cepstral coefficients (mfccs), a very popular representation in speech recognition (see for instance [15]), and also widely used in content-based music information applications. the audio was decomposed into 23 ms segments (1024 samples at 44.1 khz) with 50% overlap, and we used 100 mel bands to extract 23 mfccs plus the zero order mfcc and the frame’s log-energy, plus the delta and acceleration coefficients. this means that the audio is converted into a sequence of 25×3=75 dimensional vectors. we applied the voicebox software [2] to extract the features. in order to convert each audio excerpt into a single feature vector, the sequence of mfcc features is summarized i-etc: isel academic journal of electronics, telecommunications and computers cetc2016 issue, vol. 3, n. 1 (2017) id-2 http://journals.isel.pt using the median and logarithmic standard deviation. the median was used instead of the mean since this statistic is more robust to outliers. the log-standard deviation is given by 20 log10(σi) where σi is the standard deviation of feature i (with i = 1,. . . ,75). the reason to use the log-standard deviation instead of plain standard deviation was to convert these feature values to an order of magnitude comparable to the median feature values otherwise the standard deviation values would be a few orders of magnitude lower, and during the pca pre-processing step, these dimensions would be discarded as noise since they would not contribute in any significant way to the overall data variance. the statistics return two 75-dimensional vectors which are concatenated, so each audio excerpt is represented by a 150-dimensional feature vector. iii. method the proposed classification approach is divided into three main blocks: feature pre-processing via principal component analysis, feature transformation by linear discriminant analysis and finally a classification step performed by a k-nearest neighbor classifier. principal component analysis: pca is a standard dimensionality reduction technique, where the data is decorrelated by projecting it into orthogonal directions of maximum variance. these directions, the principal components, are obtained using a eigen-decomposition of the data covariance matrix, and in our experiments we kept enough components to explain 99.9% of the total data variance. the pca-transformed data was also whitened each data dimension was scaled in order to have unit variance. linear discriminant analysis: lda is commonly used as a pre-processing step for pattern classification. it is also a dimensionality reduction technique since the data is projected into c −1 dimensional space where c is the total number of classes (c=15 for this challenge). lda is a supervised learning method, and therefore the projection should be calculated with the training set only, otherwise we are indirectly including information about the class labels in the test set. estimating the lda projection with the whole dataset can result in overly optimistic performance values, specially when there is a relatively large number of classes and a relatively low number of examples, as in the case of this challenge dataset. we tested the performance of our system using the whole dataset to estimate the lda projection in order to assess the increase in performance compared to the “correct” evaluation procedure. the results showed a significant increase in performance, which in our perspective, is somewhat surprising. these are reported in section iv-b, along with a discussion on possible causes of such a performance discrepancy. k-nearest neighbors: k-nn is an instance-based learning, where class membership of a pattern is assigned based on a majority vote of its neighbors. k-nn is possibly one of the simplest classification methods, and therefore it is well suited for a baseline system. we tested two distance metrics with the k-nn algorithm, the cosine and the euclidean distance, and opted for the euclidean distance because it yielded slightly better accuracy results. we also ran the algorithm with different number of neighbors (from 5 to 31 using an increment of two) and chose empirically k = 9. the results reported in section iv are obtained using the euclidean distance metric, and k=9. iv. experimental results this section is divided into three subsections. in the first, we present the results obtained with our method. the experimental setup is described, the system performance is measured in terms of accuracy, either with mean or class specific values. in the second, we present the performances obtained when the whole dataset is used to estimate the lda projection. this is not the correct procedure to estimate our system performance. the intent is to have an idea of by how much the performance values are inflated. in the third part, we give a brief overview of the results on the dcase-2016 acoustic scene classification challenge. a. system performance the results presented in this section were obtained using the following experimental setup. the pca and the lda projections were estimated using only the training set. in our tests, we used 4-fold cross validation and the same data partitioning provided with the dataset. this means that a total of four lda projections where estimated with three training folds. the presented result pertain to the tests folds only. the obtained average accuracy was 77.6%. in table i reports the (average) accuracies per class. these contextwise performances vary from 52.6% (train class) to 93.6% (city center and metro station classes). table ii shows the accuracies per fold. table i accuracy per class. accuracy values obtained with the mean of all four test folds. 1. beach 76.9% 2. bus 66.7% 3. café/restaurant 79.5% 4. car 84.6% 5. city center 93.6% 6. forest path 87.2% 7. grocery store 82.1% 8. home 64.1% 9. library 87.2% 10. metro station 93.6% 11. office 92.3% 12. urban park 60.3% 13. residential area 65.4% 14. train 52.6% 15. tram 87.2% figure 1 shows the confusion matrix (obtained summing the four confusion matrices in each test fold). each line refers to the examples of a single class; the class order is the same as the one in table i. the columns refer to the classification results. g.marques et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-2 i-etc: isel academic journal of electronics, telecommunications and computers fig. 1. confusion matrix the rows represent the true classes, the columns represent the classification results. the class order is the same as the one given in table i: in the first row are the samples from the class beach, in the second from class bus, and so on. this matrix was obtained by the sum of the four confusion matrices one per test fold. a perfect classification would result in this matrix having the value 78 on the main diagonal (number of examples per class) and zeros for the rest of the entries. 60 0 0 0 1 2 0 0 1 0 1 9 3 0 1 0 52 6 2 0 0 1 0 5 0 0 0 0 11 1 0 0 62 0 0 2 8 0 2 4 0 0 0 0 0 0 3 0 66 0 0 0 0 1 0 0 0 0 7 1 1 0 0 0 73 0 0 0 1 1 0 0 2 0 0 1 0 0 0 0 68 0 0 0 0 2 7 0 0 0 0 0 4 0 0 0 64 0 1 9 0 0 0 0 0 2 2 7 0 0 2 0 50 8 0 6 0 0 0 1 0 0 9 0 0 0 1 0 68 0 0 0 0 0 0 0 0 0 0 0 0 2 0 3 73 0 0 0 0 0 0 0 0 0 0 1 0 5 0 0 72 0 0 0 0 2 1 0 0 1 13 3 0 6 0 0 47 5 0 0 2 0 2 0 1 3 0 0 0 0 1 17 51 0 1 0 11 8 0 0 2 4 0 0 5 1 0 0 41 6 0 1 0 0 0 0 7 0 0 0 1 0 0 1 68 for example, for the class beach, 60 audio excerpts were correctly classified, 9 were classified as the class urban park, and nine others were also misclassified. the results also reveal some particular error correlations among certain classes. for instance many residential area samples were misclassified as urban park (a total of 17 errors). five urban park pieces were attributed to residential area, however the excerpts of this class have a higher tendency to get confused with another class: forest path (a total of 13 errors). these mislabeling seems understandable since these acoustic scenes share some resemblances. another example of classes that have similar acoustic characteristics and a high number of errors between them are bus and train. other relations that seem to make some sense could be found such as the case of beach and urban park, or home and library, but further tests would be needed to determine if a real correlation exists. table ii accuracy per fold. the mean accuracy is 77.4%. these results were obtained using only the training set to estimate the pca projection. fold 1 fold 2 fold 3 fold 4 accuracy 79.3% 71.7% 82.6% 76.0% b. lda estimation with the whole dataset in this section, we discuss the results obtained when we used the whole dataset to estimate lda projection. this is not the correct testing methodology, since when doing this, we are implicitly including test label information in the model training process. the intent here is just to determine by how much the performances are over evaluated when using this incorrect experimental procedure. we performed some tests in order to have an idea of how the classification performance is affected by using just part or the whole dataset. the results (see table iii) show that there is no significant decrease in accuracy (less than 1%) when the pca projection is estimated with only the training set. nevertheless, when we applied the same methodology to estimate the lda projection, the results were significantly affected. table iv shows the accuracies per test fold. the mean accuracy is 90.8% which is 12.6% points higher than our baseline system. since lda is a supervised technique, an increase in performance is expected when the entire dataset is used, but in our perspective such a high bias was unforeseen. this is also an indication that there is some variability of classdependent feature distributions among folds. the partition process used for this dataset may be the cause, since it was based on recording location [13]. this division was done in order to avoid overestimating systems performances, since in this way, segments from a single recording are assigned to only one fold. we believe that the variability between folds is also due to the relatively low number of examples per class, and increasing the number of examples in the database will reduce this variation. figure 2 shows the confusion matrix. there is some similarities to the error patterns found in section iv-a. for instance, the residential area samples are still misclassified as urban park, urban park as forest path, and home as library. other errors though, like the confusion between bus and train classes, have almost vanished. table iii accuracy per fold. the mean accuracy is 78.2%. these results were obtained using the whole dataset to estimate the pca projection. the whole dataset was used to measure the increase in performance compared to using only the training set for the pca estimations (results in table ii). fold 1 fold 2 fold 3 fold 4 accuracy 79.0% 72.1% 82.9% 78.8% table iv accuracy per fold. the mean accuracy is 90.8%. these results were obtained using (inappropriately) the whole dataset to estimate the lda linear projection. fold 1 fold 2 fold 3 fold 4 accuracy 95.9% 85.9% 92.3% 89.0% c. dcase 2016 challenge results the dcase 2016 challenge, provided two datasets for acoustic scene classification. the first one, described in section ii, was created for developing purposes and the results reported in the previous sections were based on this dataset. a second dataset consisting of 390 audio excerpts belonging to one of 15 classes was also provided without the ground truth g.marques et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-2 i-etc: isel academic journal of electronics, telecommunications and computers fig. 2. confusion matrix obtained by the sum of the four confusion matrices one per test fold. the results were obtained using (inappropriately) the whole dataset to estimate the lda linear projection. 67 0 0 0 1 2 1 0 1 0 0 2 3 0 1 0 72 3 1 0 0 0 0 1 0 0 0 0 1 0 0 0 77 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 77 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 76 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 75 0 0 0 0 0 3 0 0 0 0 0 1 0 0 0 73 0 0 4 0 0 0 0 0 1 1 2 0 0 0 1 61 8 0 3 0 0 0 1 0 0 2 0 0 0 0 0 76 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 77 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 78 0 0 0 0 2 0 0 0 0 7 2 0 4 0 0 62 1 0 0 1 0 0 0 0 2 0 0 2 0 0 14 58 0 1 0 1 3 0 0 0 3 0 1 2 1 0 0 60 7 0 0 0 0 0 0 3 0 0 0 1 0 0 1 73 for evaluation purposes. our submission obtained an accuracy of 83.1%, and ranked 15thplace among 49 submissions (see table v for the top 20 results). the first place submission [8] achieved a 89.7% accuracy, with a fairly complex approach. their system was based on a combination of different scores obtained with distinct features such as spectrograms or ivectors [5] (using the audio stereo information), and also distinct classifiers such as deep convolutional neural networks or creating an universal background model for the ivectors using a mixture of gaussians. in the end, each audio piece was represented by 16 different scores, which were fused together in order to provide the class estimate. the majority of other submissions also used some sort of neural network (nn) for classification (recurrent-nn, convolutionalnn, deep-nn), and various sets of features. the second method of choice was support vector machines (svm), while the third was a combination of various classification algorithms through fusion or ensemble methods. in terms of algorithmic complexity, our method is by far the simplest, and in that sense, we believe that is interesting that such low-complexity system can obtain fairly reasonable performances. v. conclusion in this work, we presented a baseline for acoustic scene classification system composed of two dimensionality reduction transformation (pca and lda) followed by a k-nn classification algorithm.we trained and tested our method on the dcase 2016 acoustic scene classification dataset, and submitted it to the challenge provided by the organization. our approach was not the top ranked one: 6.6% points below the 1st place. nevertheless, the performance obtained (83.1% accuracy) is still a relatively high value, specially taking into consideration the simplicity of the method. these results also highlight the benefits of pre-processing the data with dimensionality reduction techniques. table v top 20 ranking positions in terms of accuracy scores for the acoustic scene classification task of the dcase 2016 challenge (for details see their web page: http://www.cs.tut.fi/sgn/arg/dcase2016/ task-acoustic-scene-classification.) rank acc. author classifier 1 89.7% e-zadeh et al. fusion 2 88.7% e-zadeh et al. i-vector 3 87.7% bisot et al. nmf 4 87.2% park et al. fusion 5 86.4% e-zadeh et al. i-vector 5 86.4% marchi et al. fusion 6 86.2% valenti et al. cnn 7 85.9% elizalde et al. svm 8 85.6% takahashi et al. dnn-gmm 9 85.4% kim & lee cnn-ensemble 10 84.6% han & lee cnn 11 84.1% bae et al. cnn-rnn 11 84.1% wei et al. ensemble 12 83.8% liu et al. fusion 13 83.6% liu et al. fusion 14 83.3% e-zadeh et al. cnn 14 83.3% pham et al. cnn 14 83.3% lidy & schindler cnn 15 83.1% bao et al. fusion 15 83.1% marques & langlois k-nn 16 82.3% mun et al. dnn 16 82.3% wei et al. ensemble 17 82.1% yun et al. gmm 17 82.1% rakotomamonjy svm 18 81.8% lidy & schindler cnn 19 81.3% ghodasara et al. svm 20 81.0% kong et al. dnn 20 81.0% nogueira svm in our tests, we also found a large variation in accuracies when the lda transformation was estimated using the whole dataset versus using only the training set. this is an indication that there is some variability of feature distribution among folds in this particular dataset. acknowledgement gonçalo marques thanks fi-sonic1 for supporting his research. references [1] dcase 2016 detection and classification of acoustic scenes and events http://www.cs.tut.fi/sgn/arg/ dcase2016/. [2] voicebox: speech processing toolbox for matlab (2005) by mike brookes http://www.ee.ic.ac.uk/hp/ staff/dmb/voicebox/voicebox.html. [3] j.-j. aucouturier and f. pachet. improving timbre similarity: how high is the sky? journal of negative results in speech and audio sciences, 1(1), 2004. [4] m. casey, r. veltkamp, m. goto, m. leman, c. rhodes, and m. slaney. content-based music information retrieval: current directions and future challenges. proceedings of the ieee, 96(4):668–696, april 2008. [5] n. dehak, p. j. kenny, r. dehak, p. dumouchel, and p. ouellet. front-end factor analysis for speaker verification. trans. audio, speech and lang. proc., 19(4):788–798, may 2011. 1http://www.fi-sonic.com/ g.marques et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-2 i-etc: isel academic journal of electronics, telecommunications and computers [6] s. dieleman and b. schrauwen. multiscale approaches to music audio feature learning. in 14th international society for music information retrieval conference (ismir), pages 116– 121, 2013. [7] r. duda, p. hart, and d. stork. pattern classification (2nd edition). wiley-interscience, 2000. [8] h. eghbal-zadeh, b. lehner, m. dorfer, and g. widmer. a hybrid approach using binaural i-vectors and deep convolutional neural networks. technical report, dcase2016 challenge, september 2016. [9] r. fisher. the use of multiple measurements in taxonomic problems. annals of eugenics, 7(2):179–188, 1936. [10] h. hotelling. analysis of a complex of statistical variables with principal components. journal of educational psychology, 24:417–441, 1933. [11] c.-h. lee, j.-l. shih, k.-m yu, and h.-s. lin. automatic music genre classification based on modulation spectral analysis of spectral and cepstral features. ieee transactions on multimedia, 11(4):670–682, 2009. [12] g. marques, t. langlois, f. gouyon, m. lopes, and m. sordo. short-term feature space and music genre classification. journal of new music research, 40(2):127–137, 2011. [13] a. mesaros, t. heittola, and t. virtanen. tut database for acoustic scene classification and sound event detection. in 24th european signal processing, budapest, hungary, 2016. [14] s. r. ness, a. theocharis, g. tzanetakis, and l. g. martins. improving automatic music tag annotation using stacked generalization of probabilistic svm outputs. in proc. of the 17th acm int. conf. on multimedia (acm-mm’9, new york, u.s.a., 2009. acm. [15] l. rabiner and b.-h. juang. fundamentals of speech recognition. prentice hall, 1993. [16] j. salamon and j. p. bello. unsupervised feature learning for urban sound classification. in ieee international conference on acoustics, speech and signal processing (icassp), pages 171–175. ieee, 2015. [17] b. sturm. a survey of evaluation in music genre recognition. proc. adaptive multimedia retrieval, denmark, 2012. [18] g. wu, j. zhu, and h. xu. a hybrid visual feature extraction method for audio-visual speech recognition. in 16th ieee international conference on image processing (icip), pages 1829–1832, 2009. g.marques et al. | i-etc cetc2016 issue, vol. 3, n. 1 (2017) id-2 i-etc: isel academic journal of electronics, telecommunications and computers multi-stage mixed frequency-time simulator for bandpass sampling receiver front-ends multi-stage mixed frequency-time simulator for bandpass sampling receiver front-ends pedro miguel cruz and nuno borges carvalho instituto de telecomunicações – universidade de aveiro departamento de electrónica, telecomunicações e informática, aveiro, portugal pcruz@av.it.pt nbcarvalho@ua.pt mikko e. valkama tampere university of technology, tampere, finland mikko.e.valkama@tut.fi keywords: band-pass sampling, mixed-domain simulator, software defined radio, wideband receiver. abstract: this paper address the implementation of a mixed-domain simulator for first-order band-pass sampling receivers, which is based on an initial frequency-domain signal treatment followed by a time-domain simulation scheme. one of the proposed applications for this type of receivers is to perform the spectrum sensing feature, which is required in actual and future cognitive radio approaches. some details about the multi-stage modelling strategy will be given focusing in each specific component of the receiver, wherein it is considered a mixed frequency-time signal treatment. moreover, it will be summarized the main features of the implemented simulator, as well as potential improvements. finally, several simulation examples obtained with the implemented simulator will be shown, in which are included the impact of a cw signal excitation in a received modulated signal and a multi-carrier signal reception scenario. 1 introduction the constant appearance of new communication standards (e.g. gsm, ieee 802.11a/b/g/n, umts, wimax, lte, lte-advanced) are imposing rough challenges to the wireless communication industry, either by the difficulty of integration of several standards in the same device or due to the incompatibility between each other in the different parts of the world. one of the most promising solutions for this contradictory situation is the deployment of software defined radio (sdr) [1] and cognitive radio (cr) [2] technologies, which demonstrate flexibility and agility in order to change carrier frequency, bandwidth, modulation format and transmitted power, but at the same time, respecting all the regulation for traditional radio transmitters maintaining a high energy efficiency as possible. a possible receiver architecture solution to the design of agile sdr/cr radios is the band-pass sampling receiver (bpsr), [2 4], which due to the constant advancements achieved in the analogue-todigital converter (adc) technology is becoming an attractive receiver scheme, figure 1. figure 1: band-pass sampling receiver block diagram. the bpsr architecture (figure 1) is constituted by an initial tuneable filter (bpf) or a bank of filters, then the incoming signal is amplified by a wideband low-noise amplifier (lna), and afterwards it is converted to the digital domain by using a sample and hold (s/h) circuit followed by an adc (frequently s/h and adc components are integrated in the same chip). finally, the incoming signal can be digitally processed taking advantage of digital signal processing to alleviate some mismatches of the analogue radio front end. in this sense, the main focus of this paper is to present a suitable simulator for this kind of complex design. i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-11 http://journals.isel.pt/index.php/iajetc mailto:pcruz@av.it.pt mailto:nbcarvalho@ua.pt mailto:mikko.e.valkama@tut.fi in this way, the proposed strategy uses a multistage approach to model the entire receiver chain considering independent models for each component (bpf, lna, adc, etc.). the developed simulator is able to represent the static behaviour of the complete receiver over its entire input bandwidth and under completely different input powers when excited by up to two different excitation signals. thus, in section 2, we start by giving several details about the models used for each component of the bpsr taking into account its most important attributes as, for instance, the particular band-pass sampling behaviour and the respective mixer-free down-conversion to the first nyquist zone (nz), as well as the increase of the noise floor due to thermal noise effects of the lna conjugated with clock and internal circuit sampling jitter. then, section 3 exemplifies a few facilities already implemented in the simulator and other that can be later on developed. in section 4, it will be shown the simulator functionality by giving several results of a modulated signal when interfered by a sinusoidal signal, which is varied in power and carrier frequency. moreover, the evaluation of a multicarrier signal reception will also be a case of study. finally, some conclusions will be drawn. 2 details on the models the fundamental concept of the bpsr is to design a receiving system with a reduced number of components taking advantage of the digital signal processing capabilities to achieve the required performance. a fundamental necessity for this simulator is to allow the inclusion of measured data within the simulation environment. based on this fact all the designed models are mostly based in laboratory measurements from real components. in the following it will be explained the major details about each specific component model. 2.1 filtering stage concerning on the filter component, if a static filter made of non-semiconductor devices is considered, it can be assumed linear unless some passive intermodulation (pim) is detected, [5]. this element is not considered in the developed model. in that way, the filter modelling can be characterized just by the measured frequency-domain s-parameters (for a certain low input power) performed with the help of a network analyser within the entire input bandwidth desired to evaluate. figure 2: two-port network showing incident (a1, a2) and reflected (b1, b2) waves used in s-parameters definition. thus, considering the simple 2-port network shown in figure 2, the equations that describe the network are: { (1) afterwards, these formulas are used to calculate the time-domain signal waveforms at the output of the utilized filter through an inverse fast fourier transform (ifft). as well the reflection coefficients, both at input and output ports, are also being employed to determine impairments under non-50ω sources or loads application. in addition, to match the frequency grid being used in the simulations, it will be necessary to interpolate and/or extrapolate the measured sparameters values due to the limited number of points provided by the network analyzer (instrument used to measure the s-parameters). this happens because in the commercially available network analysers it is only possible to measure frequency-domain s-parameters between a limited range of frequencies (fl to fh in figure 3), in which it is common to fail frequencies close to dc. in this way, in our models the values of the sparameters between dc and the lower measurable frequency (fl) have been linearly extrapolated. moreover, in figure 3 it is shown the processing scheme that is necessary to be employed in the measured s-parameters. here, it is assumed that measured frequency response is hermitian, i.e., the negative frequency response can be considered as the conjugate of the positive frequency response. figure 3: processing scheme employed in the frequencydomain measured s-parameters to produce a time-domain output through an inverse fft. p. cruz et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-11 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc after that, the calculated frequency-domain signal is converted to a time-domain version by means of an ifft, allowing the application of the subsequent nonlinear parts in the time-domain. it should be referred that a similar reasoning will be applied in the succeeding component models. 2.2 low-noise amplifier stage regarding the second component (lna) much has been written about nonlinear distortion in such type of devices, for instance in [6]. in small-signal operation, nonlinear behavior is often approximated by a simple polynomial, for instance a taylor series can be used when the device is memoryless. on the other hand, when in largesignal operation, the transistor starts to clip the output signal due to the fact that it will saturate and it can be approximated by a large-signal transfer function, often by a describing function approach [6, 7]. besides from that nonlinear behavior it is also important to determine non-ideal adaptations to 50ω sources and loads and thus, include into the model information about s-parameters characteristics. as a result, taking into account the previous statements, the implemented model in the simulator considers a mixed-domain signal treatment. firstly, the s-parameters of the device are measured with a common network analyzer for a given low input power, figure 4. this information is accompanied by the measured noise figure (nf) over the entire frequency range necessary to be simulated, figure 5. afterwards, it is followed by a nonlinear behaviour (figure 6) based on a power series, expression (2), which is defined based on the actual performance of the device for a certain input frequency. ( ) (2) thus, in order to implement the previous sequence of functions there is the necessity to follow the following sequential steps:  apply a fourier transform (fft) to the input signal (coming from the filter)  apply the normalized vector (complex values) of the measured s-parameters to the obtained frequency-domain signal  create a new variable of white gaussian noise with a given level (input noise floor plus maximum noise figure of lna) and filter this frequency-domain signal with the shape of the measured nf  add the noise and the input signal vectors  apply an ifft on the previous signal to obtain the correspondent time-domain version  apply the nonlinear behaviour (in this case, a power series) to the previous signal figure 4: measured forward transfer function (s21) from the lna within the bandwidth of interest. figure 5: measured noise figure for the lna within the bandwidth of interest. figure 6: measured lna nonlinear characteristic for a low input frequency over imposed with the applied 5 th -order taylor polynomial approximation. 0 20 40 60 80 100 120 140 160 180 23 23.1 23.2 23.3 23.4 23.5 23.6 23.7 23.8 23.9 24 frequency [mhz] s 2 1 [ d b ] 0 20 40 60 80 100 120 140 160 180 6 6.1 6.2 6.3 6.4 6.5 6.6 6.7 6.8 6.9 7 frequency [mhz] n o is e f ig u re [ d b ] -30 -25 -20 -15 -10 -5 -5 0 5 10 15 input power [dbm] o u tp u t p o w e r [d b m ] linear lna meas approx. p. cruz et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-11 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc in future developments of this simulator the described model could be improved and make use of other much feasible options to model the nonlinearity of an lna. for example, considering a device with memory, a volterra series [8] analysis can be considered, since it will incorporate dynamic, baseband effects and thus approximating better the overall behaviour of the lna mainly for smallsignal. as well, if we are willing to cover small and large-signal operations, system-level models as xparameters [9] should be considered. 2.3 analogue-to-digital converter stage the last component to be modelled is the wideband s/h plus adc, being in this specific component that resides the major innovation and significance to the developed bpsr simulator. the new scheme employed to model the adc component is proposed in [10] and shares a few similarities with the work presented in [11] for a sigma-delta adc. in [10] it is proposed a systemlevel model composed of different independent blocks to characterize the entire behaviour of a wideband s/h plus adc. the first block is intended to describe the rf impairments of the input circuit, and account mainly for the rf mismatch. in radio frequency designs this mismatch and frequency characteristics can be expressed by the abovementioned s-parameters. these are mainly characterized by the mismatch loss (s11) at the input matching circuit, which might demonstrate quite different performance with the frequency from the input to the output ports. the second block will model the main non-ideal processes that occur in the sampling stage of the a/d conversion. in this block it is included the degradation of the signal-to-noise ratio (snr) due to the aperture jitter of the sampling clock and adc internal jitter, equation (3). ( ) (3) then, it is followed by a relevant block that performs the mechanism of band-pass sampling, i.e., folding back any input signal inside its bandwidth to the first nyquist zone (nz), as shown in figure 7. finally, a sub-block to model the complete nonlinear behaviour of the adc is used, supported again in a power series approach. figure 7: measured transfer function characteristic (s21) of the adc, showing the concept of band-pass sampling within the first four nyquist zones. the ending block of the system-level model is the quantization segment that is designed to perform the conversion of the analogue sampled signal into an output digital word. this quantization phenomenon will originate a certain amount of quantization error, which can be represented by the value of the least significant bit (lsb). on the other side this lsb value is dependent on the reference voltage of the adc and also on its number of bits. the mentioned quantization error will impose a certain snr on the output signal, which is ideally approximated by the equation (4): (3) finally, all the blocks are interconnected and the input signal is treated based on a multi-stage mixed frequency-time scheme using a similar fashion as in the previous models. 3 simulator capabilities in our vision this simulator will be interesting to be used as an educational tool, where it would make practical the development and evaluation of new compensation schemes for imperfections and impairments present in the receiving chain, allow the study of novel mitigation methods to overcome interference problems arising from a plenty of sources, and so on. looking to figure 8 it is possible to observe the actual state of the implemented simulator. there it is seen a simulator constituted by four panels divided into an introductory panel, the configuration panel, the simulation results panel and the signal demodulation panel. p. cruz et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-11 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc (a) (b) (c) (d) figure 8: illustrative appearance of the implemented simulator: (a) initial introductory panel, (b) configuration panel, (c) simulation results panel, and (d) signal demodulation panel. the introductory panel is just the front cover for this simulator. in the configuration panel the user should insert the required parameters for each component, being available the option to execute a pre-configured example. the simulation results panel presents the input and output signals at both time and frequency domains, and is accompanied by a few more information regarding input and output powers, snr values, etc., and has the functionality to save the obtained results into a file for further use. finally, the signal demodulation panel shows the results (constellation diagram and error vector magnitude) for the demodulation of the signals present at carrier one and two. in summary, the most relevant features implemented in this simulator can be enumerated as follows:  based on matlab / simulink but it is a simple executable program (*.exe)  several signal excitations are available (onetone; multi-tone; qpsk; 16-qam; 64-qam) for one and/or two-channel operation  demodulation of i/q modulated signals with gain/phase compensation algorithm • evm for signal in channel-1 with an interference signal in channel-2  multi-carrier operation allowing the signals to be band-pass sampled and demodulated  simple compensation scheme (frequency behaviour compensation) devoted to minimize the multi-band and large bandwidth signals impairments • correction of non-flat filter attenuation and non-ideal phase performance p. cruz et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-11 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 4 illustrative simulation results in this section, it will be illustrated the functionality of the described multi-stage mixed-domain simulator by using a bpsr architecture, figure 1, using data from real components. initially, it was considered that no band-pass filter is used to select the band to be received and thus, the entire simulated bandwidth is amplified in the lna and folded back to the first nz when being sampled at the adc stage. in that sense, the first configured component was an lna that take values from a commercially available wideband (2 – 1200mhz) lna, which has a 1db compression point close to +11dbm, an approximated gain of 23db, and a noise figure near to 6db. this information can be seen in little detail in the low-noise amplifier section (2.2). this was followed by a commercially 12-bit pipeline adc that has a linear input range of around +10dbm (2vpp) with an analogue input bandwidth (-3db) of 157mhz, as can be merely understood in figure 7. the adc component has been sampled by a sinusoidal clock centred at 90mhz, with a constant noise floor and lower second and third harmonics. because of the used clock it was decided to simulate an input bandwidth of 180mhz in order to cover the first four nz’s (each one with 45mhz). 4.1 one-tone excitation the first test to evaluate the performance of this simulator was to excite the previous described dut with a sinusoidal (cw) signal centred in each nz and then, study the attainable signal-to-noise ratio (snr) and spurious-free dynamic range (sfdr). in that way, we have determined those figures of merit when the signal power at the dut input is varied from -80dbm to -8dbm with the results shown in figure 9. as can be seen the performance in the four nz’s is very similar for both snr and sfdr values when operating in a small-signal region. moreover, observing the snr results close to maximum input power is worst for higher frequencies (upper nz’s) due to adc and sampling clock jitter accounted in the model, which will degrade the final snr, see equation (3). in addition, it is noted a sudden inversion in the sfdr curve for higher input powers due to the nonlinear distortion generated in the lna and that surpasses the quantization noise level generated in the adc component. figure 9: obtained snr (top) and sfdr (bottom) values for the bpsr design when evaluated at different nz’s. 4.2 modulated signal with cw interference the second test attempts to assess the impact of a cw interference in a qpsk and 16-qam modulated signals when this interference is swept in power and varied in frequency. the spectra’s for each situation is presented in figure 10. in this simulation we maintained the qpsk (symbol rate of 5mbps) and 16-qam (symbol rate of 6mbps) signals in the 1 st nz (11.5mhz) with a fixed input power close to -40dbm. regarding the cw interference it is put in the 2 nd nz, more precisely at 69mhz, for the power sweeping case. when varying its carrier frequency it is set to an input power of -20dbm. the obtained results are illustrated in figure 11, where it can be seen that a cw interference does not cause any issue to signal demodulation but except when it saturate the adc device for an input of 10dbm. as regards to the cw frequency variation it will completely damage the signal demodulation when the input frequency is folded over the desired signal. these impacts can be also perceived in figure 12 analyzing the constellation diagrams and spectra’s. -80 -70 -60 -50 -40 -30 -20 -10 0 10 20 30 40 50 60 70 pin dut [dbm] s n r [ d b c ] 1st nz 2nd nz 3rd nz 4th nz -80 -70 -60 -50 -40 -30 -20 -10 0 10 20 30 40 50 60 70 80 90 100 110 pin dut [dbm] s f d r [ d b f s ] 1st nz 2nd nz 3rd nz 4th nz p. cruz et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-11 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 10: spectrum of the qpsk/16-qam signal with a cw interference when sweeping its input power (left) and varying its carrier frequency throughout the several nz’s (right). figure 11: error vector magnitude obtained for qpsk and 16-qam modulated signals under cw interference: sweeping input power (left) and varying carrier frequency (right). figure 12: results for cw power sweeping with constellation diagram at (a) -15dbm, (b) -11dbm, (c) -10dbm, and (d) spectra at -10dbm (fully clipped). results for cw frequency variation with (e) spectra at fcw = 80mhz being folded over the modulated signal and (f) respective constellation diagram. -80 -70 -60 -50 -40 -30 -20 -10 0 5 10 15 20 25 30 35 40 45 50 input power of cw interference [dbm] e v m [ % ] evm qpsk evm 16-qam 0 20 40 60 80 100 120 140 160 180 0 5 10 15 20 25 30 35 40 45 50 carrier frequency of cw interference [mhz] e v m [ % ] evm qpsk evm 16-qam p. cruz et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-11 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 4.3 multi-carrier operation the last test consists in the computation of the evm for two modulated signals being simultaneously received. in this situation we have used a 16-qam signal in channel one centred at 11.5mhz (1 st nz) and carrying a symbol rate of 6mbps. the channel two have a qpsk signal with 3mbps of symbol rate and situated at 128mhz (3 rd nz) being directly folded back to 38mhz. the integrated power for channel one is -40dbm and for channel two is around -20dbm. the obtained constellation diagrams are shown in figure 13 corresponding to an rms evm of 3.99% for the 16-qam signal and 0.41% for the qpsk signal. figure 13: spectra for the multi-channel operation case (top) and constellation diagrams obtained for 16-qam signal (middle) and qpsk signal (bottom). 5 conclusion in this paper a different simulator approach for band-pass sampling receivers have been described in depth, which is based on a multi-stage mixed frequency-time design. the presented simulator may become a valuable educational tool to evaluate the performance of such a type of receiving architectures that are intended for wideband sdr systems. several examples have been shown based in realworld scenarios and with the obtained results it is then possible to improve the overall signal reception by using several compensation and mitigation schemes, which may be implemented and tested on top of the provided simulator. acknowledgement the authors would like to acknowledge the support of this work within the cost action ic0803 – rfcset, and to thank portuguese science and technology foundation (fct) for the phd grant given to the first author (sfrh/bd/61527/2009). references [1] mitola, j., “the software radio architecture”, ieee communications magazine, vol. 33, no.5, pp. 26-38, may 1995. [2] vaughan, r., scott, n., and white, d., “the theory of bandpass sampling”, ieee trans. on signal processing, vol. 39, no. 9, pp. 1973-1984, sept. 1991. [3] tseng, c. and chou, s., “direct downconversion of multiband rf signals using bandpass sampling”, ieee trans. on wireless communications, vol. 5, no. 1, pp. 72-76, jan. 2006. [4] akos, d., stockmaster, m., tsui, j. and caschera, j. “direct bandpass sampling of multiple distinct rf signals”, ieee trans. on communications, vol. 47, no. 7, pp. 983-988, july 1999. [5] lui, p., “passive intermodulation interference in communications systems,” electronics & communication engineering journal, june 1990. [6] wambacq, p., and sansen, w., “distortion analysis of analog integrated circuits,” the springer international series in engineering and computer science, 1998. [7] pedro, j.c., and carvalho, n.b., intermodulation distortion in microwave and wireless systems and circuits, artech house, london, 2003. [8] schetzen, m., the volterra and wiener theories of nonlinear systems, r. e. krieger publishing, 1989. [9] verspecht, j., and root, d., “polyharmonic distortion modeling,” ieee microwave magazine, vol. 7, no.3, pp. 44-57, june 2006. [10] cruz, p.m., carvalho, n.b., and valkama, m., “measurement-based modelling of analogue-todigital converters under rf impairments,” under submission to iet circuits, devices and systems. [11] brigati, s., francesconi, f., malcovati, p., tonietto, d., baschirotto, a., and maloberti, f., “modeling sigma-delta modulator non-idealities in simulink,” proceedings of the 1999 ieee international symposium on circuits and systems, vol. 2, pp. 384387, july 1999. 4 3 2 1 0 1 2 3 4 4 3 2 1 0 1 2 3 4 in phase q u a d ra tu re rx tx 1.5 1 0.5 0 0.5 1 1.5 1.5 1 0.5 0 0.5 1 1.5 in phase q u a d ra tu re rx tx p. cruz et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-11 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc cyber physical system applied in poultry production cyber physical system applied in poultry production vitor vaz da silvaab, adepartment of engineering of electronics, telecommunication and computers, isel/ipl instituto superior de engenharia de lisboa, portugal bcts-uninova, portugal vsilva @ deetc.isel.ipl.pt abstract — a use case of poultry counting on a production line in portugal is presented as a first step for industry 4.0 application. along the poultry production line there are several transformation and translation steps in which the fowls are subjected to, and their quantity in the line may change; some are removed or fall out and others are recovered and inserted in precise places of the line. those changes can be used as an indicator of the status of the production line at that may influence the improvement of its efficiency. this paper presents a working solution of five wireless devices along a poultry production line and the complete system where data is stored locally for maintenance and supervisor decisions, and also data is stored at a remote database for management purposes which can be accessed by a mobile application. keywords: iot, industry 4.0, poultry, production line, portugal i. introduction also known by ‘smart industry’, ‘intelligent industry’, ‘smart factory’ or ‘smart manufacturing’, the industry 4.0 is related to the industrial internet, and since 2016 the industrial internet consortium (iic) and industry 4.0 platform, “plattform industrie 4.0” [1]. the portuguese strategy to develop industry in the digital area was launched in january 2017 as “indústria 4.0”. it’s aim is to put portugal at the forefront of the 4th industrial revolution by focusing on 3 axes: digitalisation, innovation and training [2]. the european commission document, digital transformation monitor, country: portugal indústria 4.0, dated may 2017 states: “concentrated on identifying the real needs of the portuguese industry, with a particular focus on smes as drivers of change, 120 portuguese companies participated in the design of the strategy. during the next 4 years 60 public and private funded measures will be implemented. the measures are divided in 6 strategic pillars: human capital qualification; technological cooperation; start-up i4.0; financing and investment incentive; internationalisation; and standards and regulation.” [2]. portuguese government released its strategy by 2017.01.30 [3]. the financial budget is of 2.26 billion euros through portugal 2020 programme of 4 years. an instrument “vale indústria 4.0” was created that offers 7,500 euros to 1500 sme (small and medium enterprises) that make a disruptive change on their model of business by contracting sites for electronic commerce or software for fabric management. the european commission expects that the results’ in portugal “impact over 50,000 companies and train 200,000 workers on digital competences” [2]. the internet of things (iot) is a generic term for all devices that can be connected to the internet and are enabled to sense or act upon the physical world [4,5]. it is shaping the development of technologies in the ict (information communication technology) domain [6]. a specific iot is the industrial iot (iiot), a cyber-physical system (cps) in which real and virtual worlds merge as a cyber-physical production system (cpps), where a secure framework is necessary [6,7]. the project presented in this paper began by the end of 2016 and the present installation has been running uninterruptedly since 2017.10.09. the place of installation is a poultry production plant in portugal but the exact place cannot be divulged. it was installed as a prototype. the motivation that led to this solution relates to the production plant in which it was installed. poultry arrives in boxes defined by their weight and not by the number of fowls. the result of the production is not only the weight but the number of fowls or eventually in sets of parts. during the production process, poultry has to be removed from the line due to quality control or they fall out because of the mechanical handling process. the poultry removed is weighted and eventually counted. the clients usually expect a number of fowls within an average weight value. (e.g.: 5000 poultry with approximately 1.5 kg each). to know the number of fowls in specific points of the plant line is essential for the maintenance and the number of poultry in and out of the system is important for the management. although this counting could eventually be made by the infrastructure that is already in the plant by adding equipment to its system, that solution could become very expensive. this document reflects the state of the art of this type of poultry production in portugal where poultry counting is made by devices off the plant’s automation and control. it is an add-on system that does not interfere with the production plant but by acting as an observer, it can produce valuable indicators for management and maintenance decisions. i-etc: isel academic journal of electronics, telecommunications and computers iot-2018 issue, vol. 4, n. 1 (2018) id-7 http://journals.isel.pt several attempts by other parties to count the fowls at this production plant failed. the only knowledge the author was aware of the difficulties encountered by those parties came from conversations with the workers that were involved in those previous projects. that led to a rough design of solutions in which video cameras, acoustic sensors and lasers were also considered; mechanical counting which involved touching the poultry was set aside due to hygienic reasons. the acoustic sensors were used in this project as a first approach mainly because of the environment, and then they were set aside because of several reasons; there were four sensors at each counting point and at that time the algorithm could not distinguish completely one fowl from the other due to overlapping when the fowls were too big. a single laser option was then implemented and it gave better results. to count the poultry and determine the line speed, two optical sensors were then chosen as the best solution. the proposed solution’s system diagram can be seen in figure 1. figure 1 – system diagram the project shown in figure 1 consists of several cyber physical devices that are connected to an intranet structure. devices send the data to a computer that gathers and stores it in a local database. a user interface at that computer shows relevant information for production line supervisor and maintenance decisions. that same computer also sends data to be stored on a remote database. the management has access to the remote database for results and graphical views. this introduction precedes the following organization of this paper: in section ii a description of the poultry production line where the pilot installation is working, the hardware components, sensors and microcontroller modules, presented in section iii cps infrastructure, sequenced by the cps logical level in section iv. details of data in graphical form and its analysis are found in section v, results and discussion. the paper ends with conclusions and future expectations, acknowledgment to the poultry facility and the people involved and references related to this work. ii.poultry production the poultry production line is composed by several different machinery that handles each fowl in sequence and individually. the poultry is hanged by the legs on hooks. all hooks are similar in form and dimensions, and are equally spaced on the same chain. there is a closed loop chain that runs around a set of machinery. whenever necessary to transport poultry from one closed loop chain to another, special machinery called transporter does the interface between those two closed loop chains. the hooks may differ in spacing and form on different closed loop chains. closed loop chains can differ very much from each other in length, from hundreds of meters to a few kilometres. on figure 2 are examples of the hooks used in this use case. a) b) c) figure 2 – poultry hooks; distance between legs x distance between consecutive hooks a) 95 x 60 mm, b) 95 x 30 mm, c) 105 x 30 mm to fit the whole chain in the production space the chain has turning points with a considerable radius to redirect the chain and hooks. at those places the hooks may oscillate as they also do when leaving machinery after being processed by it. the hooks shown in figure 2 have two kinds of markings; the one on the top is where the sensor detects the hook presence and the ones at the bottom are where the legs are supposed to be present. the hooks are positioned side by side on all of the closed loops chains (other production lines can have different hooks and can be positioned one in front of the other). the chain height from the floor may differ along the production process, as it has to be at eye level for hook uploading or extraction, and above a certain height so that a clear space for people or tall objects to go underneath, or even to be used as a temporary buffer space with the chain spiralling up or down for example for cooling purposes. chain line velocity can change along production time, a mean value for this use case is around 8500 bph (birds per hour). to maintain a continuous flow, the chain speed and hook phase is synchronized by the transporter on both chains. failure can occur on the transportation process and it may increase with the production chain speed; the fowl may fall or may hang on the hook only by one of the legs. some of the fallen poultry can be hanged up again by workers on the empty hooks, so that the sequential process for that poultry is executed. some poultry which is considered unfit by quality control can be withdrawn from the hooks or not hanged at all; they are considered as rejected. poultry can be fed into the system from different production sources. to distinguish one source from another, called a set, a gap of 10 or more empty hooks is ensured for that purpose. when removing poultry due to quality control rejection, a gap can be formed that will be identified as a set marking. although it does not affect the production line v. silva | i-etc iot 2018, vol. 4, n. 1 (2018) id-7 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt accountancy it will influence the automatic set counting, but not the total count of poultry. iii.cps infrastructure the infrastructure can be described as two main components: the sensors, which are specific to the environment and can be changed by similar ones as long as they have the same interface, and the microcontroller hardware module, that can also be changed by similar ones as it happened during the development phase. a. sensors environment can be very harsh for sensors as they must withstand not only the reality of what they are sensing as also the washing process that occurs when the place where they sense has to be cleaned very frequently. sensors may have a sparse living period; environmental damage, lose calibration, increase error or any other change that may affect its reading. photoelectric diffuse-reflective sensors were chosen. these types of sensor illuminate the target with a light beam and read the reflected light. the output is an on/off signal that indicates the success of the reflected intensity. they can be set at a distance from two to eighty centimetres. to ensure that no physical contact is made between the poultry and the sensor, the sensors are placed around 25 cm apart. the light beam is harmless if viewed unintentionally. as almost any point between and away from machinery can be set as a place for the sensors, the best places are away from the turning points or other oscillating areas and above the workers height. due to the type of sensor, the transparent surface which the light beams transverse must be free of unwanted obstacles. degradation of that surface occurs with the environmental conditions as humidity and temperature fluctuations and also the chemicals used for the environmental cleaning and washing procedures. figure 3 shows two different sensors used. a) b) figure 3 – photo reflective sensor a) visible, b) infrared two different diffused photo reflective sensors were used and shown in figure 3, a) is a visible red light [9] and b) near infrared light [10]. both have a plastic casing which did withstand the environment. the diffused visible light does give a hint of where the target is, and the infrared is invisible to the eye which makes it more difficult for good positioning. the positioning can be verified by placing a small obstacle on the place where the target should be. leds on the sensor and on the device, which is presented further down, indicate the presence of the obstacle in the sensors view. the sensors close a circuit when the target is present. the following specification is common for both sensors presented: the supply voltage is in the 10 to 30 volt range, and the maximum load current is 100 ma. a 12 v power supply is used for the sensors. the sensor output is configured as a npn which means that it must have a pull up resistor. the sensor output is equal to the power supply voltage (0 ma) when there is no object detected and it has a low voltage (about 0.5 v) when the object is present sinking at most 100 ma. the device needs two sensors to work with, and any combination of the sensors presented can be used, or any other sensor that works as the specification described above. the sensors are placed on the same vertical line and detect objects on the marks shown in figure 2. b. microcontroller module in the production line plant there is different machinery that uses wireless communication for their functioning, and there are several routers for communication purposes. the area where the sensors were set is sufficiently small for bluetooth to be used. that setup was used in the beginning so as not to interfere with the wi-fi communication infrastructure. after proof of concept, communication was changed to use the wifi infrastructure. the prototype devices built and used are shown in figure 4. a) b) figure 4 – prototype devices a) with both bluetooth and wi-fi b) with wi-fi and rtc (real time clock) module the first prototype with bluetooth is presented in figure 4a) and it is based on a ultra-low-power with fpu arm cortex-m4 mcu 80 mhz with 256 kbytes flash [11]. the software uses the mbed os-5 system and is programmed in c++ [12]. it has a bluetooth hc-05 module. at first the bluetooth seemed to be a good idea because all devices were within range, and the wi-fi system was dedicated to the already existing communicating modules of the plant. it ensured that this solutions’ bandwidth would not interfere with the existing communications. the counting was verified and it worked for some time till a decision to include another v. silva | i-etc iot 2018, vol. 4, n. 1 (2018) id-7 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt sensor outside the defined range occurred. instead of extending the range with the bluetooth solution the wi-fi option was considered and a router was installed for the use of these devices. so a wi-fi esp8266-e01[13] module was added to the and used instead of the bluetooth. an improved prototype was built and shown in figure 4b). this is based on the esp8266-e12 which has the wi-fi module and memory for code and data, which makes it a more attractive solution. also a rtc ds3231 (real time clock) was added (even though the rtc function could be implemented within the processor) [14]. the systems’ software for this latter prototype was the arduino ide (integrated development environment). the wi-fi library though was not stable enough and due to that reason the esp8266-e12 was set aside for this project. a pcb was built and the final prototype is shown in figure 5. it consists on the nucleo l432kc (cpu), the esp8266-e01 (wi-fi) and ds3231 (rtc) modules. communication between the cpu and wi-fi module is done with a serial uart at 115200 bps, and the communication between the cpu and the rtc is done through an i2c bus. the software is downloaded through the usb connection. the configuration is fixed on the present installation, but a serial interface using the usb can be used to configure the module. at this stage, ota (over the air) configuration or software download is not available. figure 5 – final prototype and in use at the installation. shown in figure 5 are the external connections, from left to right: the blue is a 5 v input, the red a 12 v input, the yellow is the connection to the legs sensor and the grey the connection to the hook sensor. the 5 and 12 v power supplies are connected to the ups (uninterrupted power supply) of the plant; otherwise a local battery could hold the system functionality for at least a day. iv.cps logical level communications on the plant can fail due to several situations as excessive traffic, inadequate configuration, new equipment installed, change of infrastructure and maintenance procedures, and many others. due to a possible lack of communication, the devices must store the values read from the sensors. with this in mind the device software is structured as shown in figure 6. each device can communicate by wi-fi through udp (user datagram protocol) and tcp (transmission control protocol) sockets. the hook flow is sent by udp to the maintenance computer (mc) viewing programme by every nth hook. figure 6 – device program structure for a configured n=40 and with a chain velocity of 9600 bph, there will be a frame every 15 seconds. that frame is for visualization purposes at the mc, and as it is sent over udp means that there is no communication recovery if the frame is not received. the device identifies a set, (interval of more than 10 empty hooks) and restarts the counting for a new set. the set boundary values are stored in local flash memory (set begin and set end) with its associated timestamp. whenever possible, a complete set data is sent to the mc through a tcp socket. the tcp ensures that there is data recovery if a communication error occurs. the set values are stored in a local database by the mc computer as shown in figure 7. figure 7 – maintenance computer (mc) view programme structure all frames sent by the device have a timestamp from the rtc. received data by the mc is stored in a sqlite database. access to previous data can be retrieved from that database and displayed at the mc. data from the local database is then sent to a remote database for storage and management purposes. the main reason for a local database instead of writing the received device data to the remote database is twofold. one that there may be communication difficulties between the plant and the remote database, and the other is that maintenance can have access to data whenever it is necessary. with this solution there is no data flow from the remote database back to the plant. the remote database is at a web server as displayed in figure 8. v. silva | i-etc iot 2018, vol. 4, n. 1 (2018) id-7 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt figure 8 – web server programme structure the web server and its database have two main purposes: to store data from the mc, and eventually other plants under the same management, and also to access the database from the management view (mv) programme. the command interface written in php (hypertext preprocessor), in figure 8, has different functions that only select data from the database and present it to the user in html (hypertext markup language). access to the command interface is for authorized users only. the system has parameters that can be changed by a configuration program only available to the maintenance officer, for example to add more devices to the production plant, replacing a faulty device by another or removing it. v.results and discussion data is produced almost continuously each production day. there are five devices installed at the production plant and for each hour in full production with a mean of 8000 bph assuming two sets per hour produces in total about 80k bytes of data. this means that one complete 8 hour shift will produce 640k a day, 15m a month and about 180 mbytes a year. sqlite database’s total capacity of 140 tera bytes would be capable of holding more than 800 thousand years of data. to do a backup of the sqlite database is very easy, it is a file and a copy can be made to an adequate type of media. the remote database backup would be controlled by the system administrator. data representation is made at the mc and mv platforms. shown in figure 9 is a detail of a sensors data in graph form. figure 9 – sample of a graphic from a device’s data a graphic using sensor data is displayed in figure 9. time resolution of the graphic is 15 seconds. the blue signal at the top is the mean chain speed. it shows a change in speed from 8600 to 8400 bph. the artefact that is clearly noted around every 11 minutes is the result of the chain having a joint that closes it as a continuous loop. the hook sensor detects a variation but the algorithm behaves correctly ignoring the sensed artefact as its lack can be seen on the count and fault signals. the lower red line indicates the number of empty hooks. the green signal indicates the continuous counting of the poultry. it seems to be a straight line because the number of faults is negligible compared to the total number of fowls; the worst case would be one less the maximum gap between sets (9 in this case) that is why the green line seems unaffected. when there is a continuous count sequence of 10 or more empty hooks (red spikes) a new set is identified. as it can be seen on figure 9, there are three complete sets. showing data as a graph and in a way that other similar graphs may be compared to can help understand how the production line is working. this kind of analysis is described with simple examples shown using parts of a graph in the figures from figure 9 to figure 11. an increase in chain speed (from 8200 to 8600 bph) as shown in figure 10 by the top blue line may have caused the increase on the number of faults, as the device whose data is shown on that figure is situated about two minutes down the production line and after a transfer point between two closed loop chains. this kind of observation, if it does correlate to the instantaneous chain speed, may help the supervisor of the production line, for example to gradually increase the speed in several steps of 100 instead of a single 400 bph increase. this is an advantage to know and analyse the state of the line. this kind of analysis could be made by a machine learning system; an entity to be considered in a future version of this system, where it would easily fit along the data stream. figure 10 – an increase in chain speed (top blue line) may have caused the increase in faults (red) about two minutes down the chain. changing the chain speed is very frequent along a day (as shown by the top blue signal line on figure 11) and it does not necessarily mean a variation in faults (loss of poultry). as it can be seen by comparing the faults (red line) between both graphs of figure 11 that shows two devices that are a considerable distance apart (several kilometres within the cooling system) where the same hook is sampled again 105 minutes after on the next device. the similarity between both figure 11 a) and b) indicate that there were no faults inside the cooling system. observing the blue signal on both figures, and using the red signals on both as a guide, it is possible to find some artefacts that are on the same place, regardless of the chain speed. these artefacts are due to the hooks that, although being of the same type, do have wearing differences among them, and could be an indicator to the maintenance for a closer inspection or replacement. a similar situation is displayed in figure 9, where the artefact shows up around every 10 minutes; it is probably due to the junction of the closed loop chain of that section. v. silva | i-etc iot 2018, vol. 4, n. 1 (2018) id-7 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt presented in figure 11 are two graphical representations of data acquired by a device a) before and b) after the cooling system. the cooling system is a continuous chain loop of several kilometres that takes about 105 minutes to complete a cooling cycle, between both devices, at about 8600 bph. a) b) figure 11 – data form a device a) before and b) after the cooling system by counting the poultry and faults at both ends of the cooling system and observing that the difference value is 0 is an indication of the good functioning of the devices and sensors and also that no poultry fell inside the cooling system. at the end of the production day an excel file is automatically produced only with the necessary results for the accountant. it also allows a date range to be chosen so that monthly or yearly results can be obtained. vi.conclusions the outcome is a pilot project in poultry industry that proved the applicability and efficiency of an iot-based system and provided interesting insights (via data analytics) to improve production and maintenance alerts. it has been in continuous operation with 5 devices every day for the past 10 months at the time of this writing. one sensor had to be substituted due to its casing. as the project is on its whole a prototype the locations in the plant where the devices and sensors were placed are not completely fixed and the sensors may move due to cleaning or other vibrating sources. cleaning and maintenance is necessary for its efficient performance. to improve the installation, sensors with metallic casing should be used. the sensors should also be within the visible spectrum and focused on the target so that a visible light line can be clear which will help positioning the sensor and aid maintenance. automatic warning of sensor malfunction can also improve its attention to maintenance. a more sophisticated representation for the visualization of the data will be available in the future, which is also one of the basic requirements of the iot elements [14,15]. knowing the state of the production line by devices that are similar in functionality helps to understand where to improve and where the challenging points are. this data, in a future version, could be correlated and analysed by machine learning algorithms to determine the quality of the chain lines and support decisions in several of the production levels. acknowledgment i am very grateful for being challenged by the poultry factory for this project, to create and develop a solution, and would also like to thank the director and maintenance officer with whom i worked close by and other workers that aided the installation and use the system daily. references [1] c. finance, “challenges and solutions for the digital transformation and use of exponential technologies,” deloitte, 2015. [2] e. commission, “digital transformation monitor country : portugal ‘indústria 4.0,’” no. may, 2017. [3] república portuguesea xxi governo constitucional, “governo lança estratégia para a indústria 4.0.” [online]. available: https://www.portugal.gov.pt/pt/gc21/comunicacao/noticia# 20170130-mecon-industria-4. [accessed: 24-may-2018]. [4] l. da xu, w. he, and s. li, “internet of things in industries: a survey,” ieee transactions on industrial informatics. 2014. [5] j. wan et al., “software-defined industrial internet of things in the context of industry 4.0,” ieee sens. j., 2016. [6] d. miorandi, s. sicari, f. de pellegrini, and i. chlamtac, “internet of things: vision, applications and research challenges,” ad hoc networks. 2012. [7] n. cam-winget, a.-r. sadeghi, and y. jin, “invited can iot be secured,” in proceedings of the 53rd annual design automation conference on dac ’16, 2016. [8] l. wang, m. tã¶rngren, and m. onori, “current status and advancement of cyber-physical systems in manufacturing,” j. manuf. syst., 2015. [9] omron, “e3f1.” [online]. available: https://assets.omron.eu/downloads/datasheet/en/v1/e94e_e3 f1_photoelectric_sensor,_compact_m18_housing_datasheet _en.pdf. [accessed: 04-jun-2018]. [10] cytron technologies, “e18-d80nk user’s manual,” 2012. [online]. available: http://synacorp.my/v2/en/index.php?controller=attachment &id_attachment=506. [accessed: 04-jun-2018]. [11] stmicroelectronics, “nucleo l432kc.” [online]. available: http://www.st.com/en/microcontrollers/stm32l432kc.html. [accessed: 04-jun-2018]. [12] mbed, “mbed os 5.” [online]. available: https://os.mbed.com/. [accessed: 04-jun-2018]. [13] espressif, “esp8266-e01.” [online]. available: www microchip.ua/wireless/esp01.pdf. [accessed: 04-jun2018]. [14] maxim-integrated, “rtc ds3231 extremely accurate i2c-integrated rtc/tcxo/crystal.” [online]. available: https://datasheets maximintegrated.com/en/ds/ds3231.pdf. [accessed: 04-jun-2018]. [15] a. knud and l. lueth, “iot basics : getting started with the internet of things,” iot anal., 2015. [16] j. lee, b. bagheri, and h. a. kao, “a cyber-physical systems architecture for industry 4.0-based manufacturing systems,” manuf. lett., 2015. project wireless sensor network architecture for tunnel monitoring project wireless sensor network architecture for tunnel monitoring 1 felipe guimarães camilo, 1,2 jorge r. beingolea garay*, 1 alexandre m. de oliveira, 1,2 sergio t. kofuji and 1 fernando matta 1 laboratory of integrated systems lsi department of electronic systems engineering lsi usp 2 interdisciplinary center in interactive technologies citi usp politechnic school, university of são paulo, cep 05508-900, são paulo, sp, brazil, phone: +55 11 30919741 {fguimaraes, jorge, amanicoba, kofuji, fmatta}@pad.lsi.usp.br keywords: tunnels, zigbee, vivaldi, directional and omnidirectional antenna, wsn. abstract: this paper presents an architecture for wireless sensor networks (wsn) operating in the 2.4ghz rf band for implementation in environments of small tunnels. the study begins with an objective description of the rf architecture and with the implementation of a rssi analysis in real scenarios. the designed modules are using directional and omnidirectional antennas for each test scenario. in initial experiment is included a test on the 433mhz band. the developed wsn architecture provides a higher degree of reliability at environments with denser structures (tunnels) and enables the use of directional and omnidirectional antennas for better signal behavior considering the structure of environment to propagation. 1 introduction at present, the need to preserve the environment is vital to our planet. the continual degradation of the environment has triggered terrible environmental disasters converting public places in areas of risk of disaster such as the flooding of towns, overflow in tunnels and others, representing a high risk to city inhabitants. the technology presents itself as an important tool helping at these numerous issues that need to be answered, how to watch extensive forests [1], prevent fires [2], floods [3] landslides and avalanches. one way to prevent these disasters could be based on a continuous monitoring of nature and its phenomena through a wireless sensor networking (wsn) capturing real-time information that would serve to predict or anticipate the outcome of natural phenomena of catastrophic impact [4]. this monitoring provides a simple and instant response to any change in the environment considering parameters such as temperature, acceleration or co2 concentration, providing objective and constant information to a diagnostic center [5] enabling on time decision-making to prevent significant damage and loss of human life. however, despite these statements and the numerous advantages of the architectures of rssf, it is necessary to efficiently determine the type of architecture to be used for each environment to be monitored, considering the power consumption, robustness of the sensors hardware and suitable frequency range to operate in the most diverse and aggressive environments for monitoring [6]. in this respect, the present work aims to present a sensor-nodes architecture followed by a specific study based on received signal strength (rssi) at 2.4ghz for application in small tunnels considering technical characteristics needed to ensure the quality and reliability of wireless communication at such environment. this paper is organized in the following topics: initially, is presented, the justifications for this work. in section 2, details are given of rf *address correspondence to this author at the laboratory of integrated systems-lsi-department of electronic systems engineering, polytechnic school, university of são paulo, cep 05508-010, são paulo, sp, brazil; tel: +55 11 30919741; e-mail: jorge@pad.lsi.usp.br i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-18 http://journals.isel.pt/index.php/iajetc communication and the various frequency bands. in section 3, there is an approach to communication in tunnels and its technical fundaments. in section 4, are described the previous steps to the execution of the experiments, such as settings, hardware and test scenarios. finally in section 5, the conclusions are presented. 2 rf communication the growth of rf communication and its application in various technological fields have resulted in a rapid saturation of the communication bands, in some cases in an irregular manner, however, it is necessary to argue that even though the free frequencies do not require licenses, the devices used at these frequency bands are ruled by laws and regulations that vary from country to country. the regulations specify aspects of radio such as the allowed spectrum and power for rf systems which falls in the technical specifications of the equipment. 2.1 overcrowding at the free bands the exponential growth of electronic devices with wireless communication has resulted in a great diversity of its use and as a consequence the congestion of most bands of free use, grouped as follows: a) frequency band 433mhz; free band in europe and several other countries as a result it presents the problem of band congestion to be considered. this is considered a narrowband [7] and is not designed to have interference resistance to adjacent or coexistent channel. the use of this frequency is subject to a special regulation [8] where the equipment using this frequency must follow certain requirements with the main objective of reducing its congestion. b) frequency band 868mhz; its use is more restricted or the same as the standard norm, as a consequence it has a lower level of interference and or congestion. the 868mhz band is divided into several sub-bands which are strictly reserved for particular functions [9]. c) frequency band 2.4ghz; this band has become quite popular and the requirement is that all devices using it should consider coexistence in its construction; it must have the ability to function in the presence of other devices operating in the same frequency. the 2.4ghz band, due to the features mentioned above, is much more congested than any other free frequency band which has caused the performance, quality and reliability of the transmissions to be questioned. the free use of this frequency eventually ended up demanding the application of some technical specifications to reduce its congestion and its problems of interference, in general, the devices manufactured to operate in this band should meet the regulation [7] considering additional features such as signal filtering and frequency hopping (fhss). 2.2 rf regulation the operation of the rf communications and electromagnetic compatibility are a new, modern and technical field that has experienced rapid growth in the brazilian telecommunications scenarios well as internationally possibly as a result of the high complexity of the electronics telecommunication designs in which electromagnetic compatibility (cem) is present. the standard en 301 783 corresponds to the european standard while approved in brazil, which standardizes the use of frequency bands amateur radio bands from 420mhz up to 440mhz and 2.3ghz to 2.4ghz. rf devices used within this band must meet the technical requirements described in this standard [7]. 3 communication in tunnels of small and large size the design and choice of rf hardware for specific applications should always follow a meticulous study of the characteristics of the environment as well as of the objectives of the application becoming quite often a difficult task because of fact that the propagation phenomena adopt distinct characteristics in each environment. this paper considers the architecture for wireless sensor, firmware and analysis of rssi, a clarification of the technical details a rf communication in tunnels to small or large size. the communications in small and large tunnels features peculiar characteristics and behavior. to model the path loss is not a trivial task, being necessary to consider for this work some of the mechanisms that regulate the propagation of radio signal, which can be grouped into three categories: reflection, diffraction and scattering. the first occurs when the electromagnetic wave reaches the surface f. guimarães camilo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-18 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc of an object with dimensions much larger than the wavelength. the second occurs when there is an object with sharp edges between the transmitter and receiver. and finally the third occurs due to multiple objects of small size between transmitter and receiver. some of the basic models for "path loss" follow: a) free space model: used to predict the propagation of radio signal when the path loss between the transmitter and receiver is free and unobstructed (1). being pr (d) the power of signal received by receiver placed at a d distance from the transmitter, gt is the gain of the transmitter, gr is the receiver gain, l is loss factor of the system not related to propagation, and the wavelength in meters. this model illustrates the case where there is only the direct path between the transmitter and receiver, as a result ends up being in most cases, inaccurate. b) radius model: it is a more accurate model for which considers the direct and the reflected path in the floor space between the transmitter and receiver [10]. (fig. 1). figure 1: representation model of 2 rays in the testing scenario is referenced model (figure 1) where the received power distribution is given as a function of distance by the following formula (2): ) where ht is the height of the transmitter antenna and hr the height of the receiver antenna. if the distance between the transmitter and the receiver is relatively large ( ), the characteristics of the radio modules are abstracted by the following simplified formula (3): where ct (t for “two-ray”) is a constant and depends on the characteristic of the radio. c) estimation of losses in indoor environments: indoors propagation, such as in tunnels, large or small are much more complex than in the open because of an increased number of obstacles which commonly have dimensions that are closer to the length of the propagated signal wave and at the same time the presence of walls and floors having various features combined to be consider in the complexity of the calculations in each of the propagated test signal. this work considered not only the average loss calculus, but also a description of various attenuation components which combined impact on the propagated waves and on the power level (4). where n is the rate of power decay, the frequency in mhz, the value of total losses, distance in meters (d> 1 m), is the loss factor due to ground penetration and is the number of floors between transmitter and receiver. for a zero value of the transmitter and receiver are considered on the same level. and for n = 20, this model is identical to a free space model for outdoor environments. indoors studies requires n = 18 for a path with losses between transmitter and receiver (an exponent of "path loss" equal to 1.8). propagation f. guimarães camilo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-18 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc around corners and walls require n = 40. in the case of long distances, the reflected paths can interfere again and result in n = 40. 3.1. tunnels of small dimensions (radius <2.5m) the propagation of electromagnetic waves in small underground tunnels have been being modeled as a large and imperfect waveguide, whereas basically a waveguide is a structure with conductive walls that support transverse electric waves (te) and transverse magnetic (tm). these waves create electric and magnetic fields perpendicular to the direction of propagation. the reinforced concrete of the tunnel, however,does not maintain the theory that the walls behave as nearly as perfect conductors, in practice, the wall acts more like an insulation than as conductor. at the characterizing of the propagation scenario of a tunnel, it is necessary to consider that the dominant mode of the waveguide will have the lower cutoff frequency [11]. when no energy is propagated throughout the system a standing wave exists below this cutoff frequency. for a circular waveguide, the cutoff wavelength is the diameter multiplied by 1.71 for the dominant mode te11. at tunnels used for this purpose, the inner diameter was 1.61 m. considering primarily a circular waveguide, the cutoff frequency is approximately (3 × 108m / s) (1.06m) (1.71) or 165.5 mhz, expanding to the exterior concrete and assuming a dielectric constant of value 6, the tunnel frequency cut-off l is reduced by a factor of 16 to 67.5 mhz. using the waveguide model is expected that losses follow the propagation model of a hybrid tunnel, increasing losses with increasing distance between the transmitter and receptor. 3.2. tunnels to large dimensions (radius> 2.5m). the propagation characteristics described for small tunnels are not different for large tunnels [12] where to determine which type of rf architecture to implement a rssf is necessary to know the behavior for each type of frequency within the tunnel. 4 experiments 4 .1 . in itia l con sid era tion s it is increasingly common to need to conduct experiments in real scenarios in order to be able to capture the phenomena and errors that occur in the iteration of the wireless transmitter hardware (rf) and the propagation scenario. in this paper two real scenarios have been used for the experiments yielding great accurace results that can serve as aid in future implementations of wireless communication systems in restricted environments such as small tunnels. a) hardware rf: the prototypes built for the experiments in real scenarios base on the use of receiver / transmitter fm-trx2-24g modules of quasar-uk. [13], it contains a cc2500 rf chip from texas instruments which is an rf module for low power consumption [14], consistent with other rf transmitters operating on the zigbee stack using the same frequency band. the baud rate used oscillates between 1.2 and 500 kbps according to the type of modulation used. the transmitter operates between 2.4 2.483 ghz with a sensibility on receptor of up to -104 dbm. (fig. 2) (fig. 3). it should be noted that the choice of the transceiver modules was not solely based on economic factors, but also in technical characteristics which should be considered indoors with very dense structures (reinforced concrete). these specifications should consider the phenomena in signal propagation communication offered by the environment, a relevant feature was to considered the fhss (frequency hopping spread spectrum) that not only makes the signal more resistant to interference, but also enables the sharing of bandwidth with many other types of conventional transmitters with minimum interference, which is an essential feature in the 2.4ghz band due to factors of signal sensitivity and high congestion. f. guimarães camilo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-18 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 2: developed prototypes 2.4ghz figure 3: prototypes omnidirectional antenna in addition in the a scenario (described in section 4.2) a hardware operating at the frequency of 433mhz, the hardware produced by minteos company is composed of a radio chip from texas c1150 [15], and a 8-bit and 2 kb of programmable flash memory microcontroler atmel attiny2313 / v [16]. the base station is composed of processor atmega 1280v [17] with 4 kb of eeprom. the radio base station chip is also a c1150 texas instruments (fig 4)(fig. 5). figure 4: station base device 433mhz figure 5: device sensor 433mhz b) vivaldi antenna: this antenna belongs to the family of aperiodic radiators, continuously scaled by an exponential curve. one can say that in theory a vivaldi antenna has a characteristic semiuniform irradiation over an ultra-wide frequency range [18][19][20][21]. based on these characteristics, it was decided to choose the vivaldi antenna, which can be defined as a small planar antenna consisting of a microstrip line transition (micro transition line mtl) and a radiator in exponential opening (exponential line slot radiator eslr) respectively detailed in the following sections:  microstrip transition line: we can define the mtl as a impedance transition line constituted by a separate tape conductor, electrically isolated from a conducting surface (ground plane) by means of a structural substrate as illustrated in figure 7. this substrate usually has a high dielectric constant (ε> 2) and separates the ground plane from the conductor tape, both consisting of thincopper-film [21]. (fig 6). f. guimarães camilo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-18 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 6: microstrip transition line where h and t are the substrate and srtip conductor thickness respectively, and w is the strip conductor width [21]. figure 7: illustration of the analysis of cross-sectional view of an mtl with detail propagating quasi-tem [21] in a structure of the type microstrip line can be seen that the field lines are not contained entirely inside the substrate (fig. 7), as would happen in a waveguide, so that the mode of propagation on the microstrip line is not purely transverse electromagnetic (tem), thus defines the propagation as quasi-tem [21]. according to studies conducted by [21], if the relation between w h is greater than 1, the characteristic impedance (z0) and the effective dielectric constant (εeff) of a microstrip transmission line can be calculated by: (5) (6)  irradiation in opening exponential; after performed the calculations and studies to ensure the best impedance matching between the circuit, connectors, cables and mtl, the goal was to design the best radiator between the mtl and the propagation medium (air) in a manner in which is possible to ensure a naturally smooth transition of the signal throughout the rf link. thus it was considered using a radiator with exponential opening, which is able to ensure an impedance bandwidth relatively large [20][21][22]. figure 8 illustrates the exponential aperture and its respective function registered in the (x, y) as well as the start and end points of the curve for the design of this opening that forms the proposed electromagnetic radiator. figure 8: detail of the exponential curve of vivaldi antenna [de oliveira ,2012]. the curve of opening exponential radiator of the vivaldi antenna can be obtained according to [20], using an exponential function given by equations (7), 8) and (9), having as its start and end points p1(x1,y1) e p2(x2,y2), (7) where (8) and (9) c) firmware: the work and experiment are not limited solely to the development of a prototype for testing, but also the development of an embedded microcontroller firmware for the wireless modules. the firmware in question was developed entirely in c language with a total of 50 pages of code, and can be considered the brain of the modules (prototypes). it should be noted that the flexibility of firmware code allows modifying settings like: transmission frequency (2.4ghz 2.43ghz), output power (tx) throughput bit, parameters that must change in each application environment. the code segment responsible for setting up some of the mentioned parameters is shown in figure 9. f. guimarães camilo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-18 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 9: firmware code segment for setting communication parameters. 4 .2 . c onf ig u ra tion s the experimental network (figure 10) consists of two rf modules (transmitter / receiver) with the characteristics described above, and a programmable gateway with a sensitivity of -101 dbm, transmission rate of up to 250 kbps connected to a base station (used for collecting and filtering data). figure 10: distribution of modules in scenario b vivaldi antennas in each of the network modules (transmitter gateway) is used a micro-controller unit to command the rf module embedded. the microcontroller used is attiny26 [23] which has characteristics of high performance, low power consumption, advanced risc architecture, programmable eprom 128 \ 256 \ 512 bytes, among other technical features that maximize the performance of the prototypes developed for this work. additionally is used a module avr avrrisp kii [24] for programming the microcontrollers attiny26. 4 .3 . s cena rio o f e xp erimen ts a nd r esu lts the scenario a, corresponds to an area of approximately 100m2 belonging to the company minteos-italy in a narrow corridor to simulate the conditions of a tunnel, the performed tests corresponds to rssi values for modules over 2.4ghz in conditions with obstacles. the distance between the transmitter and receiver module is 15m, for this test is included an antenna rubbr duck (omnidirectional) [25] coupled to the transmitter and at the receiver an antenna pcb 2.4 ghz. the test was carried out to check the efficacy of the antenna coupled to the circuit (the nominal gain was equal to the actual gain), the measures being made with all the same parameters, just adding to the antenna to the circuit (fig. 11). f. guimarães camilo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-18 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 11: graphic rssi signal with antenna and without antenna. 2.4ghz figure 12: graphic rssi signal, frequency of 433mhz. the test verified the proper functioning of the circuit and coupling of the antenna, since the real gain is equal to the nominal. it is also proved that the rssi signal power at 2.4ghz (figure 11) was greater than the transmission at 433mhz (fig. 12) under the same conditions. this result proves the premise of this study, in which the transmission made at higher frequencies have better efficiency when the mediums of transmission are small tunnels or places with the same geometrical conditions. the scenario – b, corresponds to an area of 100m2, comprising one of the corridors of the building of the electrical engineering department from the polytechnic school of the university of são paulo brazil, the modules are distributed randomly with a separation distance of 10m between the modules and concentrator (gateway) and with a distance of 1m from the ground. for this test was used the antipodal vivaldi uwb antenna (fig.13) developed by the group pad in the laboratory of integrated systems of the polytechnic school usp. figure 13: geometry proposed of antipodal vivaldi antenna – frequency 2.4ghz in figure 13, = width of opening, = width of the irradiation plan, = width of the transmission line or microstri, = length of transmission microstrip and = major axis of the ellipse. figure 14: graphic rssi signal with directional vivaldi antennas and without obstacles f. guimarães camilo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-18 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 15: graphic rssi signal with directional vivaldi antennas and with obstacles the results of the tests (figures 14 and 15) verifies a considerable variation of the received power when obstructions are introduced between the receiver and transmitter. at the same time it was demonstrated the possibility of hardware circuit to be optimized so that the losses are reduced, higher gains and a longer range to be obtained being able to reach a power of110db at a distance of approximately 80m. 5. conclusions this article began with a study on characteristics of the transmission utilizing frequency bands free for use in europe: 433mh and 2.4ghz. the study aimed to identify the frequency to bring better results in transmissions in small tunnels to develop a product with the best results. after studies and initial tests it was decided for the frequency of 2.4ghz, which was more appropriate for this medium in a way in which it behaves as a waveguide within the structure of the test scenarios. after selecting module the firmware was developed. the experiments indicated successful transmission at 2.4 ghz, and the at the tests within the company minteos-italy, a higher power received in the case of 433mhz, confirming the initial hypothesis of the study. 2.4 ghz tests proved very effective considering the rssi value obtained for scenarios that use omnidirectional antenna (duck rubbr) and directional (vivaldi) for the second type of antenna as there is great need for communication over-sight throughout the longitudinal path of the tunnel, the vivaldi antenna was chosen, since it has the favorable characteristics for the link-sight application applied at communications in tunnels, namely: simple geometry (a plate of composite insulator with the faces covered with conductive film), low weight, high bandwidth, high efficiency, small size and high gain [20][26] and showed favorable results in scenarios with and without obstacles, leaving open the possibility to tailor-off the firmware developed to zigbee modules for this type of directional antennas. references o aro y or t , g.; , "case study of a simple, low power wsn implementation for forest monitoring," electronics conference (bec), 2010 12th biennial baltic , vol., no., pp.161-164, 4-6 oct. 2010. [2] takeuchi, s.; yamada, s.; , "monitoring of forest fire damage by using jers-1 insar," geoscience and remote sensing symposium, 2002. igarss '02. 2002 ieee international , vol.6, no., pp. 32903292 vol.6, 2002. [3] jong-uk lee; jae-eon kim; daeyoung kim; poh kit chong; jungsik kim; philjae jang; , "rfms: real-time flood monitoring system with wireless sensor networks," mobile ad hoc and sensor systems, 2008. mass 2008. 5th ieee international conference on , vol., no., pp.527-528, sept. 29 2008-oct. 2 2008. [4] di tada, n.; large, t.; , "information system to assist survivors of disasters," digital ecosystems and technologies (dest), 2010 4th ieee international conference on , vol., no., pp.354-359, 13-16 april 2010 [5] schmiing, m.; afonso, p.; tempera, f.; santos, r.; , "integrating recent and future marine technology in the design of marine protected areas the azores as case study," oceans 2009 europe , vol., no., pp.1-7, 11-14 may 2009 [6] garay, jorge r.b and s. kofuji, s. “an evaluation of the networks topologies star and mesh of the zigbee standard: analysis in real n ironment ” in: 5º contecsi international conference on information systems and technology management, são paulo, p. 1932-1934, 2008. [7] frequency standards 420mhz ate 440mhz e 2.4ghz. available in: http://www.ancom.org ro access in 15 de january 2011. [8] norma frequency allocations range 9 khz to 275 ghz. available in: http://www.itu.int/itu-d/study_groups/sgp_20022006/jgres09/cept2.pdf. access in: 15 of january 2011. f. guimarães camilo et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-18 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc [9] frequency standards 800mhz. available in: http://www.rfm.com/company/etsi.pdf. access in 10 of january 2011. 0 paolo santi “topology ontrol in wirele ad o and en or network ” ac comput sur v 37. pp 164-194. 2005. [11] kjeldsen, e.; hopkins, m.; , "an experimental look at rf propagation in narrow tunnels," military communications conference, 2006. milcom 2006. ieee , vol., no., pp.1-7, 23-25 oct. 2006. [12] y. wu, m. lin, i.j. wassell, "modified 2d finite-difference time-domain technique for tunnel path loss prediction," 2nd international conference on wireless communication in underground and confined areas, val-d’or canada., aug 2008. [13] devices rf. datasheet. available in: http://www.quasaruk.co.uk/acatalog/fm_transceive r_module.html [14] rádio cc2500 – texas instrument. datasheet. available in: http://focus.ti.com/lit/ds/symlink/cc2500.pdf [15] rádio cc1150 da texas instrument. datasheet available in: http://focus.ti.com/lit/ds/swrs037a/swrs037a.pdf. acessado em 15 de janeiro de 2011. [16] atmel microcontrolador attiny 2313/v. datasheet available in: http://www.atmel.com/dyn/resources/prod_documen ts/doc2543.pdf. access in 15 de january of 2011. [17] atmega 1280v. datasheet available in: vhttp://www.atmel.com/dyn/resources/prod_docume nts/doc2549.pdf. access in 15 de january of 2011. [18] greenberg, m. c.; virga, l.; hammond, c. l. performance characteristics of the dual exponentially tapered slot antenna for wireless communication application, ieee trans. on vehicular technology, v. 52, p.305-310, 2003. [19] mehdipour, k.; aghdam, m.; dana, r. f. complete dispersion analysis of vivaldi antenna for ultra wideband applications. progress in electromagnetics research, pier, v. 77, p.85-96, 2007. [20] yang, y.; wang, y.; fathy, a. e. design of compact vivaldi antenna arrays for uwb see through wall applications. progress in electromagnetics research, pier n. 82, p.401-418, 2008. [21] de oliveira, a. m. sistema transmissor cmos de radar uwb por varredura eletrônica com arranjo de antenas vivaldi. 2012. 170 f. dissertation (m.sc.) – universidade de são paulo, são paulo, 2012. [22] gazit, e. improved design of the vivaldi antenna. iee proceedings, v. 135, n. 2, p. 89-92. 1988. [23] microcontroller attiny26. atmel. datasheet. available in: http://www.atmel.com/dyn/resources/prod_documen ts/doc1477.pdf [24] avr avrrisp kii. atmel. datasheet. available in: http://www.atmel.com/dyn/resources/prod_documen ts/doc8015.pdf [25] antenna rubbr duck 2.4ghz. datasheet available in: http://www.globalspec.com/datasheets/2762/lcom/ 705f5db6-aa59-43f9-a8fe-26b847fb5176 [26] shin, j.; schaubert, d. h. a parameter study of stripline-fed vivaldi notch-antenna arrays, ieee trans. on antennas and propagation, v. 47, n. 5, 1999. photodetection, self amplification and demux operation in tandem amorphous si-c devices photodetection, self amplification and demux operation in tandem amorphous si-c devices m. vieira1,2,3, p. louro1,2, m. a.vieira1,2, a. fantoni1,2, m. fernandes1,2 1 electronics telecommunications and computer dept, isel, lisbon, portugal 2 cts-uninova, lisbon, portugal 3 dee-fct-unl, quinta da torre, monte da caparica, 2829-516, caparica, portugal keywords: optical devices, a-sic heterostructures, optical communication, multiplexing and demultiplexing applications over pof. abstract: in this paper we report the use of a monolithic system that combines the demultiplexing operation with the simultaneous photodetection and self amplification of the signal. the device is a double pi’n/pin a-sic:h heterostructure with optical gate connections for light triggering in different spectral regions. results show that when a polychromatic combination of different pulsed channels impinges on the device the output signal has a strong nonlinear dependence on the light absorption profile, (wavelength, bit rate and intensity). this effect is due to the self biasing of the junctions under unbalanced light generation of carriers. self optical bias amplification under uniform irradiation and transient conditions is achieved. an optoelectronic model based on four essential elements: a voltage supply, a monolithic double pin photodiode, optical connections for light triggering, and optical power sources for light bias explains the operation of the optical system. 1 introduction there has been much research on semiconductor optical amplifiers as elements for optical signal processing, wavelength conversion, clock recovery, signal demultiplexing and pattern recognition [1]. here, a specific band or frequency need to be filtered from a wider range of mixed signals. active filter circuits can be designed to accomplish this task by combining the properties of high-pass and low-pass into a band-pass filter. amorphous silicon carbon tandem structures, through an adequate engineering design of the multiple layers’ thickness, absorption coefficient and dark conductivities [2] can accomplish this function. wavelength division multiplexing (wdm) devices are used when different optical signals are encoded in the same optical transmission path, in order to enhance the transmission capacity and the application flexibility of optical communication and sensor systems. various types of available wavelength-division multiplexers and demultiplexers include prisms, interference filters, and diffraction gratings. currently modern optical networks use arrayed waveguide grating (awg) as optical wavelength (de)multiplexers [3] based on multiple waveguides to carry the optical signals this paper reports results on the use of a double pi’n/pin a-sic:h wdm heterostructure as an active band-pass filter transfer function whose operation depends on the wavelength of the trigger light and applied voltage and optical bias. the dynamic response can range from positive feedback (regeneration) under positive bias, to two different behaviours under negative bias: as an active multiple-feedback filter with internal gain or in a mode that preserves the amplitude of the signal, depending on the triggering light. an optoeletronic model gives insight on the physics of the system. 2 experimental details glass pin 1 (a-sic:h) 200 nm pin 2 (a-si h) 1000 nm tco tco tco applied voltage light λ1 λ2 λ3 figure 1. device configuration. the sensor element is a multilayered heterostructure based on a-si:h and a-sic:h produced by pe-cvd at i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-12 http://journals.isel.pt/index.php/iajetc 13.56 mhz radio frequency. the configuration of the device, shown in figure 1, includes two stacked p-i-n structures (p(a-sic:h)í'(a-sic:h)-n(a-sic:h)-p(asic:h)-i(a-si:h)-n(a-si:h)) sandwiched between two transparent contacts. the thicknesses and optical gap of the front í'(200nm; 2.1 ev) and thick i(1000nm; 1.8ev) layers are optimized for light absorption in the blue and red ranges, respectively [4]. experimental details on the preparations, characterizations and optoelectronic properties of the amorphous silicon carbide films and junctions were described elsewhere [5]. as a result, both front and back structures act as optical filters confining, respectively, the blue and the red optical carriers. the device operates within the visible range using as optical signals the modulated light (external regulation of frequency and intensity) supplied by a red (r: 626 nm; 51μw/cm2) a green (g: 524 nm; 73μw/cm2) and a blue (b: 470nm; 115μw/cm2) led. additionally, steady state red, green and blue illumination (background) was superimposed using similar leds. 3 light filtering the characterization of the devices was performed through the analysis of the photocurrent dependence on the applied voltage and spectral response under different optical and electrical bias conditions. the responsivity was obtained by normalizing the photocurrent to the incident flux. to suppress the dc components all the measurements were performed using the lock-in technique. 400 500 600 700 800 0.0 0.1 0.2 0.3 0.4 pi(a-sic:h)n/ito/pi(a-si:h)n -10 v +3 v p ho to cu rr en t ( μ a) a) 400 500 600 700 800 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 back cellfront cell ito/pi´(a-sic:h)n/ito/pi(a-si:h)n/ito p ho to cu rr en t ( μ a) wavelength (nm) 0.0 0.5 1.0 1.5 2.0-5v +1v b) figure 2. a) p-i’-n-p-i-n spectral photocurrent under different applied voltages b) front, p-i’ (a-sic:h)-n, and back, p-i (asi:h)-n spectral photocurrents under different applied bias. figure 2a displays the spectral photocurrent of the sensor under different applied bias (+3v<v<-10v), the internal transparent contact was kept floating in all measurements. in figure 2b the spectral photocurrent, under different electrical bias is displayed for the front, pi’ (a-sic:h)-n, and the back p-i (a-si:h)-n, photodiodes. results confirm that the front and back photodiodes act, separately, as optical filters. the front diode, based on a-sic:h heterostructure, cuts the wavelengths higher than 550nm while the back one, based on a-si:h, cuts the ones lower than 500nm. each diode, separately, presents the typical responses of single p-i-n cells with intrinsic layers based on a-sic:h or a-si:h materials, respectively. since the current across the device has to remain the same, in the stacked configuration, it is clearly observed the influence of both front and back diodes modulated by its serial connection through the internal n-p junction. 4 self optical amplification 4.1 self bias amplification under uniform irradiation when an external electrical bias (positive or negative) is applied to a double pin structure, its main influence is in the field distribution within the less photo excited sub-cell [6]. the front cell, under red irradiation; the back cell, under blue light and both, under green steady state illumination in comparison with thermodynamic equilibrium conditions (dark), the electric field under illumination is lowered in the most absorbing cell (self forward bias effect) while the less absorbing reacts by assuming a reverse bias configuration (self reverse bias effect). in figure 3, the spectral photocurrent at different applied voltages is displayed under red (a), green (b) and blue (c) background irradiations and without it (d). results confirm that a self biasing effect occurs under unbalanced photogeneration. as the applied voltages changes from positive to negative the blue background enhances the spectral sensitivity in the long wavelength range. the red bias has an opposite behavior since the spectral sensitivity is only increased in the short wavelength range. under green background the spectral photocurrent increases with the applied voltage everywhere. m. vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-12 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc a) b) c) d) figure 3. spectral photocurrent under reverse and forward bias measured without and with (λl) background illumination. in figure 4 the ratio between the spectral photocurrents @ +1v and -8 v, under red, green and blue steady state illumination and without it (dark), are plotted. 400 450 500 550 600 650 700 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 blue background green background red background +1 v -8 v i a c ( λ l )/ i( da rk ) wavelength (nm) figure 4. ratio between the photocurrents under red, green and blue steady state illumination and without it (no background). as expected, results show that the blue background enhances the light-to-dark sensitivity in the long wavelength range and quenches it in the short wavelength range. the red bias has an opposite behavior; it reduces the ratio in the red/green wavelength range and amplifies it in the blue one. the sensor is a wavelength current-controlled device, that make use of changes in the wavelength of the optical bias to control the power delivered to a load, acting as an optical amplifier. its gain, defined as the ratio between the photocurrent with and without a specific background depends on the background wavelength that controls the electrical field profile across the device. if the electrical field increases locally (self optical amplification) the collection is enhanced and the gain is higher than one. if the field is reduced (self optical quench) the collection is reduced and the gain is lower than one. this optical nonlinearity makes the transducer attractive for optical communications. 4.2 self bias amplification under transient conditions and uniform irradiation a chromatic time dependent wavelength combination (4000 bps) of r (λr=624 nm), g (λg=526 nm) and b (λb=470 nm) pulsed input channels with different bit sequences, was used to generate a multiplexed signal in the device. the output photocurrents, under positive (dot arrows) and negative (solid arrows) voltages with (colour lines) and without (dark lines) background are displayed in figure 5. the bit sequences are shown at the top of the figure. results show that, even under transient input signals (the input channels), the background wavelength controls the output signal as in figures 3 and 4. this nonlinearity, due to the transient asymmetrical light penetration of the input channels across the device together with the modification on the electrical field profile due to the optical bias, allows tuning an input channel without demultiplexing the stream. 400 450 500 550 600 650 700 0.0 2.0 4.0 green background p ho to cu rr en t ( na ) wavelength (nm) +1 v 0 v -3 v -5v -8 v 400 450 500 550 600 650 700 0.0 2.0 4.0 6.0 8.0 10.0 12.0 blue background p ho to cu rr en t ( na ) wavelength (nm) +1 v 0 v -3 v -5v -8 v 400 450 500 550 600 650 700 0.0 2.0 4.0 no background p ho to cu rr en t ( na ) wavelength (nm) +1 v 0 v -3 v -5v -8 v 400 450 500 550 600 650 700 0.0 0.5 1.0 1.5 2.0 red background p ho to cu rr en t ( na ) wavelength (nm) +1 v 0 v -3 v -5v -8 v m. vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-12 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 0.0 0.5 1.0 1.5 2.0 2.5 0 1 2 3 4 φl=0 brr v=+1v v=-8v red green blue p ho to cu rr en t ( μa ) time (ms) figure 5. single and combined signals @-8v; without (solid arrows) and with (dotted arrows) green optical bias. in figure 6 the green channel is tuned through the difference between the multiplexed signal with and without green bias. 0.0 1.0 2.0 3.0 4.0 5.0 -1 0 1 2 3 4 5 i on -i off bias off b gr v=+1v v=-8v red green blue p ho to cu rr en t ( μ a ) time (ms) figure 6. multiplexed signals @-8v/+1v (solid /dot lines); without (bias off) and with (r, g, b) green optical bias. 5 wavelength-division (de)multiplexing device in this section we report the use of the same devices to combines the demultiplexing operation with the simultaneous photodetection of the signal. the device makes use of the fact that the optical absorption of the different wavelengths can be tuned by means of electrical bias changes or optical bias variations (figure 5 and figure 6). this capability was obtained using adequate engineering design of the multiple layers thickness, absorption coefficient and dark conductivities [7]. the device described herein operates from 400 to 700 nm which makes it suitable for operation at visible wavelengths in optical communication applications. monochromatic pulsed beams together or one single polychromatic beam (mixture of different wavelength) impinge in the device and are absorbed, according to their wavelength (figure 7). by reading out, under appropriate electrical bias conditions, the photocurrent generated by the incoming photons, the input information is electrically multiplexed or demultiplexed. 0.0 1.0 2.0 3.0 4.0 5.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 r+g+b b[10101010] g[10011001] r[01111000] v=+1v v=-8v r&g&b g b r p ho to cu rr en t ( μ a ) time (ms) figure 7. single (r, g and b) and combined (r&g&b) signals under -8v (solid arrows) and +1v (dotted arrows). in the multiplexing mode the device faces the modulated light incoming together (monochromatic input channels). the combined effect of the input signals is converted to an electrical signal, via the device, keeping the input information (wavelength, intensity and modulation frequency). the output multiplexed signal, obtained from the combination of the three optical sources, depends on both the applied voltage and on the on-off state of each input optical channel. under negative bias, the multiplexed signal presents eight separate levels. the highest level appears when all the channels are on and the lowest if they are off. furthermore, the levels ascribed to the mixture of three or two input channels are higher than the ones due to the presence of only one (r, g, b). optical nonlinearity was detected; the sum of the input channels (r+b+g) is lower than the correspondent multiplexed signals (r&g&b). this optical amplification, mainly on the on-on states, suggests capacitive charging currents due to the timevarying nature of the incident lights. under positive bias the levels were reduced to one half since the blue component of the combined spectra falls into the dark level, the red remains constant and the green component decreases. in the demultiplexing mode a polychromatic modulated light beam is projected onto the device and the readout performed by shifting between forward and reverse bias. figure 7 displays the input and multiplexed signals under negative (-8v) and positive (+1v) electrical bias. as expected, the input red signal remains constant while the blue and the green ones decrease as the voltage changes from negative to positive. to recover the transmitted information (8 bit per wavelength channel) the multiplexed signal, during a complete cycle, was divided into eight time slots, each corresponding to one bit where the independent optical signals can be on (1)or off (0). under positive bias, the device has no sensitivity to the blue channel, so the red m. vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-12 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc and green transmitted information are tuned and identified. the highest level corresponds to both channels on (r=1, g=1), and the lowest to the off-off stage (r=0; g=0). the two levels in-between are related with the presence of only one channel on, the red (r=1, g=0) or the green (r=0, g=1). to distinguish between these two situations and to decode the blue channel, the correspondent sub-levels, under reverse bias, have to be analyzed. the highest increase at -8v corresponds to the blue channel on (b=1), the lowest to the on stage of the red channel (r=1) and the intermediate one to the on stage of the green (g=1). using this simple key algorithm the independent red, green and blue bit sequences were decoded as: r [01111000], g[10011001] and b[10101010 which are in agreement with the signals acquired for the independent channels. 6 the optoelectronic model. 6.1 two connected transistor model based on the experimental results and device configuration an optoelectronic model was developed [8] which is displayed in figure 8. the optoelectronic block diagram (a) consists of four essential elements: a double pin device for detection, a voltage supply to dc voltage bias; optical connections for light triggering and optical bias to dc light bias control. these four elements when connected together form the essentials of the optoelectronic circuit. optical bias double pin photodiode structure voltage supply ligth triggering 1 2 3 4 load optical bias double pin photodiode structure voltage supply ligth triggering 1 2 3 4 load double pin photodiode structure voltage supply ligth triggering 1 2 3 4 load a) p p q1 q2 n p p i n p i n n p p i p i’ n p i1,3 i2,4 p p q1 q2 n p p i n p i n n p p i p i’ n p i1,3 i2,4 p p q1 q2 p q1 q2 n p p i n p i n n p p i p i’ n p i1,3 i2,4 b) i4 q2 i1 i2 q1 c2 r2 r1 (v) c1 i3 c) figure 8. a) optoelectronic block diagram. b) compound connected phototransistor equivalent model. c) ac equivalent circuit. the monolith device is modelled by the twotransistor model (q1-q2) (figure 8b) that can be considered to constitute pnp (q1) and npn (q2) phototransistors separately. the ac circuit representation is displayed in figure 8c. its operation is based upon the following principle: the flow of current through the resistor connecting the two transistor bases is proportional to the difference in the voltages across both capacitors (charge storage buckets). two optical gate connections ascribed to the different light penetration depths across the front (q1) and back (q2) phototransistors were considered to allow independent blue (i1), red (i2) and green (i3, i4) channels transmission. four square-wave current sources with different intensities are used; two of them, i1 and i2, with different frequencies to simulate the input blue and red channels and other two, i3 and i4, with the same frequency but different intensities to simulate the green channel due to its asymmetrical absorption across both front and back phototransistors. the charge stored in the space-charge layers is modelled by the capacitor c1 and c2. r1 and r2 model the dynamical resistances of the internal and back junctions under different dc bias conditions. once the ac sources are connected in the load loop an ac current flows through, establishing voltage modifications across the two capacitors. during the simultaneous transmission of the three independent bit sequences, the set-up in this capacitive circuit loop is constantly changing in magnitude and direction. this means that the voltage across one capacitor builds up until a maximum and the voltage across the other builds up to a minimum. the system collapses and builds up in the opposite direction. it tends to saturate and then leave the saturation because of the cyclic operation. this results in changes on the reactance of both capacitors. the use of separate capacitances on a single resistance r1 results in a charging current gain proportional to the ratio between collector currents. the dc voltage, according to its strength, aids or opposes the ac currents. 6.2 optoelectronic simulation the multiplexed signal was simulated by applying the kirchhoff’s laws for the simplified ac equivalent circuit (figure 8c) and the four order runge-kutta method to solve the corresponding state equations. )( 1 1 )( 111 11 2,1 2 1 2,1 222121 11112,1 ti c ctv crcrcr crcr dt dv ⎥ ⎥ ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎢ ⎢ ⎣ ⎡ + ⎥ ⎥ ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎢ ⎢ ⎣ ⎡ −− − = (1) ( ) ( )tv r ti 2,1 2 1 0 ⎥ ⎦ ⎤ ⎢ ⎣ ⎡ = (2) with v1(t) and v2(t) the emitter-base voltages of q1 and q2 transistors and i(t) the generated photocurrent under transient conditions. matlab was used as a programming environment and the input parameters m. vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-12 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc chosen in compliance with the experimental results (figure 7). in figure 9 the simulated currents (symbols) under negative and positive bias are compared. in figure 10 and under negative bias the simulated current under green (a) and red backgrounds are displayed. the current sources are also displayed (dash lines). to simulate the red and the green backgrounds, current sources intensities were multiplied by the on/off ratio between the input channels with and without optical bias. the same bit sequence of figure 7 was used. to validate the model the experimental multiplexed signals are also shown (solid lines). good agreement between experimental and simulated data was observed. under negative bias and if no optical bias is applied, the expected eight levels are detected, each one corresponding to the presence of three, two, one or no color channel on. under positive bias or steady state irradiation the levels are amplified or reduced depending on the external control. the expected optical amplification is observed due to the effect of the active multiple-feedback filter when the back diode is light triggered. 0.0 1.0 2.0 3.0 4.0 5.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 i(r 1 =1kω,r 2 =5kω) i(r 1 =10mω,r 2 =5kω) i 1 =0.32μα i 2 =1.1μα i 3 =1.1μα i 4 =0.82μα v=+1v v=-8v p ho to cu rr en t ( μ a ) time (ms) figure 9. multiplexed simulated (symbols), current sources (dash lines) and experimental (solid lines) under negative (r1=1kω; -8v) and positive (r1=10mω; +1v) dc bias without any background. under forward bias (high r1) the device remains in its non conducting state unless a light pulse (i2 or i2+i4) is applied to the base of q2. this pulse causes q2 to conduct because the reversed biased n-p internal junction behaves like a capacitor inducing a charging current (i2+i4) across both collector junctions. the collector of the conducting transistor pulls low, moving the q1 base toward its collector voltage, which causes q1 to conduct. the collector of the conducting q1 pulls high, moving the q2 base in the direction of its collector. this positive feedback (regeneration) reinforces the q2 already conducting state and a current i2+i4 will flow on the external circuit. green irradiation (figure 10a) moves asymmetrically the q1 and q2 bases toward their emitter voltages, resulting in lower values of i3 and i4 when compared without optical bias. i1 and i2 slightly increase due to the increased carrier generation on the less absorbing phototransistors. under negative bias during the duration of the red and blue pulses (i1 or i2 on), as without optical bias, the internal junction remains forward biased and the transferred charge between c1 and c2 reaches the output terminal as a capacitive charging current. during the green pulse (i3 and i4 on) only residual charges are transferred between c1 and c2. so, only the charges generated in the base of q2 (i4) reaches the output terminal as can be confirmed by the good fitting between simulate and experimental differences of both multiplexed signals without and with optical bias. under red background the expected optical amplification in the low wavelength range is observed (figures 3a and 4) due to the effect of the active multiple-feedback filter when the back diode is light triggered 0.0 1.0 2.0 3.0 4.0 5.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 i(off) r 1 =1kω r 2 =5kω i(on) i 1 =0.32μα i 2 =1.21μα i 3 =0.54μα i 4 =0.82μα c 1 /c 2 =2.5 φ=0 φ=73mwcm-2 λ=554 nm c p ho to cu rr en t ( μ a ) time (ms) i φoff -i φon a) 0.0 1.0 2.0 3.0 4.0 5.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 i(off) r1=1kω r 2 =5kω i(on) αi 1 =0.93 αi 2 =2.7 αi 3 =0.54 αi 4 =0.5φ=0 φ=73mwcm-2 λ=624 nm p ho to cu rr en t ( μ a ) time (ms) b) figure 10. multiplexed simulated (symbols), current sources (dash lines) and experimental (solid lines): under negative (r1=10mω; -8v) and positive (r1=1kω; +1v) dc bias and green (a) and red (b) backgrounds. 6.3 transfer characteristics to illustrate the transfer characteristics effects due to changes in steady state light, dc control voltage or applied light pulses the following steps in the operation and control of the optoelectronic circuit are considered: 1 positive voltage bias control. only an ac current is flowing through the load capacitance c2. the device has high ohmic resistance (high r1) and remains in its m. vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-12 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc non conducting state unless a light pulse (i2 or i2+i4) is applied to the base of q2. this pulse causes q2 to conduct as the reversed biased n-p internal junction behaves like a capacitor inducing a charging current across r2. no amplification effect was detected since q1 acts as a load and no charges are transferred between c1 and c2 (figure 6a). 2 negative dc voltage bias control. under negative bias the device has low ohmic resistance (low r1) the base emitter junction of both transistors are inversely polarized and conceived as phototransistors, thus taking advantage of the amplifier action of neighboring collector junctions, which are polarized directly. this results in a charging current gain proportional to the ratio between both collector currents (c1/c2). the device behaves like an optoelectronic controlled transmission system that stores, amplifies and transports the minority carriers generated by the current pulses, through the capacitors c1 and c2. here, the dc voltage control creates a voltage across one or both capacitors which, when superimposed with an ac pulse, collectively saturates the circuit. no additional change in the voltage across the capacitors occurs. this results in maximum power transfer to the load. 3 optical bias control. depending on it wavelength (figures 3 and figure 4) the optical bias changes the amplitude of the ac current sources by an α factor, and so the voltages across one or both capacitors. blue, red or green irradiations move asymmetrically the bases of q1, q2 or both toward their emitter voltages (self-forward effect), resulting, respectively in lower values of i1, i2, i3 and i4 when compared with no optical bias. the circuit can leave the saturation resulting in a wavelength controlled power transfer to the load that allows tuning input channel from the stream (figure 10). results also show that the two-transistor model explains the difference between the conduction mechanisms, under positive and negative bias, helping to understand the signal decoding algorithm. under positive bias the red and green channels are immediately decoded since the collected current is due exclusively to the generated carriers at the back diode. under negative bias, depending on the ratio between c1 and c2, different charging currents across the reversed junctions have to be considered. the balance between them depends on the presence of three, two or one channel on (fig. 2-3). so, by comparing the different sublevels signals under positive and negative bias it is possible to recover the blue, the green and the red input channels. 7 conclusions a double pi’n/pin a-sic:h heterostructure with optical gate connections was presented. multiple monochromatic pulsed communication channels were transmitted together and the multiplexed signal analyzed at different electrical and optical bias. results show that the output signal has a strong nonlinear dependence on the light absorption profile. this effect is due to the self biasing of the junctions under unbalanced light generation profiles. self optical bias amplification under uniform irradiation and transient conditions is achieved. an optoelectronic model based on four essential elements: a voltage supply; a double pin photodiode; optical connections for light triggering and optical power sources for light bias explains the operation of the optical system. references [1] m. j. connelly, semiconductor optical amplifiers. boston, ma: springer-verlag, 2002. isbn 978-0-7923-7657-6. [2] p. louro, m. vieira, yu. vygranenko, a. fantoni, m. fernandes, g. lavareda, n. carvalho mat. res. soc. symp. proc., 989, a12.04 (2007). [3] michael bas, fiber optics handbook, fiber, devices and systems for optical communication, chap, 13, mc graw-hill, inc. 2002. [4] m. vieira, a. fantoni, m. fernandes, p. louro, g. lavareda, c. n. carvalho, journal of nanoscience and nanotechnology, vol. 9 , number 7, july 2009 , pp. 40224027(6). [5] m. vieira, m. fernandes, p. louro, a. fantoni, m. a. vieira, j. costa, m. barata, mat. res. soc. symp. proc.volume 1245, a08-02 (2010) . [6] m. vieira, a. fantoni, p. louro, m. fernandes, r. schwarz, g. lavareda, and c. n. carvalho, vacuum, vol. 82, issue 12, 8 august 2008, pp: 1512-1516. [7] p. louro, m. vieira, yu. vygranenko, a. fantoni, m. fernandes, g. lavareda, n. carvalho mat. res. soc. symp. proc., 989 (2007) a12.04. [8] m. a. vieira, m. vieira, m. fernandes, a. fantoni, p. louro, m. barata, amorphous and polycrystalline thin-film silicon science and technology 2009, mrs proceedings vol. 1153, a08-0. m. vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-12 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc demux operation in tandem amorphous si-c devices a two stage active circuit demux operation in tandem amorphous si-c devices a two stage active circuit m. a. vieira1,2 , m. vieira1,2,3, p. louro1,2 , a. fantoni1,2 and a. garção2,3 1 electronics telecommunications and computer dept, isel, lisbon, portugal 2 cts-uninova, lisbon, portugal 3 dee, fct-unl, lisbon, portugal keywords: optical filters, electrical simulation, wdm devices, transceivers, abstract: characteristics of tunable wavelength filters based on a-sic:h multilayered stacked cells are studied both theoretically and experimentally. curves of gain vs. wavelength illustrate the optical filter characteristics. a capacitive active band-pass filter model supports experimental data. results show that the use of a double pi’n/pin a-sic:h heterostructure as active capacitive filters depends on the wavelength and modulated frequency of the trigger light and on the wavelength of the additional optical bias. the devices combine three indispensable functions of transceivers, amplification, detection and wavelength filtering. experimental and simulated results show the device combines the properties of active short-pass and longpass filter sections into a capacitive active band-pass filter. 1 introduction the deployment of fiber to the home (ftth) access networks and the emergence of optical network installations leads to a need for mass production of bi-directional optical modules that can exchange data upand downstream through a single fiber. in order to reduce the cost per user connected to the network, production cost of the optical modules at the optical network unit must be exceptionally low [1, 2]. amorphous silicon carbon tandem structures, through an adequate engineering design of the multiple layers’ thickness, absorption coefficient and dark conductivities can accomplish this function. here, we propose photodiodes with integrated optical thin film filters reducing module cost by minimizing the number of discrete filters. since filters are defined by their wavelengthdomain effects on signals, it makes sense that the most useful analytical and graphical descriptions of filters also fall into the wavelength domain. knowing the transfer function magnitude (or gain) at each wavelength allows us to determine how well the optical filter can distinguish between signals at different wavelengths. typical wavelengths used in infrared optical network transceivers are long wavelengths for downstream data and short wavelengths for upstream communication. in this study we demonstrate the subsequently opportunity integration of two useful optical active filters, a short-and a long-pass filter, eliminating crosstalk between downstream and upstream channels transceivers, by transmitting only one of the two datastreams and blocking the other: red channel and/or green channel for downstream data and blue channel for upstream data. in this paper we demonstrate the integration of short-pass and long-pass optical thin film filters targeting applications combined with a versatile amorphous si/sic p-i’-n-p-i-n device concept. 2 experimental details 2.1 device optimization and operation i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-13 http://journals.isel.pt/index.php/iajetc voltage and optical bias controlled devices, with front and back indium tin oxide transparent contacts, were produced by pecvd at 13.56 mhz radio frequency and tested for a proper fine tuning of the visible spectrum. figure1 device configuration the active device (fig. 1), consists of a p-i'(asic:h)-n/p-i(a-si:h)-n heterostructure with low conductivity doped layers (<10-7ω-1cm-1). the thicknesses and optical gap of the thin í'(200nm; 2.1 ev) and thick i(1000nm; 1.8ev) layers are optimized for light absorption in the blue and red ranges, respectively [3]. transparent contacts have been deposited on front and back layers to allow the entrance and exit off of the light from both sides. photo generation occurs firstly in the a-sic:h absorber, and the remaining non-absorbed light goes through the a-si:h layer. as result, both front and back diodes act as optical filters confining, respectively, the blue and the red optical carriers, while the green ones are absorbed across both [4]. the device operates within the visible range using as input color channels (data) the modulated light (external regulation of frequency and intensity) supplied by a red (r: 626 nm; 51μw/cm2), a green (g: 524 nm; 73μw/cm2) and a blue (b: 470nm; 115μw/cm2) led. additionally, steady state red, green and blue illumination (background) was superimposed using similar leds. light was always impinging from the glass side. 2.2 transfer function characteristics spectral response measurements and transfer function magnitude (gain) characteristics at different frequencies (250hz-3500hz) were analyzed. the spectral sensitivity was tested through data transmission measurements under different frequencies, with and without applied steady state bias. in fig. 2, the spectral photocurrent of the stacked device (fig.1), at 250hz, is displayed under red, green and blue background irradiations (color symbols) and without it (black symbols). for comparison the normalized spectral photocurrent for the front, p-i’-n, and the back, p-i-n, photodiodes (lines) are superimposed. data confirm that the front and back photodiodes act, separately, as optical filters. each diode, separately, presents the typical response of single p-i-n cells with intrinsic layers based on a-sic:h or a-si:h materials, respectively. the front diode, based on a-sic:h heterostructure, cuts the wavelengths higher than 550nm while the back one, based on a-si:h, cuts the ones lower than 500nm. results show that the background wavelength has a strong influence on the spectral sensitivity of the all device. 400 450 500 550 600 650 700 0 20 40 60 80 100 120 n or m al iz ed p ho to cu rr en t no background red background green background blue background v=-8v 250 hz p ho to cu rr en t ( na ) wavelength (nm) 0 1 p-i'-n diode p-i-n diode figure 2 spectral photocurrent without and under different background wavelengths (symbols).the normalized spectral current for the front, p-i’-n, and the back p-i-n photodiodes separately (lines) is superimposed. under red irradiation, when compared with its value without external background, the photocurrent is strongly enhanced at short wavelengths and disappears for wavelengths higher than 550 nm, acting as a short-pass filter. under blue irradiation the devices behaves as a long-pass filter for wavelengths higher than 550 nm, blocking the shorter wavelengths. in fig. 3 the spectral gain, defined as the ratio between the spectral photocurrents, under red (αr), green (αg) and blue (αb) steady state illumination and without it (no background) are plotted at 250hz (lines). under green background it is also displayed the spectral photocurrent at 3500hz (symbols). results from both fig. 2 and fig. 3 confirm the wavelength controlled spectral sensitivity of the device, under steady state illumination. under green irradiation the spectral response depends on the frequency. at 250hz the spectral sensitivity is strongly reduced while at 3500hz the device behaves as a band-stop active filter that screens out the medium wavelength range (green) enhancing m.a.vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-13 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc only the photocurrent for wavelengths outside of that range. 400 450 500 550 600 650 700 0 1 2 3 4 5 6 7 624nm, 250hz 526nm, 250hz 470nm, 250hz 526nm,3500hz g ai n (α r ,α g ,α b ) wavelength (nm) figure 3 spectral gain under red (αr), green (αg), and blue (αb), optical bias for different frequencies. when an external bias is applied to a double pin structure, its main influence is in the field distribution within the less photo excited sub-cell [4]: the front cell under red irradiation; the back cell under blue light, and both, under green steady state illumination (fig. 1). in comparison with thermodynamic equilibrium conditions (no background), the electrical field under illumination is lowered in the most absorbing cell (self forward bias effect) while the less absorbing reacts by assuming a reverse bias configuration (self reverse bias effect). so, the sensor is a bias wavelength current-controlled device that makes use of changes in the wavelength of the background to control the power delivered to the load, acting as an optical amplifier. its gain (fig. 3) depends on the background illumination wavelength. if the electrical field increases locally (self optical amplification) the collection is enhanced and the gain is higher than one. if the field is reduced (self optical quench) the collection is reduced and the gain becomes lower than one. this optical nonlinearity makes the transducer attractive for optical communications and can be used to distinguish a wavelength, to suppress a color channel or to multiplex or demultiplex an information-modulated wave. 2.3 frequency analysis the study of the frequency influence on the device performance was analyzed through the spectral response of the device without and with steady state optical light. results are displayed in fig. 4. 400 450 500 550 600 650 700 0,00 0,01 0,01 0,01 0,02 0,03 0,03 0,04 0,04 no optical bias 250 hz 3500 hz p ho to cu rr en t ( μ a ) wavelength (nm) a) 400 450 500 550 600 650 700 0,00 0,01 0,01 0,01 0,02 0,03 0,03 red background 250 hz 3500 hz p ho to cu rr en t ( μ a ) wavelength (nm) b) 400 450 500 550 600 650 700 0,00 0,02 0,04 0,06 0,08 0,10 0,12 blue background 250 hz 3500 hz p ho to cu rr en t ( μ a ) wavelength (nm) c) 400 450 500 550 600 650 700 0,00 0,01 0,02 green background p ho to cu rr en t ( μ a ) wavelength (nm) 3500hz 3000hz 2500hz 2000hz 1500hz 1000hz 500hz 250hz d) figure 4 photocurrent variation with the wavelength for different frequencies at -8 v obtained: a) without, b) with blue, c) with red and d) with green background. m.a.vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-13 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc data from fig. 4a show that without background light the curves measured under different frequencies exhibit the same trend with two peaks located at 500 nm and 600 nm. the signal is reduced with the increase of the frequency. under blue steady state illumination (fig. 4c) the spectral response exhibits a different trend with a single peak located at 600 nm. this is due to the strong attenuation of the short wavelengths. the red steady state illumination (fig. 4c) has the opposite effect with a single peak at 500 nm. under green background the spectral response shows two different regimes depending on the operation frequency. in the low frequency range the signal is similar to the trends obtained under red steady state light, while at higher frequencies it follows the behaviour obtained without background light. fig. 5 shows the spectral gain as a function of the frequency under red (α r), green (α g) and blue (α b) backgrounds at 624 nm (αr, red channel), at 526 nm (αg, green channel) and at 470 nm (αb, blue channel). results show how the device reacts to the background wavelength and to the modulated input color channels (pulses) between 250 hz and 3500 hz. under red and green irradiations two frequency regimes can be considered. one, for frequencies lower than 2000 hz, were either under red and green backgrounds the green and the red channel gains are very low (<<1) and the blue gain is strongly enhanced (>>1) under red background or reduced (<1) under green irradiations. the other regime, for frequencies higher than 2000 hz, the gain increases with the frequency, gradually under red and quickly under green steady state illumination. under blue, the gain increases slowly with the frequency being higher than one for the red and green channels and lower for the blue one. consequently, under red irradiation (fig. 5a) the transfer function has extra gain at short wavelengths (blue channel), than at longer wavelengths acting as a short-pass filter whatever the frequency. under green background (fig. 5b), the manipulation of amplitude is achieved by changing the frequency of the modulated lights. at high frequencies the device is a band-stop active filter that works to screen out wavelengths that are within a certain range (green channel), giving easy passage only all wavelengths below (blue channel) and above (red channel). in the low frequency regime all the amplitudes are quenched. under blue steady state optical bias (fig. 5c) the device behaves as a long-pass active filter that transmits and enhances the long wavelength photons (red and green channels) while blocking the shorter wavelengths (blue channel). 0 750 1500 2250 3000 0 1 2 3 4 5 6 blue channel green channel red channel v=-8v red background g ai n (α r r ,α r g , α r b ) frequency (hz) a) 0 750 1500 2250 3000 0,0 0,5 1,0 1,5 2,0 2,5 3,0 blue channel green channel red channel v=-8v green background g ai n (α g r ,α g g , α g b ) frequency (hz) b) 0 750 1500 2250 3000 0 1 2 3 4 5 6 linear fit blue chanel green chanel red chanel v=-8v blue background g ai n (α b r ,α b g , α b b ) frequency (hz) c) figure 5 spectral gain as a function of the frequency at 624 nm (red channel), at 526 nm (green channel) and at 470 nm (blue channel) under red (αr), green (αg) and blue (αb) backgrounds. a) short-pass filter , b) band-stop filter, c) long-pass filter. 3 numerical simulation in order to understand the light filtering properties of the device, under different electrical and optical bias conditions, a simulation program asca-2d [6] m.a.vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-13 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc was used having as input parameters the experimental data. for the a-sic:h and a-si:h absorbers an optical band gap of 2.1 ev and 1.8 ev and a thickness of 200 nm and 1000 nm were chosen, respectively. the doping level was adjusted in order to obtain approximately the same conductivity of the typical thin film layers. 0,0 0,2 0,4 0,6 0,8 1,0 1,2 1,4 1015 1016 1017 1018 1019 1020 1021 1022 λ r =650 nm (-6<v<+1 ) λ b =450 nm (0>v>-6) pi(200nm)n/pi(1000nm)n r ec om bi na tio n /c m -3 s1 ) position (μm) a) 0,0 0,2 0,4 0,6 0,8 1,0 1,2 1,4 1015 1016 1017 1018 1019 1020 1021 1022 pi(200μm)n/pi(1000μm)n -6v λ g =550 nm (-6<v<+1 v) r ec om bi na tio n (c m -3 s1 ) position (μm) b) figure 6 recombination profiles (straight lines) under: a) red (λr =650 nm) and blue (λb =450 nm) b) green (λg =550 nm) optical bias and different applied voltages. the generation profiles are shown (symbols). in fig. 6a the recombination profiles (straight lines) under red (λr =650 nm) and blue (λb =450 nm) irradiation are displayed at different electrical bias. in fig 6b it is shown the same profiles but under green bias (λg=550 nm). the generation profiles are also displayed (symbols). simulated results show that the thickness and the absorption coefficient of the front photodiode are optimized for blue collection and red transmittance, and the thickness of the back one adjusted to achieve full absorption in the greenish region and high collection in the red spectral range. as a result, both front and back diodes act as optical filters confining, respectively, the blue and the red optical carriers, while the green ones are absorbed across both. under negative applied voltage, in fig. 7a it is reported the electric field profile under different wavelengths backgrounds of the optical bias. in fig. 7b, under red background, the three red, green and blue channels were added and the electrical field profile displayed. 0,0 0,2 0,4 0,6 0,8 1,0 1,2 1,4 -4,0x107 -3,5x107 -3,0x107 -2,5x107 -2,0x107 -1,5x107 -1,0x107 -5,0x106 0,0 5,0x106 no bias λ g =550 nm λ r =650 nm -6 v λ b =450 nm e le ct ric al fi el d (v /m ) position (μm) a) 0,0 0,2 0,4 0,6 0,8 1,0 -7x107 -6x107 -5x107 -4x107 -3x107 -2x107 -1x107 0 1x107 λ r =650 nm red background no channels λ chg =550 nm λ chr =650 nm -6 v λ chb =450 nm e le ct ric al fi el d (v /m ) position (μm) b) figure 7 electric field profiles within the p-i-n/p-i-n tandem structure. a) under different wavelengths backgrounds (λr,g,b) and without it. b) red background (λr ) and different color channels (λchr, λchg, λchb) results show that the balance between the electrical field adjustments due to the non uniform absorption throughout the structure depends on the generation/recombination ratio profiles at each background wavelength. the shallow penetration of the blue photons into the front diode, the deep penetration of the red photons exclusively into the back diode or the decay of the green absorption across both controls the sensor behavior. the external background interferes mainly with the less absorbing cell (the front under red, the back under blue and with both under green irradiations). both the front and the back diodes are optically and electrically in series. under steady state irradiation, to sustain the current across the device, the current at the less absorbing diode has to be adjusted through an increase of the electrical field and thus it becomes reverse biased (fig. 7a). the superposition of a color m.a.vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-13 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc channel will affect locally this field. under red background (fig. 7b), the blue channel increases the field intensity in the front diode and even reverse it at the internal interface increasing carrier collection. the red and the green channels change the field in an opposite way. the sensor is a bias wavelength current-controlled device that makes use of changes in the wavelength of the background to control the power delivered to the load, acting as an amplifier of the optical signals. its gain (figs. 3 and 5) depends on the background illumination wavelength. if the electrical field increases locally (self optical amplification) the collection is enhanced and the gain is higher than one. if the field is reduced (self optical quench) the collection is reduced and the gain becomes lower than one [7]. this optical nonlinearity makes the transducer attractive for optical communications and can be used to distinguish a wavelength, to suppress a color channel or to multiplex or demultiplex an information-modulated wave. 4 capacitive active band pass filter model 4.1 two section circuit based on the experimental results and device configuration an optoelectronic model supported by the complete dynamical large signal ebers-moll model was developed [8, 9] and is displayed in fig. 8. the electrical model of the device looks fairly complex, but when broken down can be divided into two sections (fig. 8a): a short-pass filter (front phototransitor, q1) and a long-pass filter (back phototransistor, q2) sections. so, it can be made out of a short-pass and a long-pass filter when connected both active filter sections in parallel (fig. 8b).the charge stored in the space-charge layers is modeled by the capacitor c1 and c2. r1 and r2 model the dynamical resistances of the internal and back junctions under different dc bias conditions. the operation is based upon the following principle: the flow of current through the resistor connecting the two transistor bases is proportional to the difference in the voltages across both capacitors (charge storage buckets). to allow independent blue, red and green channels transmission four square-wave current sources with different intensities are used; two of them, α1i1 and α2i2, with different frequencies to simulate the input blue and red channels and the other two, α1i3 and α2i4, with the same frequency but different intensities, to simulate the green channel due to its asymmetrical absorption across both front and back phototransistors. i´p np nn ip q1 q2 ib,ig ir,ig a) b) r1 ( v) α2 igpinα2 ir r2 q2 c2 α1 ib q1 c1 α1 igpi ´n c) figure 8 a) two connected transistor model, b) two active capacitive filter sections and b) ac equivalent circuit. once the ac sources are connected in the load loop an ac current flows through, establishing voltage modifications across the two capacitors. during the simultaneous transmission of the three independent bit sequences, the set-up in this capacitive circuit loop is constantly changing in magnitude and direction. this means that the voltage across one capacitor builds up until its maximum and the voltage across the other builds up to a minimum. the system collapses and builds up in the opposite direction. it tends to saturate and then leave the saturation because of the cyclic operation. this results in changes on the reactance of both capacitors. the dc voltage, according to its strength, aids or opposes the ac currents. so, when the pi’npin device is reverse-biased, the base-emitter junction of both transistors are inversely polarized and m.a.vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-13 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc conceived as phototransistors, taking, so, advantage of the amplifier action of adjacent collector junctions which are polarized directly. this results in a current gain proportional to the ratio between both collector currents. under positive bias the internal junction becomes reverse-biased and no amplification effect is observed. 4.2 optoelectronic state model taking into account fig. 8b the time periodic linearized state equations are given by: )()( 111 11 2,1 2 2 1 1 2,1 222121 11112,1 ti c ctv crcrcr crcr dt dv ⎥ ⎥ ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎢ ⎢ ⎣ ⎡ + ⎥ ⎥ ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎢ ⎢ ⎣ ⎡ −− − = α α ( ) ( )tv r ti 2,1 2 1 0 ⎥ ⎦ ⎤ ⎢ ⎣ ⎡ = where α1 and α2 coefficients determine how the background affects the state change. based on figs 2-4., under red background, α1 >1 (αb+ αg >1) and α2<1(αr+ αg <1). the opposite will occur under blue irradiation. under green background both are balanced. in fig. 9 it is displayed the block diagram of the optoelectronic state model for a wdm pi´npin device under different electrical and optical bias conditions. λ1, λ2, λ3 represent the color channels, r1 and r2 the dynamic internal and back resistances and α1 and α2 are the coefficients for the steady state irradiation. to solve the state equations the four order runge-kutta method was applied and matlab used as a programming environment. the input parameters were chosen in compliance with the experimental results. λ1 λ3 λ2 α1/c1 α2/c2 ∑ ∑ −1/r1c1 1/r2 ∫ dt 1/r1c2 1/r1c1 ∫ dt −1/r1c2-1/r2c2 ∑ i(t) + i2(t) i1(t) ++ + + + v1 . . v2 + v1 v2 λ1 λ3 λ2 α1/c1 α2/c2 ∑∑ ∑∑ −1/r1c1 1/r2 ∫ dt∫ dt 1/r1c2 1/r1c1 ∫ dt∫ dt −1/r1c2-1/r2c2 ∑∑ i(t) + i2(t) i1(t) ++ + + + v1 . . v2 + v1 v2 figure 9 block diagram of the optoelectronic state model for a pi´n/pin device. the amplifying elements, α1 and α2, can provide gain if needed and attenuate unwanted wavelengths (<1) while amplifying (>1) desired ones. the values and the strategic placement of the resistors determine the basic shape of the output signals. under negative bias the device has low ohmic resistance (low r1) since the base emitter junction of both transistors are reverse polarized and conceived as phototransistors. this results in a charging current gain proportional to the ratio between both collector currents (α2 c1/ α1 c2). taking into account figs. 34, under red background, α1 >1 and α2<1. the opposite will occur under blue irradiation. under green background both are balanced. under positive bias the internal junction becomes reverse-biased and no amplification effect is observed. 4.3 model validation based on the optoelectronic model and experimental results, a multiplexed signal was simulated by applying the kirchhoff’s laws for the ac equivalent circuit (fig. 8c) and the four order runge-kutta method to solve the corresponding state equations. matlab was used as a programming environment and the input parameters chosen in compliance with the experimental results [7]. 0,0 1,0 2,0 0,0 5,0 10,0 15,0 20,0 25,0 i g i b ir i(off)r 1 =1kω r 2 =5kω v=-8v i(on) φ=0 φ=73mwcm-2 λ=624 nm p ho to cu rr en t ( μ a ) time (ms) a) 0,0 1,0 2,0 3,0 4,0 5,0 0,0 0,5 1,0 1,5 2,0 2,5 3,0 3,5 4,0 4,5 i g i b i r i on i off -i on i offr1=1kω r 2 =5kω φ=0 φ=73mwcm-2 λ=554 nm p ho to cu rr en t ( μ a ) time (ms) i φ,off -i φ,on b) figure 10 multiplexed simulated (symbols), current sources (dash lines) and experimental (solid lines) transient responses: under negative dc bias and red (a) and green (b) backgrounds. m.a.vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-13 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc to simulate the short-pass and ban-stop filters, a multiplexed signal under red and green backgrounds were compared with the correspondent signals without background (symbols). to validate the model the experimental multiplexed signals are shown (solid lines) under the same conditions. the current sources, ir, ig, ib, are also displayed. for their intensity the amplitude of the input experimental channels, without background, were used. to simulate the red and the green backgrounds, current sources intensities were multiplied by the spectral gain ( gr bgr , ,,α ) at the correspondent frequency (fig. 5). a good agreement between experimental and simulated data was achieved (fig. 10). as expected, under red background it is observed the enhancement in the short wavelength (red channel) while quenching the longer wavelengths. under green background the device screens out the medium wavelengths (green channel) and gives a slightly increase on the red and blue wavelengths. depending on the background wavelength, the device behaves like an optoelectronic controlled transmission system that transmits and/or process intelligence (data) in a manner that permits the subsequent recovery of that information. it filters (amplifies/blocks/screens) the carriers generated by the light pulses (current sources), through the capacitors c1 and c2. the manipulation of amplitude is achieved by changing background wavelength at a given modulated frequency. this allows tuning an input channel or to optically demultiplex a polychromatic channel. 5 conclusions a light-activated pi’n/pin a-sic:h device that combines the demultiplexing operation with the simultaneous photodetection and self amplification of an optical signal is analyzed. a numerical simulation supports the self bias amplification effect. results demonstrate that the multiplexed output waveform presents a nonlinear amplitude-dependent response to the wavelengths of the input channels and of the optical bias, acting as an active filter. depending on the wavelength of the external background it acts either as a shortor a longpass band filter or as a band-stop filter. a capacitive active band-pass filter model is presented and gives insight into the physics of the device. an algorithm to decode the multiplex signal was established. references [1] y.iguchi et al., “novel rear-illuminated 1.55μm – photodiode with high wavelength selectivity designed for bidirectional optical transceiver”, proc. 2th int. conf. on inp and related mater, 317 (2000). [2] c. petit, m. blaser, workshop on optical components for broadband communication , edited by pierre-yves fonjallaz, thomas p. pearsall, proc. of spie vol. 6350, 63500i, (2006). [3] m. vieira, a. fantoni, m. fernandes, p. louro, g. lavareda, c. n. carvalho, journal of nanoscience and nanotechnology, vol. 9, , number 7, july 2009 , pp. 4022-4027(6). [4] p. louro, m. vieira, yu. vygranenko, a. fantoni, m. fernandes, g. lavareda, n. carvalho, mat. res. soc. symp. proc., 989 (2007) a12.04. [5] m. vieira, a. fantoni, p. louro, m. fernandes, r. schwarz, g. lavareda, and c. n. carvalho, vacuum, vol. 82, issue 12, 8 august 2008, pp: 1512-1516. [6] a. fantoni, m. vieira, r. martins, "simulation of hydrogenated amorphous and microcrystalline silicon optoelectronic devices" mathematics and computers in simulation, vol. 49. pp. 381-401 (1999). [7] m. vieira, a. fantoni, p. louro, m. fernandes, r. schwarz, g. lavareda, and c. n. carvalho, “selfbiasing effect in colour sensitive photodiodes based on double p-i-n a-sic:h heterojunctions “vacuum, vol. 82, issue 12, 8 august 2008, pp: 1512-1516. [8] m. a. vieira, m. vieira, m. fernandes, a. fantoni, p. louro, m. barata, “voltage controlled amorphous si/sic phototransistors and photodiodes as wavelength selective devices: theoretical and electrical approaches” amorphous and polycrystalline thin-film silicon science and technology — 2009, mrs proceedings volume 1153, a08-03. [9] m. a. vieira, m. vieira, j. costa, p. louro, m. fernandes, a. fantoni, ”double pin photodiodes with two optical gate connections for light triggering: a capacitive two-phototransistor model” in sensors & transducers journal vol. 9, special issue, december 2010, pp.96-120. m.a.vieira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-13 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc performance analysis of wdm-pon architecture for performance analysis of wdm-pon architecture for wireless services distribution in future aircraft networks d. coelho1,2, j.m.b oliveira1, l.m.pessoa1,2, h.m. salgado1 and j.c.s. castro1 1inesc tec(formerly inesc porto), porto, portugal 2faculdade de engenharia, universidade do porto, porto, portugal (dcoelho,joao.b.oliveira,lpessoa,hsalgado,jcastro)@inescporto.pt keywords: wdm, wi-fi, signal-to-noise ratio, error vector magnitude and intermodulation distortion. abstract: in this work, an in-depth analysis concerning the transmission performance of ieee802.11g/n (wi-fi) signals in a wdm-pon system is presented. it is considered that the optical/electrical transceivers are based on low-cost 850 nm vcsels and pin photodiodes. system modelling includes the impact of noise generated in the optical path, such as relative intensity noise (rin), shot noise, photodetector thermal noise, clipping and intermodulation distortion. an analytic analysis based on volterra series is conducted and mathematical expressions for both the evm and snr are derived. the theoretical analysis is also compared with experimental results. among several conclusions, it is observed that the laser intermodulation distortion, clipping and rin are the most relevant factors. 1 introduction in the last decade, wireless communications have experienced a great expansion and the demand for higher data traffic to accommodate the new services, like voip (voice over ip), iptv (internet protocol television) or video on demand (vod) and peer-topeer (p2p) have increased quickly. one application domain in which wireless networks can obtain additional use is in aviation, since there is the need for connectivity and network services “every time” and “everywhere”. another important factor is that commercial aircraft operators are currently looking for ways to attract more customers by increasing the value of their service offerings to passengers [1]. in the future, in-flight entertainment (ife) services should offer to passengers high speed wireless internet connectivity, using their own personal computers as if they were earthbound [1]. entertainment services can comprise digital video and audio services, such as high definition (hd) video on demand, music, and satellite hdtv. figure 1, depicts a scheme of a possible future optical fiber based network suitable for aircraft cabins, where the communication signals between the aircraft, ground and satellite stations are represented. head end access point ground station satellite figure 1: scheme of the aircraft communication system. pons are quickly becoming one of the most popular access systems in telecommunications. wireless and fixed access convergence over pons, are also topics of much debate [2]. they can be used to provide wireless communication services in these networks due to its high bandwidth and without electromagnetic interference. moreover, using wdm there is a simplification, compared to tdm pons, of the network topology by allocating different wavelengths to individual optical network terminator (ont). i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-16 http://journals.isel.pt/index.php/iajetc pons can transparently delivery multiple services such as ieee 802.11(wi-fi), global system for mobile communications (gsm), wimax or ultra-wide band (uwb). ieee 802.11 or wireless-fidelity (wi-fi) is a set of standards for implementing wireless local area network (wlan) data communication in the 2.4 and 5 ghz frequency range. ieee 802.11g works in the 2.4 ghz frequency band and uses ofdm signals with 20mhz of bandwidth and power signal of 0dbm/mhz. the subcarriers can be qpsk, 16-qam or 64-qam. the ieee 802.11n standard adds the use of the multiple-input multiple-output (mimo) technology and operates in the 2.4 ghz and 5.0 ghz bands with 40-mhz bandwidth. orthogonal frequency division multiplexing (ofdm) is a multicarrier modulation scheme that is well known due to its robustness in multipath fading channels. yet, due to its high peak-to-average power ratio, ofdm is susceptible to nonlinear distortion from components such as optical modulators and directly modulated diode lasers [5]. a low cost solution for these systems involves the use of directly modulated vertical cavity surface emitted lasers (vcsels). these lasers are characterized by a vertical low divergence, circular beam patterns, low threshold currents (a few ma) and high bandwidths (several ghz). their vertical wafer growth process enables in-wafer testing, and is well suited for large scale production. these are reasons that make vcsels desirable for low cost directly modulated systems in these types of widespread commercial applications [3]. this article is divided in seven sections. section 2 describes briefly the wdm-pon topology. section 3 addresses the interference generated by source nonlinearities. section 4 presents the laser model. section 5 presents discusses the system performance analysis including all noise contributions and laser distortions at the receiver. the results are presented in section 6 with the comparison between simulation and experimental results. finally, section 7 highlights the main conclusions. 2 wdm-pon wavelength-division-multiplexed passive optical networks (wdm-pon) offer many advantages such as large capacity, easy management, network security, and upgradability. the usage of an array waveguide grating (awg) to multiplex/de-multiplex the upstream and downstream wavelengths, respectively, provides a dedicated point-to-point optical channel between each ont and the optical line terminator (olt), although this concept involves the sharing of a common point-to-multipoint physical architecture [4]. since a wavelength mux/demux is used instead of an optical-power splitter, the insertion loss is considerably smaller and effectively independent of the splitting ratio. in addition, since the receiver bandwidth for each ont is matched to its dedicated bandwidth, there is no additional penalty related to the number of users on the pon [4]. consequently, the signal-to-noise ratio (snr) is essentially independent of the number of onts, allowing efficient scaling and flexibility for a wdm-pon architecture, which is suited to transport multiple wireless standards including wi-fi. a wdm-pon scheme representation is depicted in figure 2. in this architecture, each ont-olt pair is assigned a set of downstream and upstream wavelengths. figure 2: wdm-pon scheme. an additional feature of a wdm architecture is its ability to localize any fault or optical loss in the fiber plant by using a single wavelength-tunable otdr (optical time-domain reflectometer) located at the olt. awgs require temperature control to keep their optical channels locked to a wavelength grid. technology advances have allowed the recent commercialization of athermal awgs that can remain locked to a wdm-wavelength grid over temperature ranges experienced at the passive-node location [4]. 3 intermodulation distortion due to the large number of electrical subcarriers of the wifi signal, a high nonlinear distortion may be expected from the electrical to optical conversion when using direct modulated laser diodes, such as vcsels. the interference resulting from source nonlinearity depends strongly on the number of channels and the distribution of channel frequencies. considering d. coelho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-16 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc the transmission of three channels ( f1, f2 and f3), figure 3 shows the harmonics generated by a nonlinear device. amplitude frequencyf 2 f 1 f 3 f 1 2f 1 f 3 f 1 + f 2 f 3 f 2 + f 3 f 1 2f 3 f 1 f 1 f 3 f 2 2f 1 3f 2 2f 3 f 1 + f 3 f 1 + f 2 f 2 + f 3 f 1 + f 2 + f 3 2f 1 + f 2 2f 1 + f 3 2f 3 + f 1 2f 3 + f 2 3f 1 transmission band figure 3: illustration of harmonics generated by a nonlinear optical modulator [5]. the most troublesome third-order intermodulation distortion products (imps) are those that originate from frequencies fi + f j − fk and 2 fi − f j , and since they lie in within the transmission band, they lead to interchannel interference. the interference, thus, depends strongly on the number of channels and on the allocation of channel frequencies with respect to the resonance frequency of the laser. for a n channels system with uniform frequency spacing the number of imps, r imn21 and r im n 111 of type 2 fi − f j and fi + f j − fk, respectively, coincident with channel r are given by [6]. r im n 21 = 1 2 { n − 2 − 1 2 [1 −(−1)n](−1)r } (1) r imn111 = r 2 (n − r + 1)+ 1 4 [ (n − 3)2 − 5 ] − − 1 8 [1 −(−1)n](−1)n+r (2) considering a wifi system, the total number of channels is equals n = 64. figure 4 shows the total number of third-order imps as a function of channel number. the channel in the middle of the band is the one with the large number of intermodulation products. 4 laser model 4.1 intrinsic model the laser operation is described by the relationship between the carrier density n and the photon density 0 10 20 30 40 50 60 1000 1100 1200 1300 1400 1500 channel number n u m b e r o f i m p s figure 4: total number of third-order intermodulation products for n=64. s under the presence of the injected current i. this is accomplished through a set of rate equations that explain all the mechanisms by which the carriers are generate or lost inside the active region. this set of equations is defined by [7]: dn dt = ηii qv − n τn − g0 (n − n0m)(1 − εs)s (3) ds dt = γg0 (n − n0m)(1 − εs)s − 1 τp s + βγ n τn (4) the first term in (3) is the rate at which the carriers, electrons or holes are injected into the active layer due to current i. the second term in the equation is the loss due to various recombination process (spontaneous and nonraditive emission) and the last term is due to the stimulated emission recombination that leads to the emission of light. the equation (4) states that the rate of increase in photon density is equal to the photon generation by stimulated emission less the loss rate of photons (as characterized by the photon lifetime, τp), plus the rate of spontaneous emission. the parameter v is the active region volume, g0 is the gain slope constant, ε is the normalized gain compression factor, n0m is the electron density at transparency, β is the fraction of the total spontaneous emission coupled at the laser mode, γ is the optical confinement factor, ηi is the injection efficiency, τp is the photon lifetime and τn is the carrier lifetime. the optical output power can be expressed as p=ηhνs, where η = (ηdv )/(2γηiτp), h is the planck’s constant, ν is the emission frequency and ηd is the differential quantum efficiency. in this work, the vcsel model finisair hfe4192-582, operating at 850 nm, was used. the intrinsic parameters extracted, by the frequency subtraction method [8], were then used in the simulation model (see table 1). d. coelho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-16 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc parameter value unit v 2.4x10−18 m3 g0 4.2x10−12 m3s−1 ε 2.0x10−23 m3 n0m 1.9x1024 m−3 β 1.7x10−4 − γ 4.5x10−2 − τp 1.8 ns τn 2.6 ps ηi 0.8 − table 1: intrinsic parameters of finisair hfe-4192-582. since the laser dynamics described by the rate equations are intrinsically nonlinear, harmonic and intermodulation distortion occurs during direct modulation which limits system performance. modelling of the semiconductor laser diode in such a way as to render tractable the accurate computation of intermodulation products (imps) is thus of importance for the design and dimensioning of such systems. 4.2 package and chip parasitics model in the laser model we will consider an intrinsic laser diode (ild), whose dynamic behavior is described by the previous rate equations (equation 3 and 4) and a parasitic interconnection circuit due to the laser assembly in a package. the corresponding equivalent electrical circuit of the parasitics elements of the finisair hfe-4192-582 is shown in figure 5. figure 5: parasitic model. table 2 presents the intrinsic laser parameters of the vcsel operating at 850 nm. combining the parameters obtained for the parasitic circuit and the intrinsic laser model, the global transfer function is obtained as can be seen in figure 6. the relative noise intensity characteristic of this vcsel, was obtained from the rate equation with langevin noise sources been represented in figure 7. in the range of 3 to 9 ma rin varies between -152 to -133 db/hz. parameter value unit rin 50 ω rs 42.6279 ω cs 0 ps cp 1.8068 pf lp1 7.6925 pf table 2: parasitic parameters of finisair hfe-4192-582. figure 6: frequency response of finisair hfe-4192-582. 10 0 10 1 �155 �150 �145 �140 �135 �130 �125 �120 �115 3 ma 4.5 ma frequency (ghz) r in (d b / h z ) 6 ma 9 ma vcsel laser figure 7: relative intensity noise of the vcsel. 5 architecture analysis the maximum fiber length of an optical network to be deployed inside an aircraft is considered to be 100 meters. thus, it is reasonable to neglect both the attenuation and dispersion of rf signals with frequencies up to 10 ghz [9]. as aforementioned, wdm architecture provides a dedicated point-to-point optical channel between each onts and the olt. taking this into account let us consider the point-to-point architecture shown in the figure 8. since the rf signal which arrives at the antenna is a weak signal due to the wireless attenuation, d. coelho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-16 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc the snr in the uplink is considerably lower than in the downlink. the rf uplink signal is generated by the mobile station and reaches the base station attenuated by the wireless channel. the weak rf uplink signal is then electrically amplified (g) before being converted from the electrical to the optical domain by the vcsel. in the central station, the optical signal is detected by a pin photodetector which converts the optical signal to the electrical before reaching the wi-fi receiver module (wifi rx). the rf signal detected will suffer the impact of the rin noise, shot noise, photodetector thermal noise, clipping and intermodulation distortion. vcsel wifi tx pd wireless link fiber p o i 0 mobile station g i s i rx l wifi rx figure 8: point-to-point transmission scheme. the snr for the uplink path, referred at the output of the photodiode optical receiver, can be written as [10], snr = ⟨i2rx⟩ ⟨i2rin⟩+⟨i 2 sn⟩+⟨i 2 th⟩+⟨i 2 imd⟩+⟨i 2 cl p⟩ (5) where the five current noise terms are: the rin noise current, the shot noise current, the thermal noise current from the equivalent resistance of the photodetector (pd) load and amplifier (req), the current due to the third order intermodulation distortions and the current due clipping distortions, respectively. the source thermal noise can be neglected and ⟨i2rx⟩ is the signal power at the receiver as described in [9][10][11]. ⟨i2rx⟩ = 1 2 ( rd µ √ 2 n ⟨p0⟩ )2 (6) ⟨i2rin⟩ = r 2 d ⟨p 2 0 ⟩10 rin 10 ∆ f (7) ⟨i2sn⟩ = 2qrd ⟨p0⟩∆ f (8) ⟨i2th⟩ = 4kt f ∆ f req (9) ⟨i2imd⟩= (rd⟨p0⟩) 2 2 ( µ √ 2 n )6( d111n 2 + d21n ) (10) ⟨i2cl p⟩ = 1 √ 2π λrd⟨p0⟩ µ5 1 + 6µ2 e −1 2 µ 2 (11) the rd parameter is the photodetector responsivity, p0 is the average optical power detected by the pd, ∆ f is the electrical bandwidth of the receiver, q is the electronic charge (1.6 × 10−19 coulomb), k is the boltzmann’s constant, t = 290k, f is the noise factor of the amplifier following the pd and d111 and d21 are the third-order distortion coefficients (imds fi + f j − fk and 2 fi − f j ), which depend on the laser characteristics and operation point. the µ parameter is the total rms modulation index and is equal to µ = m √ n/2, where m is the optical modulation index per subcarrier [11]. the λ parameter represents the fraction of the clipping distortion power which falls in the transmission band which is also dependent on the optical modulation index [12]. for the specific channel allocation, λ = 1.1 × 10−3 for µ = 2%. 6 simulations and results the presented analysis considers the usage of wifi signals in the 2.4 ghz frequency range, with 20 mhz of bandwidth and the use of ofdm with 64 orthogonal subcarriers (n = 64). here we have assumed that the signal directly modulates a vcsel. volterra functional series, described as a “power series with memory”, has been applied previously to assess accurately the laser distortion of the semiconductor laser [13]. the latter analysis enables one to determine adequately the third-order intermodulation coefficients of the semiconductor laser, considering the allocation of subcarriers for wi-fi. 0 20 40 60 80 2 4 6 8 10 −35 −30 −25 −20 −15 −10 subcarrier numberibias (ma) d 1 1 1 (d b ) figure 9: d111 for several bias current. d. coelho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-16 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 0 20 40 60 80 2 4 6 8 10 −40 −35 −30 −25 −20 −15 −10 subcarrier numberibias (ma) d 2 1 (d b ) figure 10: d21 for several bias current. the d111 coefficient has the major impact in the imd limitation since it increases with n2, while the d21 contribution increases with n (eq. 10). also from figures 9 and 10, it is seen that a better performance is expected when the vcsel is operated at 9 ma of bias current and a worst performance at 3 ma, when considering the imd impact on the system performance. for a bias current of 3 and 5 ma, d111 is maximum for the subcarrier number 34 and 35, and equals 0.0522 and 0.0057, respectively. the corresponding maximum values for d21 occur for subcarrier number 63 and 64, and are 0.0081 and 0.0034, respectively. the resonance of the laser may actually change the location within the band (subcarrier) where we would expect the maximum distortion to occur (middle channel for d111 and last channel for d21). the previous theoretical analysis, based on snr, is compared with experimental results. experimentally the performance of the system is assessed in terms of error vector magnitude (evm), which relates to the snr by [14]. ev mrms = 1 √ snr (12) the experimental setup used is depicted in figure 11. it is composed of a vector signal generator (rodhe&schwarz smj 100a) to generate the wi-fi signal, an electrical to optical converter (vcsel model finisair hfe-4192-582), an optical to electrical converter (81495a) and a digital serial analyzer (tektronics dsa 71254c) for the signal analysis and evm measurements. the vcsel laser has a slope efficiency of 0.075 w/a and a threshold current of 0.8 ma and the pin photodetector is considered to have a responsivity of 50 a/w. figures 12 and 13 show the results of both analytical and experimental snr as a function of the total rms modulation index, for the uplink point-to-point pin vcsel 850nm 0.075 w/a slope efficiency and 0.8 ma threshold current rodhe&schwarz smj 100a tektronics dsa 71254c 50 a/w responsivity figure 11: diagram of experimental setup. 10 −1 10 0 10 1 10 2 0 5 10 15 20 25 30 35 40 45 50 s n r (d b ) total rms modulation index (%) curves analytical circles experimental shot rin 3ma clipping imd 3.0ma figure 12: analytical and experimental snr ibias = 3 ma. 10 −1 10 0 10 1 10 2 0 5 10 15 20 25 30 35 40 45 50 55 s n r (d b ) total rms modulation index (%) curves analytical circles experimental rin 5ma shot clipping imd 5.0ma figure 13: analytical and experimental snr ibias = 5 ma. transmission scheme. the noise limiting contributions are plotted in the graph as well. a minimum snr of 20 db can be specified considering a typical sensitivity from a commercial ieee 802.11n of -74 dbm, in the 2.4 ghz band [15]. from the results we can see that the best performance in terms of snr is achieved at high bias currents, when the intermodulation distortion is lower and the performance is limd. coelho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-16 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc ited by both clipping and imd. the maximum snr value for a bias current of 3 ma is 37.12 db for a total rms modulation index of 13%, while for a bias current of 5 ma, the maximum snr is 45.78 db for a total rms modulation index of 11%. measurements of snr above 30 db were not obtained, since the results show a tendency to reach a plateau, which indicate that as the rf input power increases, the photoreceiver’s amplifier limits the power of the signal, which on the other hand limits the snr. let us consider now an analysis for a network with higher fiber length, which can be applied for other wdm-pon network architectures. note that these new results are obtained taking only into consideration the optical loss due to fiber attenuation. from equation 5 and for ibias = 5 ma the snr versus the total rms optical modulation index for several values of attenuation (α = 3, 6, 9,··· , 24 db) is plotted in figure 14. we can see that for low modulation indexes values the snr performance is limited mainly by thermal noise except for the α = 0 db case, where the rin is the leading noise source. 10 −1 10 0 10 1 10 2 0 5 10 15 20 25 30 35 40 45 50 55 total rms modulation index (%) s n r ( db ) optical attenuation increase α = 0 db α = 24 db figure 14: snr as a function of the total rms modulation index for different optical attenuation values (ibias = 5 ma). in figure 15 it is depicted the maximum snr values and the corresponding modulation indexes as a function of the optical attenuation. these modulation indexes are considered to be optimum in the snr sense. the results indicates that the maximum optical attenuation that can be considered for an acceptable minimum snr of 20 db is α = 23 db (for ibias = 5 ma). by considering a multimode fiber with an attenuation of 3 db/km at 850 nm, it corresponds to 7.7 km. by considering a minimum snr of 20 db for a reliable transmission, it is possible to determine from figure 14 the corresponding minimum modulation index (and minimum electrical power) of the rf signal that drives the vcsel. this result gives us the mini0 5 10 15 20 25 15 20 25 30 35 40 45 optical attenuation (db) s n r m a x ( db ) 0 5 10 15 20 25 10 20 30 40 50 60 70 m od ul at io n in de x (% ) fo r s n r m a x snr limit figure 15: maximum snr and optimum modulation index values as a function of the optical attenuation (ibias = 5 ma). mum rf power that can be used to directly modulate the vcsel for an snr of 20 db. figure 16 shows this result as a function of the optical attenuation. 0 5 10 15 20 25 0 5 10 15 20 25 30 m in im um m od ul at io n in de x (% ) fo r s n r = 2 0 db optical attenuation (db) figure 16: minimum modulation index value to achieve an acceptable snr of 20 db as a function of the optical attenuation (ibias = 5 ma). assuming an rf signal with power given by prf = 1 2 ⟨i2⟩ri (13) where ri is the input resistance of the vcsel considered to be 50 ω, the relation between rf power and modulation index is plotted in figure 17. considering a maximum amplifier gain (g) of 30 db and assume that the antenna noise from the ont does not severally affect the snr of the signal, the rf power at the input of the vcsel is given by: pin = pt x,max − l + g + 4 (14) where pt x,max = 13 dbm is the maximum transmitter power defined by the standard, l is the wireless fspl d. coelho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-16 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 0 20 40 60 80 100 −45 −40 −35 −30 −25 −20 −15 −10 −5 0 total rms modulation index (%) p ow er ( db m ) figure 17: rf power as a function of the modulation index (ibias = 5 ma). (free space path loss) defined by: f spl(db)=20log10(d)+20log10( f )−147.55 (15) where d is the wireless link length in meters and f is the rf signal frequency in hertz. additionally, the maximum wireless link length as a function of the optical link attenuation and optical link length for an snr = 20 db necessary for a reliable transmission are plotted in figures 18 and 19, respectively. 0 5 10 15 20 25 0 50 100 150 200 250 300 350 400 450 500 optical attenuation (db) m ax im um w ir el es s ch an ne l le ng th ( m ) figure 18: maximum wireless link length as a function of the optical attenuation for a reliable transmission. as aforementioned, the maximum fiber length of an optical network to be deployed inside an aircraft is considered to be 100 meters (α = 0 db case). looking at the results, for a reliable wi-fi transmission, the system could be deployed with just one access point (or ont), however if another services (wimax, uwb, etc) were integrated there would be necessity to increase the number of onts to achieve its snr requirements. 0 1 2 3 4 5 6 7 0 50 100 150 200 250 300 350 400 450 500 optical fiber length in km (considering only α = 3 db/km) m ax im um w ir el es s ch an ne l le ng th ( m ) figure 19: maximum wireless link length as a function of the optical fiber length for a reliable transmission. 7 conclusion in this article, we have considered the transmission of wifi signals through an optical channel. in particular, we analyze the uplink performance in a point-to-point transmission system for short range networks.a theoretical analysis was performed and a good agreement with the experimental results was obtained. both analytical and experimental results show that, for low bias currents, the intermodulation distortion is the main limiting performance factor at high modulation indexes, whereas the rin is the dominant factor for low modulation indexes. for increasing bias current, the imd distortion decreases and clipping distortion starts to dominate over intermodulation distortion, at high modulation indexes. for a maximum fiber length of 100 meters (aircraft network case), the best performance in terms of snr is achieved at high bias currents. the maximum snr value for a bias current of 3 ma is 37.12 db for a total rms modulation index of 13%, while for a bias current of 5 ma, the maximum snr is 45.78 db for a total rms modulation index of 11%. for higher fiber length wdm-pon networks, the maximum optical attenuation that can be considered for an acceptable minimum snr of 20 db is α = 23 db (for ibias = 5 ma), which correspond a fiber length of approximately 7.7 km. the detailed analysis and system performance presented is adequate for the assessment and design of radio-over-fiber systems in aircraft networks. d. coelho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-16 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc references [1] inescporto, “report on terrestrial networks and architectures,” tech. rep., daphne project, 2009. [2] d. qian, j. hu, p. ji, t. wang, and m. cvijetic, “10-gb/s ofdma-pon for delivery of heterogeneous services,” in optical fiber communication conference/national fiber optic engineers conference, p. owh4, optical society of america, 2008. [3] j. m. b. oliveira, s. silva, l. pessoa, d. coelho, h. salgado, and j. c. s. castro, “uwb radio over perfluorinated gi-pof for low-cost in-building networks,” in microwave photonics (mwp), 2010 ieee topical meeting on, pp. 317–320, 2010. [4] c.-h. lee, w. sorin, and b.-y. kim, “fiber to the home using a pon infrastructure,” lightwave technology, journal of, vol. 24, no. 12, pp. 4568–4583, 2006. [5] j. m. b. de oliveira, receiver design for nonlinearly distorted ofdm signals: applications in radioover-fiber systems. phd thesis, faculdade de engenharia da universidade do porto, 2012. [6] r. westcott, “investigation of multiple f.m./f.d.m. carriers through a satellite t.w.t. operating near to saturation,” electrical engineers, proceedings of the institution of, vol. 114, no. 6, pp. 726–740, 1967. [7] h. salgado, performance assessment of subcarrier multiplexed optical systems: implication of laser nonlinearities. phd thesis, university of wales, 1993. [8] p. morton, t. tanbun-ek, r. logan, a. sergent, p. sciortino, and d. coblentz, “frequency response subtraction for simple measurement of intrinsic laser dynamic properties,” photonics technology letters, ieee, vol. 4, no. 2, pp. 133–136, 1992. [9] i. cox, c.h., e. ackerman, g. betts, and j. prince, “limits on the performance of rf-over-fiber links and their impact on device design,” microwave theory and techniques, ieee transactions on, vol. 54, no. 2, pp. 906–920, 2006. [10] c. h. coxiii, analog optical links: theory and practice. cambridge university press, 2004. [11] a. saleh, “fundamental limit on number of channels in subcarrier-multiplexed lightwave catv system,” electronics letters, vol. 25, no. 12, pp. 776–777, 1989. [12] j. mazo, “asymptotic distortion spectrum of clipped, dc-biased, gaussian noise [optical communication],” communications, ieee transactions on, vol. 40, no. 8, pp. 1339–1344, 1992. [13] h. salgado, “modelling optical/rf impairments in ultra-wideband radio over fiber,” tech. rep., inesc porto, 2007. [14] r. shafik, s. rahman, and r. islam, “on the extended relationships among evm, ber and snr as performance metrics,” in electrical and computer engineering, 2006. icece ’06. international conference on, pp. 408–411, 2006. [15] w. n. corporation: product specifications of dnma92, an ieee 802.11n a/b/g/n mini-pci module, version 1.6, apr. 2009. d. coelho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-16 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc itec issue 13.pdf i-etc launching a new open access academic journal �������� ��� ������ �������� �������� �� ���� � �� �������� �������� ���� ��� ������� ������ ��!� �� �"�� ������#�� abstract$�!� �������� ����%�&�������#��������% ��� ������ &'��(� ��!� �����)���� ���*���� �&&�������������!�����"� &�� ����(��� ���&�%�+ &�,������� ��"���%����������(�)����open journal system���%� �� ��+� ��� ��%�"'����� �����+�����(����&���� &� �!���&�++�� &�� ���� ��%� ��+������ ��# ���� �#� -� ��!�. � ��� ���� / #�� ���� ����� �(� ��# ���� �#� �(� � �"��� -����.� ��� �"��� keywords: *�����&&��� �*�������������'���+ � ��!� 1. introduction the “isel academic journal of electronics, telecommunications and computers” (i-etc) is an international online open access journal that publishes peer ��4 �)�%� ��� &���� �� ���� "���%� 5��%�� �(� ���&���� &� �!���&�++�� &�� ���� ��%� ��+������ ��# ���� �#�� ��!�� ����"� ���%���� �����%� �� ���&&��� "��������#������)�"���#������$66 �% ��� ��� ������6,���������!�����"� &�� ����(� ��!�� ��"���%����*�������������'���+��789 � )� &�� �� ��� ����� ����&�� ��(�)���� (��� ,������� +���#�+���� ��%� ��"� �� �#� %�4�����% � ��������% � ��%� (����'� % ��� "���%� "'� ���� ��"� &� :��)��%#�� ���,�&�� ��%��� ���� ;<=� ;���������"� &�� &����� ��������� ����� ������ �������������������� *���� �&&���� �& ��� 5&� ��"� �� �#� � ���� "�&�+�� �� ��� ��������� !��� ���� " � �'� �((���%�"'�����(������%�+�%����% ((�� ����(� ������������&&���������#������� �'��(������%� +��� +�% ��&�������� ��)����>��)��"'��4��'����� ��'��% ���������������+�4 �#�������� ������&&����+�%�� �������4 % �#�������������((����(����% � �������"�&� "�%���%�������&&���� ,��������� !� �� + ?�%� +�%��� �� "�# �� �#� ��� "�� �((���%� ��� ��������� !��� ����� �&&���� �(� ������� &��� ��#�������'���������%�"'�������'+�����(�����"� &�� ���(����!� �� ����+�� +��� �?�� & ��)���������+�� +��� ����� %%���&����-) �� ��������# ����� ���(��.�(���������"� &�� ��� �(�����+����&� ��� ������&��(����&�����&��% �#��������'�&��� �*�����&&���� ������)�)�'� (�������% ((�� ����(��& ��� 5&���������� ������������ ����� ����������� � @ ��� ���� ��"� &�� ��� �(� � ��!�� )�� )���%� � >�� ��� ������ ���� ���& 5&� 4 � ��� ��� �� ����",�&� ����%�& �#���+�% ��)� &�� ��������",�&������������a��&���+ &������ &� ������%� ����� �������(����'��4� ��"��������'� ��������%����%����*��� ����� ���"�� �%�����&���� ����(� special inauguration issue vol. 1, nº 1, id-1 (2012) http://editorial.isel.pt/journals afantoni stamp a.fantoni et al. | i-etc launching a new open access academic journal i-etc, volume 1, nº 1, special inauguration issue (2012), paper id-1 �� ����)��&�%�+ &�,������� �����# 4� �+� ��'����'���#�������&����� ���� ���& ��� 5&����� � ��� ���'� �&&���� ��� ���� ���� " � �'� �(� ��"+ �� �#� ��� �� )��>� ��� �� � #������ �& ��� 5&� ����� ��4 �)�%� ��"� &�� ������� ��� ��"� &�� ��� (���� ���� ���� ��%� (��+� �������� -��� ������ �� �� �� 5���������.�)��� +�����((���������������� ���������������"�����(��)��)���%�$����#����%�(���� % ((�� ��� �(� �������� -����� �&&���� +�%��. � ��� ����� &� ���� ��� � +����&� ��� ���#��� ��%� ��� ��"� &�� ���%���'�-��"� &�� ��� ��%�����?&��� 4��'���� ��������#����%�% &���%�)�"����4��.� *�������� �������4���� ��,������� ������� �����'� �%�?�%� ����%������# 4��4 � " � �'�������� ��"� ���%���� &������ �&�� ���%�)��� ��!��) ���"�� �%�?�%��������directory of open access journals � �����%� "'� ���� ���%� =� 4��� �'� � "���'� 7b9� ��%� ��� ���� (�����&&��� "��� google scholar� )�"� ����&�� ��# ��� 7d9�� ��&�� ������#'� ) ��� ���"��� ���� �&��� ��� �(� ���� ������� ��"� ���%�������� ��!��,�������������������������%�&�%�"'���������&����# ���publish or perish 7f9���%� �������'� ���(������� +��&����%�% ((�� ����!����(��� �� �&������5����'�����(� ��"� &�� �� �������� &������"� ���%���� ��!��) �����4���������� " � �'�������%�&����#��%� �((�&���������&������&� ����(������ %��& ��� 5&�&��� &���+�� ���� %�� �#����������#�&�+��� � ������������)��&�%�+ &�,�������(�&���� ���� ���& ��� 5&� ���� �)�������������%�&������� ��� ��(5& ������%����������+��� +����� %���4 �) �#����&����� ��4 �)���� ) ��� "�� ��>�%� ��� �������� ) ��� %��� �� ��� �� &�++����� ��%� ��##��� ����� @�� �?��&�� ����� +��'� �(� ���� ���� "��� �������� ��"+ �� �#� +����&� ���� ��� �� �� ,������� ) ��� "�� '���#�������&������!�����4 �)����&����) ���"�������(������%�#�# &���'�%� 4�� ���� +���4�� ���� ���� �'� �(� ���� +����&� ��� ������� ����� ��� +����'� �&&���� ��� ��(���� ����@�� "�� �4�� ����� ���� ��"� &�� ��� �(� �� ���� �(� ����&��%� ������� (��+� ���� &��(����&�� ��!�� bg88� ) ��� "�� ��� �?&����������������(�����������������(��������5�����)��'������(���"� &�� �� � �� ������ ����� ��� ��� ����'� (��� �%�? �#� ) ��� isi thompson web of knowledge 7h9 � )� &�� &��&������� ��%� ��� #����������� +��&��(�&������ =�� ������ �)��)��&�+�������4���������"+ �� ����(�'������������(�����"� &�� ��� ���������)�"���� ��!�� 2. aim and scope of the journal !�����������(���������� # ����&���� "�� ������%���4 �)���� &�����(�������� &�����%� �?��� +������������&��"�����������(��%�+��������%����� �%���4������� ��+� ��� ��%�"'����� �����+�����(����&���� &� �!���&�++�� &�� ������%���+��������# ���� �#���!�. ���� ����/ #������ ������(���# ���� �#��(�� �"���-����.� ��� �"�� �) ��������+�������% ��� ��� "���%����)���������� ������� ������% ��� ����%4 ���'�"���% �)� &������#���������������� #�� ���� �'��(������% ��� ������ &'� ��������� ���,������ ���� &������"� ���%���� ��!����������� + ��%�"'��������� �� �#� &������ ���� ��%� �������� ���� ��&����#�%� ��� ��>�� �%4����#�� �(� ���� ,������a�� +��� +�% �� ������&��� (��� �� "������ ��%� &������� ��������� ��� �(� ��� �� ��������� �������� ���� ����)�%� ��� ����+��� +�% ��������� ���&�����������4 %������%�&�+������#���� &���� +�� �������"������ &��4�'�&�+���?� �(��+�� �������������%��� isel academic journal of electronics, telecommunications and computers http://editorial.isel.pt/journals �������� �������� �� � ��!��&��� %����������"� &�� ����(�����(����) �#��'�����(���� &��$ �� �������&����� &��� �� ���4 �)���� &��� �� �������&�++�� &�� ��� �� ������������% ���� ����%% � ��������#����� �������(�&���� "���%������� �����,���������"� ����$ �� ����&��%����&��% �#���(��������(����&��������&���� &� �!���&�++�� &�� ���� ��%���+������ ���#�� ��%���� �% &���'�������������-/ #������ ������(� ��# ���� �#��(�� �"���. �� ����&��%����&��% �#���(��������& ��� 5&�+��� �#��) �� ���������������&���� �� ���& ��� ������������ &���(�&������� �������� 3. editorial policies ���������� � reading articles published in �� !�� ���"� �#���������� �� ����� $� ���� �""��� ����� �����%�"����� � is free of charge� i-etc is an open access ,��������!� �� ��&��� %���%������� ��� ����%����������)�(������%��(( & ����% ���+ ��� ��� �(�����������a��������&�����������!������ ���&���������"� ���%� �� ��!�� ��(����'���%� ���+������'��&&��� "������ ��� ++�% ����'��(��������( ��������( �#����#������������(� articles published in the isel academic journal of electronics, telecommunications and computers� ���� �� &��'� #��� �(� ��� �� )��> � � &��� �#� �� ��%��� ���� creative commons attribution-noncommercial 3.0 unported license� 7i9�� !� �� � &����� ����)�� (���� %�)����%� �(� ���� ��� &���� (��+� ���� ��!�� )�"� �� � ��� )���� ��� ������� ��%� ���% ��� "�� ���) ����������� &� �� �������#���������� # ����)��>� ���������'�& ��%���%� �������%�(���&�++��& ������������� publishing in the �� !�� ���"� �#���������� �� ����� $����� �""��� ����� � ����%�"����� is free of charges�� ��!��%��������&���#����'���� &������&��� �#�(����!��� �& ��� 5&�4�������%� ���������(�������"+ ���%���� &���������������'�&� ��� ��(�����"� &�� �� � &��� %���%�"'������% ��� ���"���%� ����&���� ������&��� � ��'� +����&� ��� ��� ��"����� ��� ������ �(� � � ��"+ ���%� ��� ���� ,������� +���� ���� "�� ��%��� &��� %���� ��� (��� ��"� &�� ��� "'� ��'� ������ ,�������� ��� #������ � ���� +����&� ��� �����%�������4�������%'�"������"� ���%� ����'�,����������������& ��"���(��+��������������� ���� ��%� ��� ������� ����� ��� +���� ��� ��"+ ���%� ��� ����� �(� �� +����&� ��� �(� �#��� �? �� �#� &��'� #��� � ��� ���� � #���� �(� �� �� �%� ����'�� !��� isel academic journal of electronics, telecommunications and computers��&&�����(�����"� &�� ���+����&� ���(��+���'�������� a.fantoni et al. | i-etc launching a new open access academic journal i-etc, volume 1, nº 1, special inauguration issue (2012), paper id-1 ) �� �#� ��� ��"+ �� ��� &��� (��� ��4 �) �# � ��#��%����� �(� ��? � ��&� � ��� ���� �' � ���#��#� � ��� # �� ���� � &����� � �� ������������%���& ���&��% � ���� ��!������������� �#���&����%��������4 �)��'���+����&��+����&� ��� ����4 �)�%���� ������"'����������(������4 �)���� �� � �#���"� �%�j�!���+����&� ������%�����"��"� �%�%�)�������������(�����4 �)k� ������4 �)������������'+�������������������"���������4 �)����>��)������������� �� !)�����#�� ��4 �)� j� !����� ������� ����� ���� ��� ��4 � ��� ��� ��##����%� "'� ���� ��4 �)��� +�'� "�� ����� "�&>� ��� ����� ��+�� ��4 �)��� �(���� ��4 � ���� ���� + ���� ��4 � ������' ������% ����+�' &��&>�������4 ��%�4��� ����(�����+����&� ���� �(��������+����&� ��� ����"+ ���%�4 ��������� ����'���+ � �� ��&��&>�%�"'������% ������� ���� (� ��+����������&�����(���������������%� ���(��+������� ��+������ ����&� ������������� ������ ��"���(�����"� &�� ��� ������������������ ++�% ����'���,�&��%� �(�����+����&� ���&��(��+���������� +����%��&�����(������������ ���%�(��+���'������&��� ����������&� �������������� ������% ������� #���������4 �)������%�����+����&� ��� �����&�%� ��%�����4 �)��l���%����������4 �)���a�&�++����������% ����+�>�����%�& � �����$ �� �&&��������+����&� ���) ������(���������4 � �� �� �&&�����(����-+ �������+�,��.���4 � ��� �� ��>����������������"+ ����%���4 �)���������4 �)���&���#� � -+�,�����4 � ���. �� ��,�&� ����������%���4 �)����) ���"�����+���'�&����&��%�"'������% ������%� �(��+�%��"�������� 5����%�& � ���� 4. references 789����@ �� ��>' �� "���'�/ �!�&� �����+��bd �������f ����hgf�h8m�-bggh.� 7b9�����$66)))�%��,���#6 7d9�����$66�&������#��#���&�+ 7f9�/��� �# ���@��-bggn.���"� ��������� �� �����$66)))����� �#�&�+6������+ 7h9�����$66�& ��&�����+�����������&�+6+,�6 7i9�����$66&���� 4�&�++������#6� &�����6"'��&6d�g6� comparing layer 1 and layer 3 relay stations deployment in a lte network comparing layer 1 and layer 3 relay stations deployment in a lte network andré martins1,2 1instituto de telecomunicações (it), lisbon, portugal 2instituto superior técnico (ist), lisbon, portugal andre.epmartins@lx.it.pt pedro vieira1,3, antónio rodrigues 1,2 3instituto superior de engenharia de lisboa (isel), lisbon, portugal pvieira@deetc.isel.pt, antonio.rodrigues@lx.it.pt keywords: wireless communications, cooperative communications, lte, relaying, fixed-relay. abstract: the relay solution in planning of mobile networks, has the aim of increasing the network coverage and/or capacity. according to the open literature, this technique will be highly used in the next long term evolution (lte) networks. the relay station (rs) performance varies with its position in the cell, with the radio conditions to which rss and user equipments (ues) are subjected and with the rs capacity to receive, process and forward the information. the aim of this paper is to compare the performance of the layer 1 (l1) and layer 3 (l3) rs types, and to determine the ideal position in which a rs should be placed, with the aim of maximizing the ue throughput. 1 introduction over the past few years, mobile communications have suffered great technological advancements. if a few years ago a mobile user was satisfied to make a voice call with a reasonable quality, nowadays, most users demand to see high-definition video in real-time. besides, the cloud paradigm shift will become more and more relevant. due to this increasing quality of experience (qoe) demand from users, and also to offer new features with the aim of stimulating the market, the providers are forced to implement new technologies and advanced features. the 1st generation (1g) cellular mobile network was deployed in 1981, where connections between terminals had a highly variable quality, due to interference and equipment limitations. in 1990 was launched the 2nd generation (2g) of mobile communications. it is stated by many that 2g has revolutionized the world of mobile communications, and the truth is that his successor, the 3rd generation (3g), failed to migrate as many users as would be expected. 2g was also responsible for introducing the short message service (sms), which continues to be highly used. when in 1998 was launched the 3g, the concept of mobile broadband has become part of users everyday life. this generation enables downlink peak data rates up to 14.4 mbps (considering the high speed downlink packet access (hsdpa)). 3g is also responsible for supporting a mobile video call or accessing the internet with a very acceptable quality, to turn this feature within the reach of most users. towards 4th generation (4g) several telecom standards were developed: lte [1], by the 3rd generation partnership project (3gpp) family, and mobile worldwide interoperability for microwave access (wimax) (ieee 802.16e), developed by institute of electrical and electronics engineers (ieee) are the most important examples. the main goal of lte is to improve the user data rates. the theoretical throughput are up to 100 mbps in the downlink and up to 50 mbps in the uplink. in addition, the latency round trip time has decreased from 50 ms (in high speed packet access evolution (hspa +)) to approximately 10 ms. in order to fulfill the main goals, in lte appear several improvements compared with previous releases, such as bandwidth aggregation, enhanced multiple antenna technologies, coordinated multipoint transmission (comp) and relaying. this paper will focus on the ue throughput enhancement using fixed relays. for this purpose, comparisons between scenarios where relays are not i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-15 http://journals.isel.pt/index.php/iajetc used, with scenarios where a relay, either the l1 or the l3 type is used, were set. the main issues addressed in this paper are: a) what is the rs with the best performance, b) what is the ideal position to deploy both rs in order to maximize the ue throughput and c) in which conditions a rs should be used instead of its donor enb (denb). the paper is organized as follows. section 2 introduces relaying technique concept, followed by the simulation system model described in section 3. section 4 explains the simulator adaptation and, in section 5, are presented the simulation results and analysis. finally, the conclusions are drawn in section 6. 2 relaying technique as in [15], there are three schemes for cooperative mimo (co-mimo) [14] cellular systems: comp, fixed-relay and mobile-relay. in the fixed-relay scheme [16], fixed relay stations (frss) are used to send the signal to ue from the evolved nodeb (enb) and vice-versa. this type of co-mimo can be used to increase the coverage [12] and network capacity [4] in a specific area, and to improve the qoe. on one hand, this scheme is very appealing because it can not only exploit the spatial diversity gain, but also to increase the network performance or even expand the coverage area, at a low-cost. according to [10], the rs capital expenditure (capex) and operating expense (opex) values range from 1.8% to 3.8% and 5.3% to 6.15%, respectively, depending on its transmit power, compared to an enb typical cost. on the other hand, this type of equipment introduces an additional delay in the network, and can create interference problems reusing the spectrum in the rs. from the ue’s point of view, the rss are classified into two types: • type i – for the ue, a type i rs, currently in standards development for long term evolution advanced (lte-a) rel.10 [3], is treated like an enb. thus, the rs creates its own cell, i.e., transmits its own identity number, reference signals and synchronization channels. one of the requirements of lte relay solutions, is that the rss should be transparent for the ue. thus, it is ensured that the release 8/9 terminals, may be served by rss introduced in release 10. • type ii – using a type ii rs, the ue is not able to distinguish a rs from the enb, because the rs does not have its own cell identifier. when a rs of this type is used, the control information can be transmitted via enb and the user data via rs. therefore, it is possible to consider that the type i is the most suitable in order to extend the network coverage, and the type ii aims to create hotspots, i.e., improve the network’s capacity. a basic scheme of both types is illustrated in figure 1. type ii type i access link relay link figure 1: relays types. the rss may also be classified depending on the function they perform as network nodes. according to [6], the rss may be divided into three main categories: • l1 – this rs type, also known as amplify and forward (af), has the function to amplify the received radio frequency (rf) signal and send it to the next hop, which can be another enb, a rs or an ue. the associated problem with this type is that by amplifying the signal, noise and interference are also amplified; • layer 2 (l2) – the second rs type, designated by selective decode and forward (sdf) performs functions of decode and forward, having more freedom to optimize the network’s performance. it extracts the received data packets, processes, regenerates and delivers them to the next hop. hence, its processing delay is longer than the l1 rs type; • l3 – another name for this type of relay is demodulation and forward (df). a rs of this type operates in the same manner as the l2 rs but it acts like an enb, since it has own cell identifier. it achieves, therefore, better results when compared to the other two types, but it is a more expensive and complex solution. this rs type have been standardized for lte-a rel.10 [2]. 3 system model the results presented in this paper were obtained using a lte link level simulator [11], which was adapted to support the existence of a rs in the network. in order to evaluate the rs deployment benefits, simulations for the enb-rs, rs-ue and enbue links were made. in both situations, simulation is based on snapshots where the ue is positioned at the cell edge, i.e., considering the worst case. a. martins et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-15 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc in the scenario where the rs was utilized, its position ranged from 25 to 725 m, as illustrated in figure 2. in the sequence, for each distance in which the rs is deployed, the following data were simulated: signal-to-noise plus interference (snir); used channel quality indicator (cqi); hybrid automatic repeat request (harq) mechanism performance; ue bit error rate (ber); ue throughput. enb ue 750 m rs figure 2: simulation scheme. in the enb-rs link, the transmission is done with the most suitable modulation order according to the snir. one one hand, using a l1 rs, the used modulation order in the the rs-ue link is the same that was used in the enb-rs. on the other hand, when a l3 rs type is considered, the used modulation order in the links can be different, because a rs of this type has the capability of adapt the transmission parameters (e.g., order modulation and code rate) according to the received radio channel measures, reported by the ue. the simulations were implemented according to the parameters of table 1. 4 simulator improvement the used link level simulator only implements an enb to ue direct link. in order to insert one rs in the enb-ue chain, was necessary to develop a new rs module which was added to the existing simulator. this module works starting from the physical layer and considers two relay approaches, l1 and l3, already presented in section 2. concerning to l1 implementation, the task was simpler when compared with the l3 approach, since it was only necessary to split the simulator way of implementation in the number of radio links connecting the different network nodes. regarding the l3 approach, extensive simulation work was set. a l3 rs type implements cqi adaptation, in parallel with harq, transport block (tb) segmentation and decoding/coding procedures, between the incoming and outcoming signals. the ue throughput (expressed in mbps), t h, is given by general bandwidth 20 mhz frequency 2.6 ghz cell radius 750 m channel model winner ii [5] b1 (enb-rs link) b5c (rs-ue link) c2 (enb-ue link) scheduling round robin max. harq retrans. 3 background noise −174 dbm/hz transmission mode td cqi [1:15] height of roofs 15 m antenna configuration 4 × 2 (enb-rs link) 2 × 2 (rs-ue link) 4 × 2 (enb-ue link) enodeb height 25 m transmit power 37 dbm relay station height 10 m transmit power 30 dbm distance from enb [25 : 725] m types l1 & l3 user equipment height 1,5 m distance from enb 750 m table 1: simulation parameters. t h = nb nf × tsb × 10−6 (1) where nb is the number of bits in the transmitted sub-frame, tsb is the lte sub-frame duration in seconds (10−3) and nf is the sum of the transmitted frames in the link(s), including the retransmitted frames. therefore, nf can be defined as nf = ∑ ( nenb−rsf + n enb−rs rf + nrs−u ef + n rs−u e rf ) (2) with nenb−rsf and n rs−u e f the number of transmitted frames in the enb-rs and rs-ue links, respectively, and nenb−rsrf and n rs−u e rf represent the a. martins et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-15 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc sum of the retransmitted frames in the enb-rs and rs-ue links, respectively. the original simulator calculates the throughput in function of snir. so, in order to obtain the results as a function of the distance, as shown in section 5, it was necessary to relate snir with the distance. the relation between snir and distance can be calculated as [7] snir = p ( 10 pls 10 ) n0w + 6 ∑ i=1 ( pi ( 10 pli 10 )) (3) where p is the transmit power for the enb or the rs, in the enb-rs and the rs-ue links, respectively. pls corresponds to path-loss for the enb-rs, rs-ue or enb-ue links, and pli to path-loss for the kth enb interferent. finally, n0 and w is the thermal background noise and the channel bandwidth, respectively (the expression is calculated in linear values). as known, the transmission’s quality depends heavily on the propagation conditions that a certain scenario presents. therefore, it is extremely important the transmitter’s and receiver’s location, whether or not in line-of-sight (los) situation. on one hand, as the rs position is chosen by the network provider, it can be strategically placed in a position where it is in a los situation with the enb, in order to maximize its performance. on the other hand, the ue’s conditions constantly changes due to its mobility. therefore, in many cases, despite the shorter distance between the rs and the ue, they are in a non-line-ofsight (nlos) situation. in order to increase the simulation realism, the los model probability presented in [8] was implemented in both links. 5 simulation results and analysis because the rs performance is heavily dependent of the scenario conditions, the comparison between the two studied rss types is based on two scenarios with different characteristics. in the first scenario, the rs is in a los situation with the enb when it is located at a 25 m distance from each other. on the other side, in the second scenario, the enb and the rs are is a los situation at 25 m enb-rs distance and, in addition, the rs and the ue are in a los situation with the enb and the rs, respectively, when the rs is deployed at a 325 m distance from the enb. for the others snapshots, these elements are in a nlos situation between each other. the snir value, in the both links, as a function of the rs position for scenarios 1 and 2 is presented in figures 3 and 4, respectively. in both mentioned figures is also presented the minimum snir value to use the lowest cqi value, according to [13]. 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 20 10 0 10 20 30 40 50 60 70 ← lte minimun snir enb rs distance [m] s n ir [ d b ] enb rs link rs ue link figure 3: snir in the first scenario. 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 20 10 0 10 20 30 40 50 60 70 ← lte minimun snir enb rs distance [m] s n ir [ d b ] enb rs link rs ue link figure 4: snir in the second scenario. from figures 3 and 4, we can see that the snir in the enb-rs link decreases as the distance between each other increases. on the other hand, the quality of the rs-ue link is improved with the greater proximity to the ue. further analysis showed that the los situation improves the link’s quality. comparing the snir value in two presented scenarios, it is possible to see that it increases approximately 30 db in the snapshot which corresponds to the los situation, in both links. figures 5 and 6 present the used cqi in both links when a l1 rs type is used, in the first and second scenarios, respectively. in both figures, the correspondence between the cqi value and modulation order is also presented. 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 1 3 5 7 9 11 13 15 enb rs distance [m] u s e d c q i ↑ qpsk ↑ 16 qam ↑ 64 qam enb rs link rs ue link figure 5: used cqi in l1 rs test in the first scenario. a. martins et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-15 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 1 3 5 7 9 11 13 15 enb rs distance [m] u s e d c q i ↑ qpsk ↑ 16 qam ↑ 64 qam enb rs link rs ue link figure 6: used cqi in l1 rs test in the second scenario. in both figures, it is seen that the used cqi by the enb and the rs is the same, due to the utilization of a l1 rs type. it can be seen from figures, that as the enb-rs distance increases, the value of the used cqi decreases, because the enb-rs link’s quality also decreases, forcing the adoption of a lower modulation order or a higher coding rate. despite of the snir in the link rs-ue increases, the used cqi by the enb and the rs is the same, by virtue of the rs-ue link’s quality is not being considered. the main difference between the created scenarios is that in the second scenario, when the enb and the rs are separated by 325 m, due to the los situation, the snir of the enb-rs link is enough to use a cqi value of 15. as it will be shown, the link’s quality enhancement has also repercussions in the ue ber, the number of retransmitted frames and in the ue throughput. similarly to figures 5 and 6, in figures 7 and 8 is presented the used cqi value by the enb and the rs, when a l3 rs type was deployed in the first and second scenarios, respectively. 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 1 3 5 7 9 11 13 15 enb rs distance [m] u s e d c q i ↑ qpsk ↑ 16 qam ↑ 64 qam enb rs link rs ue link figure 7: used cqi in l3 rs test in the first scenario. the presented data in figures 7 and 8, indicate that the used cqi in the enb-rs link is exactly the same to the one used in the l1 rs type simulation, due to the cqi value being adapted in the enb-rs link, independently of the rs type. on the other side, in contrast to the results present in figures 5 and 6, the used cqi in the rs-ue link is the one that guarantees a block error rate (bler) value lower than 10%, whenever it is possible. from figures 7 and 8 we observe that as the enb-rs distance increases, the 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 1 3 5 7 9 11 13 15 enb rs distance [m] u s e d c q i ↑ qpsk ↑ 16 qam ↑ 64 qam enb rs link rs ue link figure 8: used cqi in l3 rs test in the second scenario. used cqi by the ue gets higher. this increase is related to the the proximity between the rs and the ue, resulting in a higher ue snir. once again, a comparison between scenarios reveals that a los situation causes a significant increase in the used cqi value. further analysis showed that the used cqi in the rs-ue link suffers a considerable increase: in the first scenario the used cqi was 4 and, in the second scenario, it was 15. it is advisable to remember that, in this snapshot, the ue snir increased approximately 30 db. another achieved result by a suitable transmission parameters adaptation is the lower ber values. as it was mentioned, the chosen cqi value will be the one that ensures an ue bler lower than 10%, whenever it is possible. figures 9 and 10 present the ue ber when the l1 and l3 rss types were used, in the first and second scenarios, respectively. 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 0 5 10 15 20 25 30 35 40 45 50 enb rs distance [m] b it e rr o r r a te [ % ] l1 rs l3 rs figure 9: ue ber in the first scenario. 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 0 5 10 15 20 25 30 35 40 45 50 enb rs distance [m] b it e rr o r r a te [ % ] l1 rs l3 rs figure 10: ue ber in the second scenario. the most striking result to emerge from the simulation results is that the ue ber, when a l3 rs type is used, is lower than when a l1 rs type is consida. martins et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-15 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc ered, in both scenarios. the l3 rs type does not always ensure an ue ber equal to zero, but it achieves a much lower error data rate than the l1 rs type. as shown in figures 9 and 10, for short enb-rs distances, i.e., long rs-ue distances, the error data rate is lower using a l3 rs type, for the reason that the ue snir decrease is balanced with a suitable cqi value to the radio propagation conditions. nevertheless, for larger enb-rs distances, the ue ber tends to be equal to zero for both rss types. the l1 rs type may also achieve a null error data rate as result of the used cqi in the enb-rs link can also being used in the rs-ue link, due to the ue snir increase. regarding to the los snapshot, the ue ber is null, for both rs types, as shown in figure 10. while in the l3 rs type test, the ue ber decrease is not significant (from 0.10% to 0%), in the l1 rs type, there is a decrease of approximately from 0.35% to 0%. from this discussion, it is also proven that the rs position plays a major role in its performance. finally, figure 10 reveals an expected result when the distance between the enb and the rs is equal to 225 m. in this snapshot, the ue ber is equal to 0.2% while, in the previous simulated distance, the ue ber was equal to 0%. if a detailed analysis in figure 6 is made, we can observe that the used cqi for the rs-ue link is higher than the used cqi in the following snapshot (275 m). this particularity may be classified as an inaccurate parameters transmission adaptation, which caused a high ue ber value. one of the ways to test the precision of the cqi value adaptation is to analyze the harq mechanism’s performance. the aim of this technique is to correct errors, using frame retransmission. therefore, the number of retransmitted frames conduces us to conclude about the cqi adaptation accuracy. if there is a higher number of retransmitted frames, the used modulation order is too high for the conditions that the radio channel presents. on the other side, if there is a high percentage of non-retransmitted frames, the chosen cqi value can be used. in order to analyze the harq mechanism’s performance, a histogram of the retransmitted frames class occurrences, considering the l1 rs deployment in the first scenario, is presented in figure 11. for each snapshot, is shown the percentage of frames that were not retransmitted, and the percentage of frames that were retransmitted one, two or three times by the rs to the ue. as showed in figure 11, when the l1 rs type is used, the number of retransmissions is always three for enb-rs distances until one half of the cell’s radius. this means that the used cqi is too high for the rs-ue link’s quality, since the ue can not success3 2 1 0 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 0 10 20 30 40 50 60 70 80 90 100 p e rc e n ta g e [ % ] #rtxenb−rs distance [m] figure 11: retransmitted frames in l1 rs test in the first scenario. fully received the data sent by the rs, not even using the harq mechanism. figure 11 also indicates that as the rs gets away from the enb, the number of retransmitted frames between the rs and the ue has a tendency to decrease, for the reason that the used cqi in the enb-rs link also decreases and the rsue link’s quality increase. as a result, the used modulation order can be the same in the both links. figure 12 depicts a histogram of the retransmitted frames class occurrences, considering the l3 rs deployment in the first scenario. 3 2 1 0 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 0 10 20 30 40 50 60 70 80 90 100 p e rc e n ta g e [ % ] #rtxenb−rs distance [m] figure 12: retransmitted frames in l3 rs test in the first scenario. the graph shows that there has been a sharp drop in the number of retransmitted frames. in addition, it is noteworthy that, despite of still existing retransmitted frames three times, the l3 rs type deployment causes a high percentage of never repeated frames for lower snir values. nevertheless, for longer distances between the rs and the ue, it is seen that all the frames have to be repeated at least once. therefore, using a l3 rs type and even with a suitable cqi value, there is no guarantee that the ue will not use the harq mechanism. similarly to figures 11 and 12, figures 13 and 14 illustrate the harq’s mechanism performance in the second scenario for the l1 and l3 rs types, respectively. a. martins et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-15 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 3 2 1 0 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 0 10 20 30 40 50 60 70 80 90 100 p e rc e n ta g e [ % ] #rtxenb−rs distance [m] figure 13: retransmitted frames in l1 rs test in the second scenario. 3 2 1 0 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 0 10 20 30 40 50 60 70 80 90 100 p e rc e n ta g e [ % ] #rtxenb−rs distance [m] figure 14: retransmitted frames in l3 rs test in the second scenario. comparing figure 11 with figure 13 and figure 12 with figure 14, at the los snapshot (325 m), we can observe that the number of retransmitted frames decreased to zero, due to the good radio propagation conditions. therefore, the simulation results indicate that the snir increase due to the los situation is enough for the ue to discard the harq mechanism. this result should be taken into account by the providers in the rs deployment phase, since it is a key factor in the rs’s performance. finally, in figure 15 it is shown the ue throughput as a function of the enb-rs distance, in the first scenario when it is served by a rs. for comparison purpose, the ue throughput when a rs is not used is also present. it is important to underline that, in both cases, the ue is located at the cell edge. 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 0 2 4 6 8 10 12 14 enb rs distance [m] t h ro u g h p u t [m b p s ] ← ue throughput w/o rs l1 rs l3 rs figure 15: ue throughput in the first scenario. as it would be expected, the ue peak data rate was achieved when the l3 rs type was tested. furthermore, in the l1 rs type context, the ue throughput only has a value different from 0 mbps, when the rs is positioned at enb-rs distances higher than one half of the cell’s radius. this is due to the fact that from this distances, the used cqi in both links tends to decrease, while the ue snir increases. consequently, the radio propagation conditions is going to reasonable for an ue throughput different than 0 mbps. in the last two snapshots (675 and 725 m), the ue throughput decreases considerably for both rs types. this is related to the choice of a lower modulation orders in the enb-rs link and the increase of the retransmitted frames between these two nodes. although the number of retransmitted frames in this link is not presented in the paper, simulation results indicate that 40% of the frames which were sent by enb to rs were retransmitted one time and 60% two times, in the last snapshot. for this reason, it is not advisable to install the rs exactly at the cell edge, due to the considerable decrease of the enb-rs link’s quality. figure 16 illustrates the ue throughput in the second scenario, as a function of the enb-rs distance when it is served by a rs. the ue throughput when it is served by the enb is also depicted. 25 75 125 175 225 275 325 375 425 475 525 575 625 675 725 0 5 10 15 20 25 30 35 40 enb rs distance [m] t h ro u g h p u t [m b p s ] ← ue throughput w/o rs l1 rs l3 rs figure 16: ue throughput in the second scenario. the results show that in the los situation snapshot, the two rss types achieve the same ue throughput. regarding to the l1 rs type, the ue throughput is 0 mbps until this snapshot, being 0 mbps again in the next one, where a nlos situation is again considered. this means that the circumstance of the rs being closer to the ue, is not a necessary condition to a maximized ue throughput. it is important to underline that the snir increased more even with a los situation than to shorter enb-rs distances. despite of the diminution of approximately 30 mbps between the los snapshot and the next snapshot, figure 16 proves that the l3 rs type always promises data rates different than 0 mbps, as seen in the previous scenario. taken together, these results suggest that the l3 rs type has always a better performance than the l1 a. martins et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-15 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc rs type, due to is ability to adapt the transmission parameters according to the rs-ue radio channel quality. nevertheless, the l1 rs type may be used when the quality of the rs-ue link is higher or equal to the one of the enb-rs link, achieving an improvement of the ue throughput even with a less complex rs and, consequently, a cheaper solution for operators. regarding to the l1 optimized position, it should be deployed at one half of the cell’s radius. on the other hand, if the l3 rs type is considered, it should be placed at three quarters of the cells’s radius, counted from the enb to the cell edge. these values correlate favorably with [9] and further support the idea of the ideal position for the l3 rs type is close to the cell edge. finally, from figures 15 and 16 we can conclude that from distances above to approximately one half of the cell’s radius, the ue achieves a higher throughput when it is served by a l1 rs type than when served by the enb. on the other side, if a l3 rs type is considered, the enb to rs handover algorithm should be calibrated to decide that the users located at distances above one fifth of the cell’s radius (counted from the enb to the cell edge), will be served by the rs instead of by the enb. 6 conclusions the purpose of the current research was to compare the l1 and l3 rss types performance, to determine the ideal position to deploy both rs types, and in which conditions the ue should be served by the rs instead of the enb. one of the most significant findings to emerge from this study is that the l3 rs type achieves a better performance when compared to the l1 rs type. nevertheless, the l1 rs type deployment should be considered since it is a cheaper solution for providers. this research has shown that at distances equals to one half of the cell’s radius, both rs achieve the same performance. in what concerns to the ideal position, the results supports the idea that it is related to the rs type. on one hand, the l1 rs type should be deployed at one half of the cell’s radius. on the other hand, if the l3 rs type is considered, it should be placed at three quarters of the cells’s radius. finally, it is also shown that the ue throughput when it is served by the rs is higher than when served by the enb from distances above to one fifth of the cell’s radius (counted from the enb to the cell edge) and one half of the cell’s radius for l3 and l1 rs types, respectively. references [1] 3gpp. requirements for evolved utra (e-utra) and evolved utran (e-utran). tr 25.913, 3rd generation partnership project (3gpp), march 2006. [2] 3gpp. rp-091434, relays for lte. rp 091434, 3rd generation partnership project (3gpp), 2009. [3] 3gpp. requirements for further advancements for eutra (lte-advanced). tr 36.913, 3rd generation partnership project (3gpp), march 2011. [4] k. balachandran, j. kang, k. karakayali, and j. singh. capacity benefits of relays with in-band backhauling in cellular networks. in communications, 2008. icc ’08. ieee international conference on, pages 3736–3742, 2008. [5] l. hentilä, p. kyösti, m. käske, m. narandzic, and m. alatossava. matlab implementation of the winner phase ii channel model ver1.1 [online]. glasgow, scotland, december 2007. [6] m. iwamura, h. takahashi, and s. nagata. relay technology in lte-advanced. special articles on lte-advanced technology ongoing evolution of lte toward imt-advanced 2, ntt docomo, sep.2010. [7] f. khan. lte for 4g mobile broadband: air interface technologies and performance. cambridge university press, 2009. [8] p. kyösti, j. meinilä, l. hentilä, x. zhao, t. jämsä, c. schneider, m. narandzić, m. milojević, a. hong, j. ylitalo, v. holappa, m. alatossava, r. bultitude, y. de jong, and t. rautiainen. winner ii channel models. part ii. radio channel measurement and analysis results. technical report, ec fp6, september 2007. [9] a. martins, a. rodrigues, and p. vieira. finding optimized positioning for fixed relay stations in a cooperative lte network. in wireless personal multimedia communications (wpmc), 2012 15th international symposium on, pages 316–320, 2012. [10] a. martins, p. vieira, and a. rodrigues. avaliação do impacto económico da utilização de repetidores fixos numa rede lte. in proc ursi seminar of the portuguese committee, 2012. [11] c. mehlführer, m. wrulich, j. c. ikuno, d. bosanska, and m. rupp. simulating the long term evolution physical layer. in proc. of the 17th european signal processing conference (eusipco 2009), glasgow, scotland, august 2009. [12] a.k. sadek, z. han, and k. j. r. liu. distributed relay-assignment protocols for coverage expansion in cooperative wireless networks. mobile computing, ieee transactions on, 9(4):505–515, 2010. [13] l. song and j. shen. evolved cellular network planning and optimization for umts and lte. taylor & francis, 2010. [14] p. vieira and a. j. rodrigues. an insight into cooperative mimo communications in wireless networks. volume 1, pages 1–1, october 2010. a. martins et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-15 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc [15] c. wang, x. hong, x. ge, x. cheng, g. zhang, and j. thompson. cooperative mimo channel models: a survey. communications magazine, ieee, 48(2):80– 87, 2010. [16] y. yang, h. hu, j. xu, and g. mao. relay technologies for wimax and lte-advanced mobile systems. ieee communications magazine, 47(10):100– 105, 2009. a. martins et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-15 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc design methologies for integrated inductor-based soft-switching dc dc converters design methologies for integrated inductor-based soft-switching dc-dc converters vitor costa pedro m. santos beatriz v. borges inst. de telec., lisboa, portugal. área dep. de eng. electrónica e telec. e de comp., isel, lisboa, portugal. vsc@cedet.isel.ipl.pt inst. de telec., lisboa, portugal. academia militar, lisboa, portugal pedro.santos@lx.it.pt inst. de telec., lisboa, portugal. dep. de eng. electrónica e de comp., ist, lisboa, portugal. mbvb@lx.it.pt keywords: high switching frequency, quasi-resonant topologies, cmos integrated circuits, power management. abstract: this paper presents a study on resonant converter topologies targeted for cmos integration. design methodologies to optimize efficiency for the integration of quasi-resonant and quasi-square-wave converters are proposed. a power loss model is used to optimize the design parameters of the power stage, including the driver circuits, and also to conclude about cmos technology limitations. based on this discussion, and taking as reference a 0.35μm cmos process, two converters are designed to validate the proposal: a quasi resonant boost converter operating at 100mhz and a quasi-square-wave buck converter operating at 70mhz. simulation results confirm the feasibility of these topologies for monolithic integration. 1 introduction power management circuits for battery powered portable electronic equipment are being more demand, imposing the research on very efficient solutions. in some of these equipments the difference between the battery and the circuit voltage are significant. for this kind of applications the use of inductor-based switching topologies in the conception of integrated dc-dc converters in cmos technology have inherent advantages, in order to achieve good performances and compactness, at low cost. the main design specifications for an inductor-based cmos dc-dc converter are usually the ratio between the input and output voltages, the output power or load current, and the output voltage ripple. nevertheless, additional variables such as the switching frequency and the inductor current ripple have also to be considered, since they have a direct impact on silicon footprint, efficiency and system electromagnetic interference (emi). the fully integration of an inductor-based dc-dc converter in cmos technology brings new challenges. the inductor and capacitance integration are restricted to low values, when compared with the discrete implementation, due to the available silicon area. the low value integrated capacitors present reasonable behaviour and are normally used on mixed-signal. however, integrated inductors on standard cmos process are only available for a few nh and specifically target for rf-cmos applications. many authors have presented solutions with the objective of fully integrate switch-mode dc-dc buck converters [1-11]. distinct solutions have been proposed. some of the solutions are based on discontinuous conduction mode (dcm) operation, in order to reduce the filter inductor to an acceptable value for cmos, below 20 nh [7,9]. however, dcm operation implies overstress on power transistors for the same output load current. other solutions make use of the stacked chip concept to perform a system in a package (sip) [8], which are not cost effective [1, 9]. finally another approach is supported on the research of new design methodologies that leads to full integration, maintaining specifications competitive when compared with other hybrid and costly solutions [1-6]. it is well known that the physical dimensions and consequently the parasitic impedances of the filter passive components are greatly reduced as the converter switching frequency increases. in that case, the increase in the switching frequency could become the key parameter for full integration. for very high switching frequency the power transistors i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-17 http://journals.isel.pt/index.php/iajetc and respective driver power losses dominate the losses on the converter [2], mainly due to the increase of the switching losses, turning the design of the power stage even more relevant [1-6, 12]. although, the study of optimization methods on the design of the converters power stage [1-6, 12], all the solutions presented are based on hard-switching. so, the study of alternative solutions that could present lower switching losses maintaining high efficiency at high switching frequency is a major challenge. among the power electronics dc-dc converters circuit topologies, the soft-switching topologies [13, 14, 15] are distinguished by their efficiency and low electromagnetic interference (emi). thus, the use of soft switching techniques appears attractive to minimize noise and switching losses [16]. the objective of this paper is to present design methods based on the analytical loss model of the power stage described on [12] for soft-switching topologies. on section 2 the zvs (zero voltage switch) quasi resonant converters are introduced discussing the necessity of defining a design procedure tailored for cmos integration. a design procedure based on the theoretical operation for a boost converter is presented, which contemplates the power transistors and respective driver circuit optimization. section 3 introduces the zvs quasi square wave converters. a design method for the buck topology with power transistors and respective driver circuit optimization is also presented. several simulation results based on a standard 0.35 μm cmos process are shown in section 4 for the two types of resonant switching converters, in order to verify and validate the viability of the theoretical approach. conclusions are presented in section 5. 2 quasi-resonant converters quasi-resonant converters are obtained from conventional pwm (pulse width modulation) converters by using resonant switches. the difference between these switches and the conventional pwm switches is the inclusion of an inductor and a capacitor, as shown in figure 1, in order to allow zcs (zero current switch) (a), or zvs (zero voltage switch) (b). as a consequence, theoretically, switching frequency can be increased with reduced switching losses and increased power density for integration purposes. a) b) fig. 1. a) buck zcs qr converter; b) boost zvs qr converter. for monolithic integration in cmos technology, the use of intrinsic mosfet transistor capacitance as the resonant capacitor is a possibility. however, this is only valid for zvs topologies. in fact, mosfet parasitic capacitances distribution in zcs converters, do not match with the specific resonant capacitor c0, as can be confirmed in figure 1 a). furthermore, maximum current on the active switch is, at least, the double of the maximum current on a pwm equivalent converter [17]. these aspects reinforce the use of zvs topologies for cmos integration. for these reasons only the zvs topologies are considered in this work. 2.1 boost quasi-resonant converter theoretical operation a detailed steady state analysis of the zvs-qr buck topology is presented in [18]. based on that analysis, a design procedure for discrete zvs-qr buck converter is developed in [15]. this study can be extended for the boost topology only with some modifications. the resonant circuit, composed by l0 and c0, dominates the operation and mainly define the converter characteristics. for better comprehension consider the following parameters of the converter: characteristic impedance z0, the resonant frequency f0, the normalized load resistance q and the conversion ratio m: 0 0 0 c l z = (1.a) 00 0 2 1 cl f ⋅⋅⋅ = π (1.b) 0 z r q l= (1.c) i o v v m = (1.d) where rl is the load resistance. rl vo s2 ll0 c0 s1 c0 control fs cvs control fs control fs c0 c rl vo vs l l0 s2 s1 v. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-17 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc the conversion ratio of the zvs-qr boost converter can be obtained, as in the case of the zvs-qr buck converter in [18]: ( )mq f f m s ,,φ 2 1 0 α π ⋅ ⋅⋅ = (2.a) with: ( ) ( )ααα cos1 2 ,,φ −⋅+ ⋅ += q m m q mq (2.b) where:       += m q arcsinπα (2.c) to describe m as function of the control variable, the switching frequency fs, it is necessary to use a numerical procedure. however it is possible a closed form solution, with some rearrangement of equation (2.a): ( ) ( )mq m f f s ,,φ 12 0 α π −⋅⋅ = (3) the regulation characteristic obtained from equation (3) shows that the load variation implies adjustments on the control variable to regulate the output voltage. this means that not only is necessary adjustments on control variable when the converter ratio varies but also when the load value changes. these changes could compromise the critical condition that guaranties zero-voltage switching, qm ≥ . designing zvs-qr converters requires the design of the resonant circuit, the filter components and the power switches that includes the power mosfet and the respective drivers. the design of the filter components is similar to that of the pwm converters, considering the minimum switching frequency. 2.2 design procedure for boost quasi-resonant converter typically, the design specifications for a cmos dcdc converter include: output voltage, vo; input voltage range, mini v to maxi v ; the load resistance range, minl r to maxl r ; maximum switching frequency, maxs f . from the design specifications, the maximum and minimum value for the conversion ratio is given by: max min i o v v m = (4.a) min max i o v v m = (4.b) considering the input voltage and load range the worst condition that could compromise the zerovoltage switching occurs for the minimum conversion ratio, mmin, and maximum normalized load, qmax, defining a boundary condition described by equation (5). min 0 0 maxmax max m r z z r v v ll i o ≥⇔≥ (5) this situation will correspond to the maximum switching frequency, maxs f , given the minimum value for the characteristic impedance. for the definition of the input voltage and load range the maximum voltage value that the nmos mosfet, acting as active switch, supports, must be taken into account. this aspect is fundamental when designing dc-dc converters in cmos technology. the relation between the maximum value and the input voltage and load ranges will be given by:         ⋅+⋅= min max max min max1 l l odsnmos r r m m vv (6) equation (6) shows that even for fixed load and input voltage the maximum voltage on the active switch is twice the maximum voltage at the converter output. this will represent a problem if the output voltage is near the maximum cmos process voltage, or for high load variations, especially if the switching frequency rises to tens or even hundred of mhz (although there is the possibility of producing high-voltage cmos compatible transistors [19], these transistors operate at lower frequencies). the maximum current for the same transistor, in the worst case, can be obtained considering the maximum conversion ratio and the maximum load current. maxmaxmax max oldnmos imii ⋅== (7) for the pmos mosfet, the maximum voltage stress and the maximum current stress will occur for πα = . v. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-17 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc odspmos vv −≅max (8.a) maxmax max2 odpmos imi ⋅⋅= (8.b) considering the situation of maximum switching frequency, corresponding to maxii vv = and maxll rr = , and that z0 assume the minimum value given by equation (5), equation (3) will be rewritten as: ( )      −++⋅ ⋅ = αα π cos1 2 1 2 min 0 max m f f s (9) in the situation of boundary condition 2 3 max π αα ⋅ == equation (9) simplifies to: ( ) minmin0 1 13 4 max mmf f s ≅ +⋅⋅ ⋅ = π π (10) the minimum switching frequency will be achieved when minii vv = and minll rr = . in this way, equation (3) assumes a new form: ( )         −++ = αα π cos1 2 2 min max min max min max 0 min max max m n min mr mr mr mr m f f l l l l s (11.a) with:         ⋅ ⋅ += max min min max minarcsen mr mr l l πα (11.b) for cmos integration the maximum switching frequency will assume particular importance, due to cmos process limitations on frequency and converter efficiency optimization. in this way the maximum switching frequency must be chosen to meet the above requirements and will depend on the cmos process in use. resonant frequency, 0f , will be obtained after using equation (10). with the value of the resonant frequency and the characteristic impedance imposed by the boundary condition the resonant circuit components are given by the follow equations: 0 0 0 2 f z l ⋅⋅ = π (12.a) 00 0 2 1 zf c ⋅⋅⋅ = π (12.b) an extension of the design method used for hardswitching converters for the design of the power switches and respective driver circuits, presented on [12], can be used for zvs-qr converters. c g s p 2 p c gd p2p c g s n 2 p cgd n2p c g b n 2 p c g b p 2 p c d b n 2 p c d b p 2 p p 2p n2p c g s p 2 n cgd p2n c g s n 2 n cgd n2n c g b n 2 n c g b p 2 n c d b n 2 n c d b p 2 n p2n n2n c g s p 3 p c gd p3p c g s n 3 p cgd n3p c g b n 3 p c g b p 3 p c d b n 3 p c d b p 3 p p 3p n3p n1 c d b n 1 c g s n 1 c g b n 1 c g d n 1 c d b p 1 c g s p 1 c g b p 1 p 1 c g d p 1 r o n p 1 r o n n 1 c g s p 3 n cgd p3n c g s n 3 n cgd n3n c g b n 3 n c g b p 3 n c d b n 3 n c d b p 3 n p3n n3n switching frequency variation control vo vo vo vo vo vo vgn vgn vgn vgnvgp vgp vgp vgp lf cf l o a d pmos gate-driver nmos gate-driver wp3p= b(b+1)wp1 ap 2 wp3n= b(b+1)wn1 an 2 wn3n= (b+1)wn1 an 2 wn3p= (b+1)wp1 ap 2 wp2p= bwp1 ap wp2n= bwn1 an wn2n= wn1 an wn2p= wp1 ap vi l0 c0ext fig. 2. zvs-qr boost converter model including parasitic impedances and transistor sizes. v. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-17 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 2.3 power stage loss model the power loss model is obtained considering the proposed method for buck hard-switching dc-dc converters on [12]. this method could be extended, with a few modifications on the power mosfet and the respective drivers design model, for buck zvs-qr dc-dc converters. a boost zvs-qr converter model that includes the parasitic impedances is shown in figure 2. the losses of the driver circuit of the pmos power device, p1, shown on figure 2, are the same as for the buck hard-switching dc-dc converters, [12]. for the pmos power transistor the energy associated to the switching losses per unit will be slightly different and given by: ( ) ( )2 0001 gpopmosgdpmosgspmosgbp vvccce −⋅++= (13) considering the contributions from the driver circuit and the power losses from the pmos power transistor, the total power loss associated to p1 is obtained from: shingtotalswitcpprmspmos p pmos totalp fewi w r p ⋅⋅+⋅= 11 2 1 0 1 (14.a) with: spmosdriverphingtotalswitcp eee += 11 (14.b) where pmosr0 is the pmos on resistance per unit length, wp1 is the p1 width, fs the switching frequency and rmspmosi the pmos rms current. for the nmos transistor, n1, and the respective driver circuit, similar equations are obtained: shingtotalswitcnnrmsnmos n nmos totaln fewi w r p ⋅⋅+⋅= 11 2 1 0 1 (15.a) with: snmosdrivernhingtotalswitcn eee += 11 (15.b) where: ( ) 2 0001 gnnmosgdnmosgsnmosgbn vccce ⋅++= (15.c) from (14) and (15) it can be concluded that the conduction losses are proportional to the width of power mosfet. in this way it is possible to optimize the losses on the power transistors and respective drivers, as in [12] for hard-switch converters. for the definition of the tapering factor, tf, it has to be consider the losses in the driver circuits, as in [12] for hard-switch converters, remaining the functional behaviour of the converter. so, a different solution based on [3] was used, where tf is defined as five to ten times less than the charge of the resonant capacitor, c0, corresponding to the first operating interval of zvs-qr boost converter, the short time interval. 3 quasi-square-wave converters as in the case of the qr converters, the zvs-qsw and zcs-qsw converters are also obtained from conventional pwm (pulse width modulation) with some additional reactive components, with a different circuit topology. zvs-qsw converters present lower voltage and higher current stress on the switching devices. on the other hand zcs-qsw converters have lower current and higher voltage stress [20]. the zvsqsw topologies present the advantage of the addition of only one capacitor, when compared with the qr or zcs-qsw topologies, because theoretically the additional inductor can be placed in parallel with the filter inductor. thus, zvs-qsw topologies appear as an alternative to the use of high-voltage transistors in zvs-qr topologies. another advantage is that the parasitic capacitors associated to the two power transistors are in parallel and contribute both to the resonant capacitor. however, it is necessary to take into account some drawbacks: the higher variation of the inductor current, which leads to an increase of the electromagnetic interference (emi) when compared with the quasi-resonant topologies. to obtain a qsw converter it is necessary to manipulate the low-frequency storage elements in the correspondent pwm topology, followed by the insertion of resonant tank elements. certain pwm converters cannot be transformed into their corresponding zvs-qsw or zcs-qsw topologies, unless new low frequency storage elements are added to the original pwm converter (e.g. the zcs buck converter) [20]. this solution is only attractive for the topologies that do not need the additional low-frequency elements, as the zvs buck converter, presented in figure 3. v. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-17 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc v s v o rl c l f s control c 0 s 1 s2 fig. 3. buck zvs-qsw converter. in this converter the output filter inductor is also used as the resonant inductor. therefore, this solution appears more attractive then the correspondent qr topology, presented in figure 1. b), since the resonant inductor is not present. 3.1 buck quasi-square-wave converter theoretical operation a detailed steady state analysis of the buck zvs-qsw converter theoretical operation is presented in [20]. using this steady state analysis and performing an energy balance for one operation period, considering no losses, it is possible to obtain a set of five equations which resolution is not trivial. in order to solve this problem, a semi closed method to obtain the conversion ratio for the design of a buck zvs-qsw converter was developed in [21]. using the result obtained in [21] the frequency conversion ratio is given by: ( )      −⋅−⋅⋅ −⋅⋅⋅ = 12 )1(4 2 2 2 2 2 0 m i i q m q mm f f l m s π (16) assuming that m is known and that the parameter q is function of the load and of l0, the only unknown variable is im (maximum current in the inductor). the ratio between the maximum and average currents in the inductor, lmi iil /=α , results from the solution of (17): 0 23 =+⋅+⋅+⋅ dcba lll iii ααα (17) where: ( )mq a −⋅ −= 12 1 (17.a) ( )mq m q b −⋅ += 1 1 (17.b) ( ) ( )mm qm m mm m c −⋅⋅ ⋅−⋅ +−⋅ ⋅ − +      − = 12 12 12 1 11 arccos 2 (17.c) qmd ⋅= (17.d) from equation (17) three solutions are obtained for li α , but only one is in agreement with the normal behaviour of the converter. substituting in (16) the valid li α it is possible to obtain the conversion ratio, m, as function of the normalized switching frequency for different values of the normalized load, q. as in the zvs-qr boost converter the load variation implies adjustments on the control variable to regulate the converter output voltage. this variation forces a critical condition to guarantee the zero voltage switching in the zvs-qsw converter. in this case, independently of the normalized load, 5.0≥m . nevertheless, a design procedure is needed to design a zvs-qsw buck converter in cmos technology. the zero voltage switching critical condition implies the limitation of these converters to applications where the conversion ratio is above 0.5. designing zvs-qsw converters have the same requirements as in the zvs-qr converters. it is necessary to design the resonant circuit, the filter components and the power switches that include the power mosfet and the respective drivers. because of that a new design procedure is developed for cmos zvs-qsw converters in the next section. 3.2 design procedure for buck quasi-square-wave converter some authors have proposed design methods for zvs-qsw converters, [22]. however these methods are not appropriated for cmos integration. a new proposal was made in [21] but without the design procedure for the power stage. in this way a more detailed and improved method is proposed in this paper. the cmos monolithic integration of zvs-qsw converter implies a careful attention over the definition of circuit parameters. as a matter of fact the linear approximation made to obtain equation (16), is only valid if the normalized switching frequency is low [21]. this implies that q should have the smallest value as possible (from equation (16)). considering all the dependencies of q, including the switching frequency, the load and v. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-17 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc characteristic impedance z0, and taking into account the parameters of the cmos process, the following design method is proposed: 1. define the normalized load, q, that with the conversion relation, m, of the converter to be designed, corresponds to a small ratio between the switching frequency and the resonant frequency. 2. with q defined, obtain the characteristic impedance, z0, considering the maximum load resistance, maxl r , corresponding to the boundary condition. obtain the ratio between im and il using equation (17). determine the ratio between the maximum switching frequency, maxs f , and the resonant frequency using equation (16). 3. design the power transistors and associated drivers using an extension of the method used in [12], function of the load current, io, and the maximum switching frequency, maxs f , which is defined considering the compromise between the occupied area and the converter efficiency, taking into account the limitations inherent to the specific cmos process. obtain the resonant frequency with the information of step 2. 4. obtain the intrinsic parasitic capacitors of the power transistors that contribute to the resonant capacitor, c0. the sum of these capacitances must be smaller than the resonant capacitor determined after steps 2 and 3. if not, return to step 1 and relax the specification of q. with this method it is possible to guarantee a normal behaviour of the converter in the worst case, defining as in the case of the zvs-qr converters, ranges of variation for the input voltage and output load. after the definition of the maximum normalized load, qmax, the characteristic impedance could be obtained by: max 0 max q r z l = (18) using equations (16) and (17) and considering the situation of mmin, it is possible to obtain the values corresponding to the situation of maximum switching frequency, the worst condition to validate the restriction on the ratio between the switching and resonant frequencies. using the same equations and making maxmm = and 0/min zrq l= it is possible to obtain the relation between the minimum switching and resonant frequencies, defining in this way the operating interval of the converter, as function of the ranges of the input voltage and of the load. in the zvs-qsw buck converter the filter inductor is shunted by the resonant inductor. so, the design of the filter components is reduced to the design of the filter capacitor. the filter capacitor value is obtained in a similar way as for the pwm converters considering the minimal switching frequency operation and assuming a ripple current on the inductor given by half of im: 2 m l i i ≅∆ (19) the efficiency analysis of the converter is measured considering the situation of maximum switching frequency (with qmax and mmin), corresponding to the worst case for dynamical losses on the power stage. 3.3 power stage loss model the power losses model of the power transistors and respective driver circuits is similar to the power loss model obtained for the zvs-qr boost converter. the losses on the zvs-qsw buck converter are characterized in the same way as for the zvs-qr boost converter. the parasitic capacitances considered are the same because of the zero voltage switching. in this way, the equations from (13) to (15) are valid and can be applied. for the definition of the tapering factor, tf, the solution adopted is the same as in de zvs-qr boost converter. 4 design procedures validation this section presents simulation results of a zvs-qr boost converter and a zvs-qsw buck converter that validates the design procedures proposed in this work. the zvs-qr boost converter was design for the following specifications: [ ]v32.108.1 l=iv , vo = 3.3v, [ ]ma109l=oi , ll ii ⋅=∆ 1.0 and ripple output voltage below 1%. taken as reference a maximum theoretical efficiency of 80% the maximum switching frequency for the cmos v. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-17 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc process is 100mhz. with that value and using the design methodology exposed in section 2, the parameters obtained for the converter are: l0 = 90nh, c0 = 4.18pf, lf = 2.9µ h, cf = 1.9nf. the optimized transistors dimensions are wnmos = 339.6µ m, wpmos = 552.9µ m with lnmos = lpmos = 1µ m. the resonant capacitor will be formed by a nmos parasitic capacitor and an additional shunt capacitor. fig. 4. transient analysis detailed waveforms of output voltage (vout), inductors currents (il and il0), power transistors control signals (vg_nmos and vg_pmos) and resonant capacitor voltage (vc0) of the boost qr converter. output voltage full transient response waveform. the waveforms presented in figure 4 shows some drawbacks on the use of zvs-qr converters for integration purpose. the current in the resonant inductor and the voltage in the resonant capacitor oscillate in the last interval of the switching period. the parasitic resonance results from the accumulated energy in the pmos parasitic capacitance, cgd, when acting as the switching transistor. the accumulated energy is injected in a resonant circuit formed by the resonant inductor, l0, and the capacitor resulting from the sum of cgd with cdb in the pmos power transistor. this resonance will be described by the following equations: ( ) ( )( ) ll itt z v ti p +−⋅⋅= 30 0 sen 0 ωσ (20.a) ( )( ) oc vttvv p −−⋅⋅= 30cos ωσ (20.b) where σ v represents the cgd accumulated energy resultant voltage, pc v represents the voltage in the pmos total parasitic capacitance, the sum of cgd with cdb, and p z 0 represents the characteristic impedance of the parasitic resonance circuit. this parasitic resonance causes a resonant voltage on the pmos drain that presents a high voltage stress, not supported by the cmos process. this imposes the use of a high voltage pmos transistor. the other solution could be the use of a diode in the pmos place, like in the discrete implementation, but it will lead to less efficiency. this will be a significant drawback in the use of zvs-qr topologies for cmos integration. using the section 3 design procedure, a zvs-qsw buck converter was designed with the following characteristics: [ ]v2.58.4 l=iv , v3.3=ov , ma30=oi , oo vv ⋅=∆ 05.0 , l0 = 249nh, c0 = 1.3pf, cf = 990pf, mhz70 max =sf . the optimized transistors dimensions were wnmos = 238.1µ m, wpmos = 653.9µ m with lnmos = lpmos = 1µ m. the waveforms presented in figure 5 are in good agreement with the expected results. after the transient, the converter tends to stabilize at the desired values for output voltage and average inductor current. fig. 5. transient analysis detailed waveforms of inductor current, l0, output voltage, vo and pmos and nmos control signals. output voltage full transient response waveform. the theoretically efficiency, considering only the losses of the power transistors, is approximately 80%. the value obtained for the efficiency in the simulated circuit was about 7% less. the difference obtained is mostly caused by the overlap conduction on the inverters of the power transistors driver circuits. 5 conclusions two design procedures for the implementation of resonant switching converters in cmos technology are introduced. a zvs-qr boost converter and a zvs-qsw buck converter were conceived using the proposed design methods. the results obtained show a validation of the proposed design methodologies. some drawbacks for the implementation of the zvsqr converters in cmos are revealed for the initial assumptions considered. the use of high-voltage pmos transistors is proposed to solve these v. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-17 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc drawbacks. nevertheless the parasitic resonance could cause a higher emi. the results obtained from the zvs-qsw buck converter encourage this solution. the introduction of non-overlap circuits in the last inverters of the inverters chain of the power transistors driver circuits could reduce the switching losses. the parasitic resonance in the cmos integrated zvs-qr converter reduce the advantage of lower emi when compared with the cmos integrated zvs-qsw converter. in this context the utilization of cmos integrated zvs-qr converters will be conditioning to the study of solutions to reduce the parasitic resonance. references [1] volkan kursun, silva g. narendra, vivek k. de and eby g. friedman, “analysis of buck converters for on-chip integration with dual supply voltage microprocessor”, ieee transactions on very large scale integration (vlsi) systems, vol. 11, n. 3, jun 2003. [2] volkan kursun, silva g. narendra, vivek k. de and eby g. friedman, “low-voltage-swing monolithic dc-dc conversion”, ieee transactions on circuits and systems-ii: express briefs, vol. 51, n. 5, may 2004. [3] a. j. stratakos, s. r. sanders, r. w. brodersen, “a low-voltage cmos dc-dc converter for a portable battery-operated system”, ieee power electronics specialists conf., june 1994, vol. 1, p.p. 619-626. [4] g. villar, e. alarcón, j. madrenas, f. guinjoan and a. poveda, “energy optimization of tapered buffers for cmos on-chip switching power converters”, ieee int. symp. on circuits and systems, vol. 5, may 2005, p.p 4453-4456. [5] s. musunuri and p. l. chapman, “improvement of light-load efficiency using width-switching scheme for cmos transistors”, ieee power electronics letters, vol. 3, n. 3, sep. 2005, p.p 105110. [6] t. takayama and d. maksimovic, “a power stage optimization method for monolithic dc-dc converters”, ieee power electronics specialists conf., june 2006, p.p 1-7. [7] m. wens, m.s.j. steyaert, "a fully integrated cmos 800-mw four-phase semiconstant on/offtime step-down converter," power electronics, ieee transactions on, vol.26, no.2, pp.326-333, feb. 2011. [8] k. onizuka, k. inagaki, h. kawaguchi, m. takamiya, t. sakurai, "stacked-chip implementation of on-chip buck converter for distributed power supply system in sips,"solidstate circuits, ieee journal of , vol.42, no.11, pp.2404-2410, nov. 2007. [9] m. wens, m. steyaert, "a fully-integrated 0.18μm cmos dc-dc step-down converter, using a bondwire spiral inductor," custom integrated circuits conference, 2008. cicc 2008. ieee, vol., no., pp.17-20, 21-24 sept. 2008. [10] j. ni, z. hong, and b. liu, “improved on-chip components for integrated dc-dc converters in 0.13 μm cmos”, in proc. 35th eur. solid-state circuits conference, pp. 448–451, sep. 2009. [11] j. wibben and r. harjani, “a high-efficiency dcdc converter using 2nh integrated inductors”, ieee j. solid-state circuits, vol. 43, no. 4, pp. 844–854, apr. 2008. [12] v. costa, p.m. santos, b. borges, p. m. santos, “a design methodology for integrated inductor-based dc–dc converters”, elsevier microelectronics journal, vol. 43, nº 6, june 2012, p.p 401-409. [13] vatché vorpérian, richard tymerski and fred c. y. lee, “equivalent circuit models for resonant pwm switches”, ieee transactions on power electronics, vol. 4, nº2, april 1989. [14] dores costa j.m., “design of linear quadratic regulators for quasi-resonant dc-dc converters”, ieee power electronics specialists conf., june 2001, vol. 1, p.p. 422-426. [15] wojciech a. tabisz, pawel m. gradzki and fred c. y. lee, “zero-voltage-switched quasi-resonant buck and flyback converters experimental results at 10 mhz”, ieee transactions on power electronics, vol. 4, nº2, april 1989, p.p 194-204. [16] olivier trescases and wai tung ng, “variable output, soft-switching dc/dc converter for vlsi dynamic voltage scaling power supply applications”, ieee power electronics specialists conf., vol. 39, nº1, june 2004, p.p. 4149-4155. [17] k. h. liu, r. oruganti and f. c. lee, “quasiresonant converters – topologies and characteristics”, ieee transactions on power electronics, vol. 2 jan. 1987, p.p 62-71. [18] k. h. liu and f. c. lee, “zero-voltage switching technique in dc/dc converters”, ieee power electronics specialists conf., june 1986, p.p. 58-70. [19] p. m. santos, v. costa, m. c. gomes, b. borges, m. lança, “high-voltage ldmos transistors fully compatible with a deep-submicron 0.35µ m cmos process”, elsevier microelectronics journal, vol. 38, nº 1, jan. 2007, p.p 35-40. [20] vatché vorpérian, “quasi-square-wave converters: topologies and analysis”, ieee transactions on power electronics, vol. 3, nº2, april 1988. [21] v. costa, p. m. santos, b. v. borges, “design method for integrated cmos quasi-square-wave dc-dc converters”, proceedings of dcis2005, xx conference on design of circuits and integrated systems, lisboa, nov. 2005. [22] d. maksimovic, “design of the zero-voltageswitching quasi-square-wave resonant switch”, ieee power electronics specialists conf., 1993 rec., p.p 323-329. v. costa et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-17 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc stochastic theater: stochastic datapath generation framework for fault-tolerant iot sensors 1 stochastic theater: stochastic datapath generation framework for fault-tolerant iot sensors rui policarpo duartea, mário véstiasb, carlos carvalhoc, joão casaleiroc ainesc-id, instituto superior técnico, universidade de lisboa, portugal binesc-id, instituto superior de engenharia de lisboa, portugal ccedet, instituto superior de engenharia de lisboa, universidade nova, portugal rpd@inesc-id.pt mvestias@deetc.isel.pt cfc@cedet.isel.ipl.pt joao.casaleiro@cedet.isel.ipl.pt abstract— stochastic computing has emerged as a competitive computing paradigm that produces fast and simple implementations of arithmetic operations, while offering high levels of parallelism, and graceful degradation of the results when in the presence of errors. iot devices are often operate under limited power and area constraints and subjected to harsh environments, for which, traditional computing paradigms struggle to provide high availability and fault-tolerance. stochastic computing is based on the computation of pseudo-random sequences of bits, hence requiring only a single bit per signal, rather than a data-bus. notwithstanding, we haven’t witnessed its inclusion in custom computing systems. in this direction, this work presents stochastic theater, a framework to specify, simulate, and test stochastic datapaths to perform computations using stochastic bitstreams targeting iot systems. in virtue of the granularity of the bitstreams, the bit-level specification of circuits, high-performance characteristics and reconfigurable capabilities, fpgas were adopted to implement and test such systems. the proposed framework creates stochastic machines from a set of user defined arithmetic expressions, and then tests them with the corresponding input values and specific fault injection patterns. besides the support to create autonomous stochastic computing systems, the presented framework also provides generation of stochastic units, being able to produce estimates on performance, resources and power. a demonstration is presented targeting klt, typical method for data compression in iot applications. keywords: iot, fpga, fault-tolerant computing, stochastic bitstreams, approximate computing. i. introduction data intensive digital signal processing (dsp) applications for near real-time image and video processing, neuromorphic [21] and bio-inspired systems [17], are characterized for their regularity in their datapath. their computations are mainly based on multiplications followed by accumulations, and by the fact that they can tolerate some errors in their computations. to alleviate edge servers from the work of computing basic, but essential, dsp and machine learning (ml) functions, there is interest in delegating such computing to the internet of things (iot) device. however, in the iot context, devices are often required to operate under heavy power and area constraints and subjected to harsh environments, struggle to provide high availability and fault-tolerance. to overcome such limitations, this work proposes to make use of a different computing paradigm that blends well with the iot context, and offers direct analog sensor interface without analog-to-digital fig. 1. illustration of two scenarios for a typical iot application, with stochastic computing. converters (adcs), fault-tolerance and savings in resources and power. stochastic arithmetic has emerged as an alternative computational paradigm able to provide approximate computations requiring less hardware, towards a circuit design with simpler but massively parallel components, trading off precision for computation time [11]. applications like neural networks [24], [23], highthroughput bayesian inference [15], image and video processing [19], finite impulse response (fir) [5] and infinite impulse response (iir) [20] digital filters, and autonomous cyber-physical systems [10] are characterized for their regularity in their datapath. their computations are mainly based on multiple multiplications followed by accumulations. moreover, many of these applications do not require exact results and can tolerate some deviations in their computations. the operation of stochastic computing (sc) is suitable for reconfigurable devices such as field-programmable gate arrays (fpgas) given that the bit level specification of stochastic bitstreams makes it favorable for implementation on these devices. figure 1 illustrates the scenario where a dependable iot system requires the implementation of fault-tolerance mechanisms at all levels of the system, even though it only acquires data from the sensor and communicates the data to the edge servers. the proposed approach, using sc, performs computations directly over the acquired stochastic bitstream, thus alleviating the computational load at the edge, and reducing the require fault-tolerance mechanisms at the iot and edge levels. however, the majority of research conducted on sc is confined to a set of applications, which are highly customized, specific to certain applications and difficult to extend its adoption. furthermore, the benefits of sc are not always clear due to the resources of the supporting elements and the clock latency to process long bitstreams. often, the benefit of sc is i-etc: isel academic journal of electronics, telecommunications and computers iot-2018 issue, vol. 4, n. 1 (2018) id-8 http://journals.isel.pt shadowed by the latency and resources required to interface a traditional computing systems. as an inspirational example, the bayesian inference system presented in [7] requires 597 logic elements (les) to be implemented, of which, only 42 are spent on the datapath for the bayesian machine. the remaining 555 les are spent on conversion of 13 stochastic bitstreams. the contribution in [18] presents a comparison of parallel binary versus stochastic implementations for neural networks on reconfigurable hardware. the authors concluded that even though stochastic bitstreams require more clock cycles to compute than binary, the advantage of compact realizations in hardware surpasses that through comparison of geometric mean of the two metrics. therefore there is a need for a methodology to make this assessment at an earlier stage of the design process. the main claim addressed in this paper is to ease the definition and evaluation of a stochastic datapath (sd) to compute, at the iot-level, mathematical expressions as alternative to other time consuming and prone-to-error design approaches, and without having to delve into the technicalities of highlevel synthesis (hls). this work presents stochastic theater, a highly customizable and scalable framework that given a problem’s specification as mathematical expression, it generates the corresponding sd, and its supporting blocks, targeting reconfigurable logic. this work is intended to facilitate automated architectural changes via unified and regular interfaces, and design-space exploration often sought in research due to the long execution times. moreover, this work provides an estimate of resources, power and performance metrics. this enables the usage of the sc in stand-alone stochastic systems or accelerators for heterogeneous and system-on-a-chip (soc) platforms. stochastic theater is a novel framework for prototyping of sc systems on fpgas, and it is an improvement over previous work found in [8]. moreover, the karhunen-loève transform (klt) algorithm can be implemented as inner product, which is the same for a fir filter, hence demonstrating its wide range of applicability of the proposed work. stochastic theater offers the following features: • scalable and fully automated process through the execution of configuration scripts, to enhance the use of stochastic computing; • generation of custom stochastic computing elements, e.g. arithmetic units with more than 2 inputs; • support for fault injection, through simulation, to predict the behaviour of the stochastic system when operating under fault conditions; • supports for self-timed ring-oscillator (stro) to minimize temporal correlations of bitstreams, according to [9]; • support for different fpga device families. currently cyclone iii, iv and v from altera are supported, but can be easily ported to devices from other vendors. the flow of stochastic theater is illustrated in figure 2. it begins with a high-level specification from the user as a mathematical expressions described in python, which are translated into a computational stack, as sequences of inter-connected operands and operators. the framework then generates the register transfer level (rtl) in very high speed integrated fig. 2. flow of the proposed framework to simulate, synthesize and evaluate stochastic computing systems on fpgas. circuits (vhsic) hardware description language (vhdl) for custom sc arithmetic units, the sd which implements the desired functionality, and the supporting blocks. this paper is organized as follows: section ii is devoted to introduce sc and presents the most relevant research contributions incorporated in the proposed framework. section iii presents the details about the proposed framework along with the inner workings of the proposed framework, to generate vhdl entities and the datapath for the mathematical expression to be implemented. a demonstration of a stochastic system with the first implementation of the klt implemented on an fpga is in section iv. analysis on the outcomes are in section v conclusions and final remarks are in section vi. ii. background j. von neumann introduced sc in [22] as a method to design probabilistic logic circuits and synthesize robust systems from unreliable components. in [12], gaines has introduced the use of stochastic bitstreams to represent operators with high levels of error tolerance. a. stochastic bitstreams by definition, a stochastic signal is the result of a continuous-time stochastic process which produces two values: 0 and 1. according to [11], a unipolar stochastic bitstream is a sequence of stochastic signals over time whose value is within [0; 1] and defined as the number of ones (o) over the total number of bits (t). in bipolar representation, the value is within [−1; 1] and is also encoded as a ratio but followed by a negative bias and a scale factor of 2. on stochastic bitstreams there are no weights in the representation, as in typical binaryradix representation, thus all bits have the same contribution for the encoded value. for example, the same sequence of 8 bits 01110110 represents 5/8 = 0.625 in unipolar and 2 ∗ (5/8 − 0.5) = 0.25 in bipolar. figure 3 illustrates the aforementioned stochastic bitstream. on top, there is the clock signal, to ensure synchronism; and on bottom the encoded value. rui policarpo et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt fig. 3. example of a stochastic bitstream encoding 0.625 and 0.25 in unipolar and bipolar encodings, respectively. fig. 4. block diagram of the unipolar stochastic units: a) multiplier (top-left), b) adder (top-right), c) negation (bottom-left) and d) squarer (bottom-right). b. stochastic arithmetic stochastic arithmetic supports basic arithmetic computations like addition and multiplication, as illustrated in figure 4. details on stochastic arithmetic units can be found in the survey presented in [2] which covers the most common arithmetic units. the unipolar stochastic multiplication is only the result of a logic and of its stochastic inputs. the complement is the negation of the bitstream. bipolar multiplication is achieved through an xnor operation. addition, or more precisely average, is obtained via a round-robin multiplexation of the stochastic inputs, which depends on a n-module counter corresponding to n inputs in the multiplexer. the square of a stochastic bitstream is the equivalent to a multiplication of a bitstream by itself, delayed by a clock cycle. the clock cycle makes the pseudo-random bitstreams to be uncorrelated. the multiplication is now of two independent streams but with the same value, resembling the power of two operation. the implementation of n-ary add or multiply operators is achieved by adding additional inputs to the logic circuits. in terms of fpga implementation it means that the number of inputs of a le can be evaluated simultaneously. for more complex operators, such as exp, tanh and abs, there are realizations of stochastic operators using finite state machines (fsms). implementations of such units can be found in [4], [14]. table i illustrates the sensitivity to temporal correlations, for the case of multiplications of two bitstreams. in this case both input streams encode the value 0.5, which is represented by the same number of 0s and 1s on the bitstream. however, because of the alignment of the bits, the multiplication, which is achieve at the logic and, produces a bitstream with only 0s, encoding value 0, rather than the expected value of 0,25. to improve the statistical quality of the stochastic bitstreams on the datapath, this work adopts the stro proposed in [7]. bit num 8 7 6 5 4 3 2 1 a 0 0 1 1 0 0 1 1 b 1 1 0 0 1 1 0 0 a.b 0 0 0 0 0 0 0 0 table i example of a multiplication of two correlated bitstreams, producing a bad result. fig. 5. detail on process of generating a pseudo-random bitstream for a given binary-radix value between 0 and 1. the main reasons to consider a stro instead of a global clock source are: different clock signal for each stochastic unit; all generated clock signals with the variation of voltage, temperature, location on the device and its degradation. all synchronous stochastic units have an instance of this unit to generate its clock signal. c. interface and other supporting blocks forasmuch as most systems usually use parallel binaryradix representation, it is therefore required a converter from-to stochastic bitstreams to ensure inter-operability. the process of generating the stochastic bitstream is illustrated in figure 5, where a specific binary-radix value (val) is compared with the output of a uniform pseudo-random generator, usually a linear feedback shift-register (lfsr) [3]. whenever the pseudo-random number is smaller, it produces a 1 and 0 otherwise. after each pseudo-random sample the ratio between the number of ones and the total number of bits will be towards val. in this example, the encoded value is 9/16 = 0, 5625. the conversion from stochastic-to-binary is based on the integration of the 1s on a bitstream, which is accomplished using a binary-radix counter. a second counter is required to count the total number of bits. figures 6 and 7 show the details of the conversion units. d. fault-tolerance the graceful degradation of stochastic bitstreams is referred to the impact of bit-flips on the bitstream. in such occurrences, fig. 6. block diagram of a binary-to-stochastic unit. rui policarpo et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt fig. 7. block diagram of a stochastic-to-binary unit. and regardless of the position of the bit on the bitstream, the value of the error associated with each bit-flip is the same as the least significant bit, in binary-radix. on this account, [19] has applied the concept of stochastic logic to a reconfigurable architecture that implements image processing operations on a simulated datapath. the authors show that the quality of the results degrades gracefully with the increase of errors on the bitstream. e. disadvantages the main disadvantages of sc are: a linear increase in the precision of typical binary representations, for stochastic computations it imposes an exponential increase in the length of the bitstream; sensitivity to temporal correlations; and the supporting blocks are usually the performance bottleneck, rather than the arithmetic units. iii. stochastic theater: framework for specification of stochastic datapaths this work proposes a method to specify it as a mathematical expression, defined as a list of operands and operators, organized in a stack to resemble reverse polish notation (rpn), or postfix notation. the advantages of such representation are: simplified representation without parenthesis, hence fewer operations are needed, faster introduction by the user and with fewer mistakes [13], [1]. in rpn the operators follow the operands. the strength of this notation is the support of nary operators, which is compatible with the aforementioned stochastic operators. example: the computation of 1 − 2× 3 is defined in rpn as 1 2 3×−. essentially, the framework recognizes the different operands and operators of a mathematical expression, and then generates the corresponding vhdl. this regular form is easily mapped into an fpga, exploiting the parallelism offered. from this data structure it is possible to identify the requirements for a system, namely: the number of input, internal and output signals; the different types, number of input operands and data dependencies of the operators used. the data structure is organized as a tree of computations which maintains the data dependencies in the datapath. these mathematical expressions can be variable in size and type of operations. considering the following example of a function to be implemented to compute data from 4 sensors: func = 1 n (i0 × i1 + i2 × i3 × i4) (1) it has the corresponding rpn stack representation: func = i0 i1 × i2 i3 i4 × × + n / (2) to ease the stack manipulation, it is split into a set of partial computations stored in intermediate variables: aux0 = i0 × i1 (3) aux1 = i2 × i3 × i4 (4) which the original expression can be replaced with: func = 1 n (aux0 + aux1) (5) and to facilitate the generation of the vhdl source file describing the datapath to implement this expression, it is expressed as a list of operands and operators in python, e.g. sums and multiplications, resembling rpn. this regular form is easily extracted and can be efficiently mapped into an rtl specification, exploiting the parallelism offered by fpgas. the user input for equation 2 in python can be described by the following list of computations, which itself can be comprised of other lists, or operands, and operators: aux1 = [’i0’, ’i1’, ’*’]; aux2 = [’i2’, ’i3’, ’i4’, ’*’]; f = [aux1, aux2, ’+’]; which results in the following python variable: >>> f [[’i0’, ’i1’, ’*’], [’i2’, ’i3’, ’i4’, ’*’], ’+’] variables t1 and t2 are lists of strings, which represent partial computations. these variables can be of any size. the last element, or tail, of the list holds the representation of the operation. in this example, the operands are: + or *. the remaining elements are the operands. it is also possible to define operations which depend on the results of previous computations, e.g. f is defined as the sum, or average, of t1, and t2 . table ii lists the stochastic combinatorial and sequential operators supported so far. table ii list of the stochastic operators supported. operator codification sum (average) + multiplication * negation square pow2 complement not the inputs and outputs of the sd correspond to the number of variables and are determined by the framework. to complete the specification of a datapath it is necessary to indicate the length and type (unipolar/bipolar) of the bitstream. to serve this purpose there is a variable in python which holds this configuration. however, the architecture of the sd is independent of the bitstream’s length. one of the key strengths of the proposed framework is that given any mathematical expression, regardless the complexity of the mathematical expressions, the system maintains its regularity. the framework integrates the translation of the rui policarpo et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt fig. 8. top-level architecture of the circuit design to test the stochastic datapaths, including the supporting units. expression of the sd into a stack. apart from the core of the sd, which is different for all expressions, all other supporting units have the same architecture, such as data sources and sinks for the stochastic bitstreams, varying only the number of bits, or the length of the bitstreams supported. the generated sd was planned to be autonomous or part of a larger system, as illustrated in figure 8. the sd is in the middle and the rest of the circuit is formed by the supporting units to do the computations. the system is interfaced via the input and output bitstreams, and also the fsm’s control signals, namely clk, enable and reset. in particular, the fsm is responsible for the generation of the control signals for all units in the design. it also controls the burn-in period to compensate the clock cycles required by the fsm-based stochastic arithmetic units. a. stochastic arithmetic units typical parallel binary-radix representation all basic operators are either unary or binary, with 1 or 2 operands, respectively. however, has n-ary sc operators support for more than two operands. therefore, it is required to create the customized components, as it is difficult to account for all possible operators with any number of operands in advance. therefore, the framework determines the number of arguments for multiplications and sums and then generates the required stochastic arithmetic components. in more detail, it iterates over the aforementioned list of computations to retrieve the different operators and then generates the vhdl entity matching the operation and the number of inputs. in sc, each operator can have a diverse number of operands. therefore, it is necessary to generate custom arithmetic units according to the mathematical expression. moreover, the number of variables considered is unknown, so it is also necessary to create the interfaces to support any number of inputs and outputs. the vhdl source files are created, and used to synthesize the design and generate the fpga configuration file, or to simulate the system. b. interfaces for iot conversion between binary-radix and stochastic bitstreams is the major limitation in interfacing typical digital systems. even tough it offers many parallel operators there are not many inputs available. connecting the sd from the rest of the supporting elements allows to integrate it in other systems, capable of interfacing fig. 9. analog-to-stochastic bitstream conversion circuit (from [10]). fig. 10. example of a 3-input stochastic adder with a stro to minimize correlation between the bitstreams. with stochastic bitstreams, such as [10]. in this work the authors have created a cyber-physical system which interfaces analog sensors and actuators without the need to have either analog to digital and binary-to-stochastic converters, to acquire input data; and stochastic-to-binary and digital to analog converters to drive the actuators. in essence, the generation of a bitstream from a binary-radix value requires more resources than an analog interface, but the analog interface requires a dedicated input pin. c. state-of-the-art attributes to hold on to the novel advancements in sc, the framework already includes a few research novelties to demonstrate its adaptability. the incorporated features which mitigate some of the limitations in sc and improve the quality of the results 1) burn-in period: there are stochastic arithmetic units are based on fsm, thus the values at their outputs are not instantly produced. therefore, to account for such units, a counter is included to introduce enough latency. the outputs become valid once the counter reaches the threshold value. 2) independent and uncorrelated units: as mentioned previously, correlation between bitstreams leads to weak results. to reduce such correlations [9] introduced stros to generate spread-spectrum, individual and uncorrelated clock sources for each synchronous stochastic unit. moreover, the authors also claim reduction in the power dissipated in the clock trees, without the penalty of introducing synchronizers, or alternative components, typical of asynchronous circuit designs [16]. this feature, which consists of a configurable length ring oscillator can be instantiated to provide the clock signal to all synchronous components in a sd, as illustrated in fig. 10. d. simulation platform to facilitate the system development and verification, the proposed framework supports both rtl and gate-level simulation. this functionality is granted by a vhdl top-level rui policarpo et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt entity, automatically and specifically created for each system. this top-level entity automatically interfaces the generated sd, by connecting its inputs and outputs. regarding the input stimulus for the simulation, it uses the values from the problem specification. the same information is then used to validate the obtained results at the end of the simulation. one of the flagships of stochastic computing is its inherent resilience to faults. doing fault-tolerance tests requires a rather complicated laboratory environment or native support from the fpga to change configuration bits according to existent models. therefore, to alleviate the designer from such, the proposed framework supports simulation of the complete system using fault-injection models. the actual simulation process itself is supported directly by modelsim, through the execution of a custom script which compiles all source files, executes the simulation and displays the waveforms for all signals. figure 11 depicts the waveforms for some of the signals from a simulation of a stochastic system. it is worth noticing the buses which aggregate and organize the bitstreams in the design. e. fault injection fault injection is performed at rtl level on the sd or the complete system, and they can be configured to upset the system as transient or permanent faults. in either case the faults injected are stuck-at 0 or 1 faults. the fault injection is supported only at the simulation level through the use of modelsim scripts. each fault is characterized by an identifier of a net from the circuit, simulation time of occurrence and logic level of the fault. all faults are generated by a python script before running the simulation following a specific probability distribution, e.g. weibull or normal distributions. >>> a = 5. # shape >>> s = np.random.weibull(a, 1000) f. evaluation platform the proposed framework provides a test platform to run any sd generated by it on an fpga. it creates a fully functional autonomous stochastic system, containing the sd derived from the mathematical expression. the system supports sd of any size, being limited by the resources available on the fpga device. this test platform manages the input and output signals required by the sd, along with the required conversions to be accessed by the host computer. figure 12 depicts the system to be implemented on the fpga. on the edges there are the conversion blocks, and in the middle the unit corresponding to the sd. 1) architecture: the test platform circuit is constituted by the circuit under test (i.e. a simple arithmetic unit or a sd), the bitstream generators, and the output calculators. it includes the units for the generation of the stochastic bitstreams from binary values previously stored in block random access memorys (brams), and the result converters back to binary and its storage in other brams. in more detail, each of these units supports many parallel bitstreams. 2) process: the process of evaluating a sd starts with the configuration of the fpga with the bitstream. thereafter it is ready to exchange data with the host computer. the whole process is controlled by the host computer via tool command language (tcl) scripts, and illustrated in fig. 13. the fsm controls the test process. it waits for the indication from the host computer to start the generation of the input bitstreams and starts counting the burn-in period, from binary values stored in brams. after the burn-in period is over, the fsm starts the conversion of the output bitstreams. iv. karhunen-loève transform a. background the klt, also known as principal component analysis (pca), is an algorithm widely used in machine learning to reduce the dimensionality of data sets of many correlated variables, and is formulated as follows. given a set of n data xi ∈ rp , where i ∈ [1,n] an orthogonal basis described by a matrix λ with dimensions p×k can be estimated that projects these data to a lower dimensional space of k dimensions. the projected data points are related to the original data through the formula in (6), written in matrix notation, where x = [x1,x2, ...,xn ] and f = [f1,f2, ...,fn ], where fi ∈ rk denote the factor coefficients. f = λt x. (6) the original data is described from the lower dimensional space via (7): x = λf + d (7) where d is the error of the approximation. the objective of the transform is to find a matrix λ that has the mean-square error (mse) of the data approximation minimized. a standard technique is to evaluate the matrix λ iteratively as described in steps (8) and (9), where λj denotes the jth column of the λ matrix. λj = arg max e{(λtj xj−1) 2} (8) xj = x − j−1∑ k=1 λkλ t k x (9) where x = [x1x2...xn ], x0 = x, ‖λj‖ = 1 and e{.} refers to expectation. fig. 14 presents a sample of the input data set. the data set consists of 68 images of 20x25 pixels. to compress the images from their dimension of 500 pixels down to 100 pixels, the λ matrix is required to have 500x100 elements. this application resembles the image compression in surveillance systems. b. implementation and results the klt algorithm is based on the dot-product operation, which can be implemented using different circuits. figure 15 shows the datapath for: a) rolled and b) unrolled architectures of a dot-product based circuit, to implement the datapath of one projection vector from a zp to zk klt. the circuit receives data from the input stream, identified with x. the samples, from the input stream, for each dimension p, are multiplied by the corresponding projection vector fig. 11. waveforms of a stochastic system modelsim simulation. fig. 12. rtl of a test circuit for a stochastic datapath. fig. 13. actions performed in the test process. fig. 14. faces used as input data for the klt designs. a) b) fig. 15. schematics of the datapath of a dot-product, for a projection vector of a klt circuit: a) rolled and b) unrolled architectures. λpk. the output of the multiplier is connected to an adder to do the accumulation. the final result is placed in the output stream, identified with fk. in this experiment, to compare binary-radix against sc, it was considered the unfolded architecture, which is the one that maximizes the parallelism offered by fpgas. considering also 9-bit binary-radix representation, it corresponds to a 512bit bitstream. however, the length of the bitstream has no influence on the sd. the implementation of the aforementioned klt example in a complete parallelized would require to compute 100 streams with 500 multiply-accumulate operations, thus it is necessary to evaluate what is the maximum level of parallelism, and determine if the adoption of the sc is the most favorable approach. the introduction of the expression for the klt is generated in python to explore different possible implementations for the problem via the following script: [frame=single] inp = 0 fig. 16. resources required to implement klt using binary-radix (red) and sc (blue) for different numbers of input and parallel streams, with radix conversion units. outp = [] expr = [] mac_inputs = 2 num_streams = 10 for i in range(0,num_streams): for j in range(0,10,mac_inputs): for k in range(0,mac_inputs): outp.append("in" + str(inp)) inp = inp + 1 outp.append("*") outp.append("+") expr.append(outp) figure 16 presents the comparison of the resources required by both types of implementation using different numbers of inputs and parallel streams. for the sd implementation,on the left, the inputs are associated with converters, which penalizes the solution by requiring 18% extra resources. the plot on the right shows the results for the same implementation but without considering the conversion of the inputs, leading a sd solution which in the worst case requires 10% of the resources for the binary-radix solution. the results for the synthesis, in terms of resources, of the binary-radix and sc were modeled, using a linear approximation, to reduce the number of synthesis required to perform the evaluation. v. analysis a. design automation the present work could be of benefit for the engineer that is not familiarized with sc and is considering adopting it to use the processing power of a sensor node and improve the reliability of the system. by automatically producing and evaluating traditional datapaths in their sc implementations, not only is possible to compare resources but also evaluate its performance when operating variation of the operating conditions (power, temperature). fig. 17. resources required to implement klt using binary-radix (red) and sc (blue) for different numbers of input and parallel streams, without radix conversion units. b. bitstream time overhead in sc each value is encoded as a bitstream over time instead of a parallel set of bits at one. therefore, in the case-study presented the latency to produce a valid computation is given by the clock cycles to reach the length of the projection vector plus the time to perform multiplication and go through the adder tree, which is given by equation 10: t = pp rojlen + tmult + taddert ree (10) for both cases pp rojlen is the same as the number of inputs. however for the particular case of sc, tmult requires 2w l clock cycle and so does taddert ree, but delayed by one clock cycle. for typical parallel binary tmult is 1, considering a fully combinatorial multipliers, and taddert ree is log2(projlen). the tradeoff is given in terms of the size of the projection vector and the wordlength (wl) adopted. thus, the latency sc implementation would only produce results faster if log2(projsize) is greater than 2w l. a parallel binary system is able to produce a result per clock cycle, whereas an sc system requires 2w l clock cycles. vi. conclusions this work introduces stochastic theater, a framework to specify, simulate, synthesize and test sds on fpgas, targeting iot devices. it combines the bit-level specification for the processing of the stochastic bitstreams, and massive parallelization supported by the logic elements existent on fpgas. the proposed framework also introduces support for simulation of stochastic systems when in the presence of faults. this paper also presents an evaluation of the proposed framework by producing system designs to implement different arithmetic expressions. future work involves including support for process, voltage and temperature (pvt) variation as in [6] to enable further research low-power designs for stochastic computing. this stochastic framework has been implemented mainly in python and vhdl. acknowledgment this work was supported by national funds through fundação para a ciência e a tecnologia (fct) with references uid/cec/50021/2013 and ptdc/eei-hac/31819/2017. the authors would like to thank altera university program for the donation of the fpga board. references [1] s. j. agate and c. g. drury. electronic calculators: which notation is the better? applied ergonomics, 11(1):2 – 6, 1980. [2] armin alaghi and john p. hayes. survey of stochastic computing. acm trans. embed. comput. syst., 12(2s):92:1–92:19, may 2013. [3] peter alfke. efficient shift registers, lfsr counters, and long pseudo-random sequence generators, july 1996. [4] b.d. brown and h.c. card. stochastic neural computation. i. computational elements. computers, ieee transactions on, 50(9):891–905, sep 2001. [5] yun-nan chang and k.k. parhi. architectures for digital filters using stochastic computing. in acoustics, speech and signal processing (icassp), 2013 ieee international conference on, pages 2697–2701, may 2013. [6] r. p. duarte and c. bouganis. arc 2014 over-clocking klt designs on fpgas under process, voltage, and temperature variation. acm trans. reconfigurable technol. syst., 9(1):7:1– 7:17, november 2015. [7] r. p. duarte, j. lobo, j. f. ferreira, and j. dias. synthesis of bayesian machines on fpgas using stochastic arithmetic. 2nd international workshop on neuromorphic and brainbased computing systems (neucomp 2015), associated with date2015, design automation test europe 2015, march 2015. [8] r. p. duarte and h. neto. stochastic processors on fpgas to compute sensor data towards fault-tolerant iot systems. in dependable and secure computing (dsc), 2018 ieee conference on, dec 2018. [9] r. p. duarte, m. vestias, and h. neto. enhancing stochastic computations via process variation. in field programmable logic and applications (fpl), 2015 25th international conference on, pages 519–522, aug 2015. [10] r. p. duarte, m. vestias, and h. neto. xtokaxtikox: a stochastic computing-based autonomous cyber-physical system. in proceedings of the 1st ieee international conference on rebooting computing (icrc), 2016. [11] b. r. gaines. techniques of identification with the stochastic computer. in in ”proc. international federation of automatic control symposium on identification, progue, 1967. [12] b.r. gaines. stochastic computing systems. volume 2, page 37, 1965. [13] d. m. kasprzyk, c. g. drury, and w. f. bialas. human behaviour and performance in calculator use with algebraic and reverse polish notation. ergonomics, 22(9):1011–1019, 1979. [14] p. li, d. j. lilja, w. qian, m. d. riedel, and k. bazargan. logical computation on stochastic bit streams with linear finitestate machines. ieee transactions on computers, 63(6):1474– 1486, june 2014. [15] mingjie lin, ilia lebedev, and john wawrzynek. highthroughput bayesian computing machine with reconfigurable hardware. in proceedings of the 18th annual acm/sigda international symposium on field programmable gate arrays, fpga ’10, pages 73–82, new york, ny, usa, 2010. acm. [16] a.j. martin and m. nystrom. asynchronous techniques for system-on-chip design. proceedings of the ieee, 94(6):1089– 1120, june 2006. [17] p. merolla, j. arthur, f. akopyan, n. imam, r. manohar, and d.s. modha. a digital neurosynaptic core using embedded crossbar memory with 45pj per spike in 45nm. in custom integrated circuits conference (cicc), 2011 ieee, pages 1–4, sept 2011. [18] nadia nedjah and luiza de macedo mourelle. reconfigurable hardware for neural networks: binary versus stochastic. neural computing and applications, 16(3):249–255, may 2007. [19] weikang qian, xin li, m.d. riedel, k. bazargan, and d.j. lilja. an architecture for fault-tolerant computation with stochastic logic. computers, ieee transactions on, 60(1):93– 105, jan 2011. [20] n. saraf, k. bazargan, d.j. lilja, and m.d. riedel. iir filters using stochastic arithmetic. in design, automation and test in europe conference and exhibition (date), 2014, pages 1–6, march 2014. [21] m. suri, o. bichler, d. querlioz, g. palma, e. vianello, d. vuillaume, c. gamrat, and b. desalvo. cbram devices as binary synapses for low-power stochastic neuromorphic systems: auditory (cochlea) and visual (retina) cognitive processing applications. in electron devices meeting (iedm), 2012 ieee international, pages 10.3.1–10.3.4, dec 2012. [22] j. von neumann. probabilistic logics and synthesis of reliable organisms from unreliable components. in c. shannon and j. mccarthy, editors, automata studies, pages 43–98. princeton university press, 1956. [23] jieyu zhao. stochastic bit stream neural networks. phd thesis, london university, 1995. [24] fan zhou, jun liu, yi yu, xiang tian, hui liu, yaoyao hao, shaomin zhang, weidong chen, jianhua dai, and xiaoxiang zheng. field-programmable gate array implementation of a probabilistic neural network for motor cortical decoding in rats. journal of neuroscience methods, 185(2):299 – 306, 2010. on the feasibility of gpon fiber light energy harvesting for the internet of things on the feasibility of gpon fiber light energy harvesting for the internet of things j. casaleiro a,b , c. carvalho a,b , p. fazenda a , r. p. duarte c,d a adeetc, instituto superior de engenharia de lisboa (ipl-isel), cedet, lisboa, portugal b uninova/cts, instituto de desenvolvimento de novas tecnologias, caparica, portugal c dei, instituto superior técnico (ist), lisboa, portugal d inesc-id, lisboa, portugal jcasaleiro@deetc.isel.ipl.pt cfc@cedet.isel.ipl.pt pfazenda@cedet.isel.ipl.pt rpd@inesc-id.pt abstract — the emerging concept of smart cities demands for a large number of electronic devices, like sensors and actuators, distributed over several public spaces and buildings. the internet of things (iot) has a key role in connecting devices to the internet. however, the significant number of devices makes the maintenance task of the entire network difficult and expensive. to mitigate this problem, considerable research efforts have been made to develop energy-aware devices capable of self-sustainable operation, by harvesting their energy from various sources. in this paper, we study the possibility of harvesting energy from the light flowing in the gigabit passive optics network (gpon) to supply low-power devices. since most cities already have a working gpon installation, using this installation to interconnect and power iot devices can be a viable and less expensive solution, instead of installing new dedicated networks. this is also an interesting solution to convey communications and energy to low-power applications where access to the power grid is unfeasible. this study is focused in the 1550 nm wavelength, whose available optical power, in residential premises, is between -7 dbm and +2 dbm. with this range of optical power, and with a 30% efficiency photodiode, we show, for the worst-case scenario of the gpon, how it is possible to harvest 62 µw of energy at the maximum power point (mpp). keywords: energy harvesting, photodiode, optical fiber, gpon, iot, wireless sensor networks, smart cities. i. introduction powering electronic devices using an optical fiber is a widely studied topic. the earliest work in this domain, which consisted on remote powering an alarm, was presented by deloach et al. [1]. since then, various power-over-fiber (pof) systems have been proposed in the literature [2]-[21]. the development of devices capable of being powered from the same optical fiber they use for communications is a very challenging task. however, the motivation for addressing this challenge is sustained by several advantages. the power supply block that harvests energy from fiber is immune to all forms of electromagnetic interferences, short-circuits and electrostatic or atmospheric discharges [2]. moreover, optical fibers also have low attenuation and are capable of working up to considerable distances, in excess of 20 km [3], thus constituting an interesting solution to be used in remote locations where power from the grid is not available. the solutions that currently exist use optical sources with high power and dedicated fibers for powering purposes. some pof applications exist such as powering and reconfiguring remote nodes with and without batteries [5]-[6], powering optical splitters [7]-[8] and monitoring and signal measuring systems [9]-[10]. however, most of these applications make use of proprietary communication protocols that are not compatible with the existing telecommunication networks. this incompatibility makes it difficult, or even impedes, the establishment of low-cost sensor networks. for smart city applications, where the implementation of large-scale sensor networks will be needed, the use of the gpon can be a viable and less expensive solution than installing dedicated networks. most cities have installed gpons that provide 2488 mbps links for television, internet and telephone into residential buildings. the use of the existing networks for sensor interconnection has the advantage of simultaneously providing a communication channel and device powering, while avoiding the use of additional electric cables. the use of the latest generation networks to power sensors, as well as other devices, has been yet little explored [11]-[12]. therefore, this paper presents a study on the possibility of using the light flowing in these networks, namely gpon, to supply energy to devices with low-power requirements. in gpons three communication bands are used: (1) the upstream band, with wavelengths between 1260 nm and 1360 nm, (2) the downstream band, between 1480 nm and 1500 nm, and (3) the rf video-overlay, between 1550 nm and 1560 nm. the latter band is used for broadcasting analog and digital television channels. this band is the most suitable for energy harvesting because the optical power, available in the 1550nm wavelength, ranges between -7 dbm and +2 dbm. within this range, devices can extract between 60 μw and 475 μw. this paper is organized as follows: section ii presents the state-of-the-art and discusses the existing pof methods that are i-etc: isel academic journal of electronics, telecommunications and computers iot-2018 issue, vol. 4, n. 1 (2018) id-9 http://journals.isel.pt currently used to power electronic devices, using optical fiber. section iii describes the methodology and the components that were used in the practical setup for measuring the amount of energy that can be harvested from the rf video-overlay band. section iv presents the experimental results and a discussion on the feasibility of the proposed approach and section v concludes this paper. ii.background and related work the use of optical fiber telecommunication networks with powering features began to be researched in the late 1990s. the powering of devices on the client side, e.g. telephone or modems, was the focus of that research. it was demonstrated that the application was not feasible because of the relatively big amount of power required by the equipment on the client side [6], [11], [17]. however, one must stress out that from this research it has resulted the testing of a system that provided the client with 0.5 w of power. a similar application for local optical fiber networks was presented by miyakawa [18]. pof is currently being used in remote sensor systems and power distribution networks [13]. two different architectures are being used for optical power transmission: wavelength division multiplexing (wdm) and space division multiplexing (sdm). some applications use the combination of both [3][4]. in sdm, depicted in fig. 1 a), one or more fibers are used solely to power a device, while communication signals are allocated on another fibers [4]. recently, for sdm, the use of double-clad fibers has been proposed to allow for the transmission of data in the core (in single mode) and the power in the second outer layer [14]-[15]. this architecture has the advantage of avoiding the need for using optical filters to split power and data, and interference between power and communications. the optical power transmitted is limited by the characteristics of the fiber. as an example, a sdm application is provided by yasui et al. [16], where a system was developed to operate in high voltage environments, while providing 2 w continuously. a) b) fig. 1. a) sdm architecture and b) wdm for remote power supply. in wdm, shown in b), the channels are separated into several different wavelengths over the same fiber. one of the wavelengths is used for powering purposes and, at least, two more wavelengths, for sending and receiving data. at the terminal devices, the various wavelengths are separated using dielectric filters and the optical energy from the energy carrier is converted into electrical energy, using a photodiode. comparing both architectures, wdm has the advantage of using only one fiber for both pof and data. in the past, peña et al. [19] showed how remote devices can be powered (extracting up to 205 mw) with wdm. in another application, nango et al. [20] showed how to power remote sensor nodes for measurements of the electric field radiated in anechoic cameras. nevertheless, pof applications over the gpon are severely limited by the energy optimization of the network [21] and mainly by the characteristics of the single mode fibers being used [4]. in particular, the nonlinearities existing in this kind of fiber, such as the brillouin and raman scattering, introduce a low return loss, limiting the maximum power in single mode fibers to 20 dbm. the safety limits for laser emission into the fiber are defined in the iec/en 60825 standard and the maximum allowed levels are about hundreds of milliwatts, so as not to render the fiber inoperable because of micro curves [4], [6], [17]. the iot research domain has been focusing on energy harvesting solutions based on the light [22], radio frequency (rf), and vibrations [23]-[25]. the use of the energy coming from the normal operation of the fiber has not drawn too much attention [26]. the main reason is because the optical power in data transmission, using wavelengths ranging from 1310 nm to 1490 nm, is quite small, -30 dbm. however, as it was already said, the power of the rf video-overlay service, in the 1550 nm wavelength, is substantially higher (-7 dbm to +2 dbm). moreover, since this is a unidirectional service, the service provider cannot alter or adjust the power as a function of the number of clients using this service. therefore, this opens a novel pof opportunity for low-power iot applications. iii. description of the iot node blocks the iot node proposed in this paper consists of a harvester, a power management unit (pmu), an energy storage device and a load, as shown on the right-hand side of fig. 2. to harvest the energy from the fiber optics network, a pin photodiode is used. the energy obtained by the harvester can be stored, or it can be transferred directly to the load. the storage uses a capacitor sized for the specific energy requirements of the application. as a proof of concept, a microcontroller and a led are used as a load. one assumes that the node will work in a duty-cycling mode, meaning that the device has an active period, ton, and an idle period, toff, as shown in fig. 3. the operation duty-cycle is given by � = �� �� + ���� = �� � . (1) using a simplified energy balance approach, one can design the storage capacitance needed for a specific application. j. casaleiro et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-9 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt pin photodiode (1550 nm) power management unit (pmu) load olt volt wdm λ =1550 nm λ =149 0 nm λ =1310 nm optical splitter internet gpon storage rf tv fig. 2. schematic illustration of the gpon (on the left) with the optical line terminal (olt) and the rf video-overlay olt (volt) combined with a wavelength division multiplexer (wdm) and an optical splitter. on the right-hand side is the system block diagram of the iot node with the harvester (pin photodiode), pmu, storage and the load. fig. 3. power and energy budget for duty-cycling mode. thus, if the incident power and the temperature are constant, the energy supplied by the harvester can be described as � = � �� �� =�� �� × �, (2) where ph is the instantaneous output power of the harvester and t is the cycling period. regarding the load power supply, two energy levels must be considered. when active, the energy required by the load is given by �� = � ������ �� = ��� × �� , (3) where pla is the instantaneous load power when active. when the load is idle, its energy consumption is given by �� = � ����� �� = ��� ���� = ��� !� − �� #, (4) where pli is the instantaneous load power when idle. considering a capacitor as the storage element, the energy it can accumulate is given by % = &' ()*' , (5) where c is the storage capacitance and vc is the voltage level of the storage element. it should be noted that the node cannot consume all the stored energy, because a minimum voltage, vcmin, is required to supply the pmu. thus, the usable energy in each cycle is given by 0% = &' (!)*123' − )*145' #, (6) where vcmax is the threshold voltage level to protect the storage element from overcharging. this threshold can be configured in the pmu. to obtain a self-sustainable node, the following criteria are required: 89 ∙ � ∙ � + 0% ≥ ��, when active9 ∙ !1 − �# ∙ � ≥ �� + 0% , when idle, (7a) (7b) where η is the efficiency of the pmu. substituting (1), (2), (3) and (4) into (7) and assuming that all the stored energy will be delivered to the load, results in 89�� �� + 0% = ��� �� , when active9�� !� − �� # = ��� !� − �� # + 0% , when idle . (8a) (8b) solving (8a) to eus results in 0% = !��� − 9�� #�� . (9) substituting (6) into (9) and solving to c, one obtains the minimum capacitance for the storage element, given by ( = '!ijklmin#���opqrst lopquvt . (10) adding (8a) to (8b) gives 9�� � = ��� �� + ��� !� − �� #, (11) from which the duty-cycle is obtained: � = 9�� − ������ − ��� . (12) it is clear from (12) that the usable power ηph must be larger than the idle load power, i.e. ηph > pli for the system to be self-sustainable. a. harvester characterization and modeling the photodiode is used in photovoltaic (pv) mode, working as a pv cell. in this mode, the photodiode generates a current (photocurrent) proportional to the incident light power that it receives in its active area. the equivalent electrical model of a pv cell consists of a light-induced current source (i1), in parallel with a diode d (acting as a voltage limiter), a shunt resistance (rp) and a series resistance (rs), as shown in fig. 4. fig. 4. equivalent electrical model of a pv cell. j. casaleiro et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-9 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt the series resistance is due to the resistance of the metal contacts, ohmic losses in the front surface of the cell, impurity concentrations, and junction depth. from the electrical model (fig. 4), the output current (iout) is given by ^_`a = b cdcdecfg h &̂ − %̂ bijklmnopfulmnvqr − 1g − olmncd s, (13) where is is the limit of the current in the diode under high reverse bias, q is the electron elementary charge (1.60217657 × 10-19 c), k is the boltzmann constant (1.380648813 × 10−23 j/k), t is the ambient temperature, expressed in kelvin, and n is the emission (or ideality) coefficient, which equals 1 for an ideal diode. moreover, if the diode did not exhibit breakdown, the maximum reverse current that one could get through the diode, with an infinite reverse bias, would be is. another definition for it, is that it is the “dark saturation current”, i.e. the diode leakage current density in the absence of light. measurements were obtained using the fiber triplexer itr-d3t-sd6-2 from the source photonics manufacturer. using a triplexer has the advantage of having wdm that separates light for device powering (in the 1550 nm wavelength) and light for the data, as shown in fig. 5. fig. 5. internal block diagram of the optical fiber triplexer itr-d3t-sd6-2, from the source photonics manufacturer. the pin diode of the triplexer is usually used to demodulate the rf video-overlay signal and was not specifically designed for energy harvesting purposes as a pv cell is. thus, this photodiode needs to be studied for this kind of assignment. as such, to assess the power that can be harvested from the gpon, the pin photodiode was characterized, for several optical power levels. by measuring the output voltage and current of the pin photodiode for several load values, we obtained its i-v characteristics, which is shown in fig. 6. the pin photodiode characteristics were obtained for distinct optical power levels, measured using a jdsu olp-35 optical power meter. from fig. 6, the open circuit voltage of the harvester can be obtained. this is an important parameter because it sets the minimum cold start voltage of the pmu. fig. 6. photodiode i-v characteristic measured values for various incident optical power values. fig. 7. photodiode p-v characteristic for various incident optical power values. from the i-v characteristics it is possible to obtain the p-v characteristics, shown in fig. 7, which are important to determine the maximum power point (mpp). fig. 7 shows the output power (�_`a = ^_`a × )_`a) extracted from the harvester as a function of the output voltage. by looking into fig. 7, one can inspect which is the ratio between the voltage of the mpp and the open circuit voltage (voc). this ratio (k), for each incident optical power, is computed and documented in table i. table i ratio between vmpp and voc 0 0.1 0.2 0.3 0.4 0.5 0.6 0 25 50 75 100 125 150 v out (v) i o u t (  a ) -6 85 dbm -10 5 5 dbm -14 2 4 dbm -17 0 5 dbm -19 4 7 dbm 0 0.1 0.2 0.3 0.4 0.5 0.6 0 10 20 30 40 50 60 70 v out (v) p o u t (  w ) -6 85 dbm -10 5 5 dbm -14 2 4 dbm -17 0 5 dbm -19 4 7 dbm optical input power pin (dbm) open circuit voltage voc (v) mpp voltage vmpp (v) ratio k = vmpp/voc (%) -6.85 0.535 0.438 81.78 -10.55 0.512 0.429 83.77 -14.24 0.481 0.400 83.29 -17.05 0.466 0.374 80.37 -19.47 0.450 0.360 80.00 j. casaleiro et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-9 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt the value of the k ratio agrees with the ones commonly known in the literature, for example, in [27]. through the analysis of fig. 6 and fig. 7, it is possible to extract some simple data concerning the performance of the photodiode. this information is summarized in table ii. an important parameter is the conversion efficiency of the harvester (ηe), which is defined as the ratio between the maximum power output of the harvester, i.e. the power at the mpp, and the optical incident power, shown at the rightmost column of table ii. table ii measured characteristics from the photodiode optical incident power pin (dbm) short circuit current isc (µa) open circuit voltage voc (v) maximum output power pmax (µw) efficiency at the mpp ηe ηe = pmax/pin(µw) (%) -6.85 146 0.535 62 30.0 -10.55 64 0.512 25 28.4 -14.24 28 0.481 10 26.5 -17.05 15 0.466 6 30.4 -19.47 8 0.450 3 26.6 it is worth to note that �45!vwx# = 10log&�{�45!|}#~ − 30 (14) and that �45!��# = 10�uv!���#o���� . (15) by analyzing the data plotted in fig. 6, along with the electrical equivalent depicted in fig. 4, one can extrapolate the values of the parameters in the model, namely, i1, is, n, rs and rp. obtaining the value of these parameters is useful to simulate the behavior of the photodiode in an electrical simulation computer program. there are essentially two ways to extract the parameter’s values. one is solving a system of algebraic transcendental equations, and a comprehensive survey about it can be found in [28]. the other way is to use optimization algorithms that determine the parameters numerically. the least mean squares method is the most popular. the parameters are calculated by minimizing the error between the measured data and the theoretical curve and beforehand, each of the parameters to be obtained is bounded by a lower and an upper value, considered to be consistent with the order of magnitude of the true parameter. the photodiode parameters were extracted using the matlab® function lsqnonlin() with a tolerance of 10-12 and a maximum number of iterations of 1000. this function solves nonlinear least squares curve fitting problems numerically. for each of the light intensities that were tested, the extraction of the five parameters that make up the model, have the results listed in table iii. in order to confirm that the extracted values are the correct ones, the photodiode output current function (iout), shown in equation (13), is plotted using the values of table iii and checked against the measured data of fig. 6, in order to verify if they match. the resulting plots are shown in fig. 8. table iii photodiode extracted parameters for the electrical model fig. 8. photodiode i-v characteristic with measured data (dots) and analytical function using extracted parameters (lines). by observing fig. 8, one can confirm that there is a close match between the set of dots obtained by experimental measurements, and the theoretical function using the extracted parameters. moreover, by using the values of the obtained parameters into the electrical model of fig. 4, and using the ltspice® electric circuit simulator, a simulation was now run for each value of incident optical power. the resulting plots are shown in fig. 9, where they are compared with the ones already shown in fig. 8. given the functions depicted in fig. 9, it can be noted that there is also a very strong match between the analytical curves and the ones obtained by the computer simulation of the electrical model. fig. 9. photodiode i-v characteristics using analytical functions with extracted parameters (blue) and electrical simulation results (dashed red). 0 0.1 0.2 0.3 0.4 0.5 0.6 0 20 40 60 80 100 120 140 160 v out (v) i o u t (  a ) measured values with parameter fit: -6 8 5 dbm -10 55 dbm -14 24 dbm -17 05 dbm -19 47 dbm 0 0.1 0.2 0.3 0.4 0.5 0.6 0 20 40 60 80 100 120 140 160 v out (v) i o u t (  a ) experimental measurements using parameters extr ac ted w ith mat lab (r) simulated results using lt spic e ( r) incident power (dbm) i1 (µa) is (pa) n rs (ω) rp (mω) -6.85 146.5 6.18 1.224 29.6 9.52 -10.55 63.9 7.43 1.243 20.7 10.0 -14.24 27.8 5.55 1.216 0.0 3.62 -17.05 15.1 37.07 1.401 100.0 10.0 -19.47 8.1 41.44 1.435 100.0 10.0 j. casaleiro et al. | i-etc iot 2018, vol. 4, n. 1 (2018) id-9 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt although at naked eye, the electrical model function (13), as well as its electrical simulation, are very close to the original measured values, it is important to quantify how close they are. as such, the plot of the relative error is shown next, in fig. 10 and fig. 11, for the analytical and the simulated functions, respectively. fig. 10. relative error between the experimental currents and the ones obtained by the analytical functions with extracted parameters. fig. 11. relative error between the experimental currents and the ones obtained by electrical simulation of the extracted model. by inspecting both fig. 10 and fig. 11, a common pattern can be identified for each input power, in which the relative error increases with vout, especially when it approaches the value of the open circuit voltage. this is because when the derived function, either the analytical or the electrical simulated one, is in its steepest zone, there is a progressive increase in the deviation from the original values. this deviation is more critical as one gets closer to the open circuit voltage. nevertheless, at the mpp (see table i), for each of the input levels, the relative error is below 3%. in addition, for the most unfavorable value of the relative error, the absolute current error is less than 1 µa. given that the gpon operates between -7 dbm and +2 dbm, the most meaningful input power values are those at higher levels of optical power. thus, the most representative function is the one at -6.85 dbm, where the relative error tends to be smaller than for the other optical levels in the same variation zone. b. power management unit the function of the pmu is to process the energy harvested by the photodiode, stepping it up and sending it to the storage element, or directly to the load, if the storage is already full. the selection criteria to be fulfilled must encompass a stepup dc-dc converter capable of working with very low voltages, below 0.5v. the voc of the photodiode is in the range of 0.4 v to 0.6v. thus, a dc-dc step-up converter, with the capability of having a cold start input voltage below the lower bound of this range, must be selected. if the demand for this worst case is met, the pmu will always start. thus, such a converter was selected, consisting in the adp5091, provided by analog devices®. this integrated circuit was chosen because of its functional characteristics and for being a relatively recent device in the market. the cold start operating input voltage of the adp5091 is 380 mv [29]. this pmu has an evaluation board whose features were sufficient and suited, on one hand, to have the voltage coming from the photodiode stepped-up, and on the other hand, to have a supercapacitor being charged with the energy harvested from the photodiode. the abovementioned board is shown in fig. 12. fig. 12. photograph of the adp5091-2-evalz evaluation board [30]. this board encompasses a adp5091 pmu, a storage supercapacitor, and additional circuitry (resistors, mainly) to configure key voltage levels. this pmu can charge storage elements such as rechargeable batteries, supercapacitors, or even conventional capacitors. the adp5091 pmu performs maximum power point tracking (mppt), which keeps the input voltage ripple in a fixed range near the mpp of the harvester. the purpose is to make the harvesting process as efficient as possible. moreover, it has sensing modes with programming regulation points of the input voltage, which allow for the extraction of the highest possible energy from the harvester. a programmable minimum operation threshold enables shutdown during a low input condition. a typical operation circuit, taken from the manufacturer datasheet [29], is shown in fig. 13. the adopted settings for the proposed approach mainly make use of the term pin (bottom right-hand side of fig. 13), whose function is to set the value up to which the supercapacitor will charge (vcmax, previously defined). using the default settings of the board, this voltage is preset to 3.5 v. 0 0.1 0.2 0 3 0.4 0.5 0.6 0.7 -15 -10 -5 0 5 10 15 20 25 v out (v) r e la ti v e e rr o r fo r a n a ly ti c a l fu n c ti o n s (% ) -6.85 dbm -10 55 dbm -14 24 dbm -17.05 dbm -19.47 dbm 0 0.1 0.2 0 3 0.4 0.5 0.6 0.7 -25 -20 -15 -10 -5 0 5 10 v out (v) r e la ti v e e rr o r fo r e le c tr ic a l s im u la ti o n s ( % ) -6.85 dbm -10 55 dbm -14 24 dbm -17.05 dbm -19.47 dbm fig. 13. typical application circuit [29]. to perform the mppt, and based upon the data gathered in table i, the value of k that was chosen to work with the dcdc converter is 82%. this value can be set by establishing a voltage divider over the input voltage. one can check in fig. 13 that, at the left-hand side, between the vin terminal and the ground terminal, there is a voltage divider, yielding a voltage to the mppt pin. the ratio of this voltage divider is the one that needs to be set, to have k ≈ 0.82 (see table i). thus, the pull-down resistor (r18 in [30]) is set to 18 mω, keeping the default value used the evaluation board. however, value of the pull-up resistor (r15) needed to be adjusted to 3.9 mω, leading to a change in the original value used by the board, so that the value experimentally determined for k could be met. in fact, when wiring together the photodiode and the board, and using the original supercapacitor shipped with it (0.1 f), it takes 8h47m to fully charge it to 3.5 v. if the original setup was used instead (with r15 = 4.7 mω), it would take 9h11m to achieve the same goal. thus, it takes 24 minutes more, as the mpp is deviated from the optimum point, confirming the importance of setting the mppt voltage divider to the value determined in table i. even though the board itself is fitted with a 0.1 f supercapacitor, this capacitor was replaced by several other capacitors with higher values, to store more energy and study the charging curve over an extended period of time. c. storage device the harvested energy is stored in a supercapacitor. this device has the purpose of storing energy using a double layer between an electrolyte and a solid, in which the inner structure is composed by two electrodes immersed into an electrolyte, which can be liquid or solid, and that are separated by a membrane. these capacitors differ from rechargeable batteries because they store energy at the surface of the electrodes, unlike batteries, that store energy thanks to an electrochemical reaction. because of this, supercapacitors can stand a higher number of charge and discharge cycles than batteries can, being suited for applications where this kind of regime is usual. the number of these cycles can be as high as a million, leading to an operational lifetime of ten years, until the capacitance value starts to show some signs of degradation [31], because of the degradation of the electrodes and of the electrolyte solution. one very appealing factor about supercapacitors is that they do not require specific charging circuits, being able to stand trickle charging. supercapacitors are inexpensive, making them very appealing to use in opposition to rechargeable batteries, as these are more expensive. examples of systems that are designed to make use of a supercapacitor to store harvested light energy can be found in [32]-[33]. there are also some applications that use both a battery and a supercapacitor [34]. these act as primary and secondary energy storage, respectively. with rechargeable batteries, there are only a few typical voltage ratings, depending on the technology being used. with supercapacitors, these ratings are much more diverse, similarly to regular capacitors. this factor is also important, not only because of the end application, but also because it can result in a smaller device, if a lower voltage rating is allowed. fig. 14 shows the supercapacitors that were selected to be used in the experiments with the energy harvesting application in this work, all of them having a rating of 5 v. fig. 14. supercapacitors selected for the application (5 f, 3 f, 1 f and 0.1 f). it must be considered that, for typical low-power energy harvesting applications, the size of the whole system is intended to be small. as such, although the capacitance value, for supercapacitors, can reach values as high as 3000 f, for practical small sized applications, due to body size restrictions, the selected supercapacitors must have lower values, as well as their voltage rating. the higher the capacitance value, the larger will be the leakage current due to self-discharge. moreover, the leakage current increases in proportion to the increase in the voltage at the terminals of the supercapacitor. the leakage current can be modeled as a resistor placed in parallel with the capacitor. according to the manufacturer’s datasheet [35]-[36], the leakage current for each capacitor in fig. 14 is presented in table iv. table iv leakage current for the supercapacitors that were used capacitance (f) leakage current @20ºc and @5.0v (μa) 0.1 3 1.0 12 3.0 16 5.0 25 iv.experimental results some experimental results have been determined so far, given that the photodiode was characterized from an experimental point of view. from the set of measurements, by varying the output conditions of the photodiode from short circuit to open circuit, it was possible to obtain the various values of current that, subsequently, allowed for obtaining the data in table i, table ii and table iii. the measurements setup is shown in fig. 15, which includes one oscilloscope (tektronics tds 2004b) and two digital multimeters (agilent keysight 34401a). the laser was generated using a jdsu (ols-38), at the wavelength of 1550 nm and a power of -7 dbm. fig. 15. photo of the measurements setup. a. performance of the supercapacitors several tests have been run to determine how long would it take to charge a supercapacitor of a given value, connected to the bat pin of the dc-dc converter, from a 0 v condition (cold start). just as described in section iii.b, the terminal voltage of the charging process is set to 3.5 v. for each of the devices shown in fig. 14, the results of their charge, using an optical incident power of -6.85 dbm, at the wavelength of 1550 nm, have been recorded. the voltage variation at terminals of the supercapacitor (vc) is shown for each unit. the horizontal axis is normalized to hours per farad, so that all the voltages representing the charge can be compared on a common basis. fig. 16. normalized voltage variation (during charge) of the supercapacitors. the time needed to charge each of the supercapacitors up to 3.5 v (starting from zero), is indicated in table v, as well as their stored energy. from fig. 16, one can conclude that the supercapacitor that has the highest rate of charge per unit of capacitance is the one with 1 f. table v charging data for each supercapacitor capacitance (f) time of charge to 3.5 v (days hours minutes) stored energy (j) 0.1 00d 08h 47m 0.6125 1.0 03d 09h 10m 6.1250 3.0 14d 16h 09m 18.375 5.0 33d 16h 18m 30.625 although, in principle, this does not agree with the data in table iv, it is to note that both the temperature and the working voltage that were used are not the same as in the datasheet, nor constant. moreover, when referring to the source of input power to the circuit (the photodiode), one must bear in mind that temperature is a variable to consider for the current generation. for those longer periods of time to get the supercapacitor fully charged, the temperature variation is bigger, given that it spans for a full 24-hour cycle, for several days. also, a common remark for any of the capacitances reported in fig. 16, is that when the voltage reaches slightly less than 2.5 v, the charging regime suffers a change. this change is directly associated with the mode of operation of the dc-dc converter, when it enters synchronous mode [29] and the output voltage is driven to follow the voltage of the storage device. b. powering an electronic application as described in the datasheet of the adp5091, the average value for the efficiency of the dc-dc converter is about 80% (assuming its worst-case scenario). to check if the approach tried in this paper is feasible, a simple electronic application was built. the power supply is obtained from the pmu to demonstrate that it is possible to harvest energy from the gpon, store it, and use it to supply a low power iot sensor node. the application consists of a pic16f1459 micro controller unit (mcu), running a program that periodically turns a led on and off using a very low duty-cycle. the timings where calculated so that this application, after starting up, could permanently remain in operation. during the interval of time when the led stays turned off, the mcu remains turned on, in a sleeping mode, maintaining the internal oscillator running and the watchdog timer to periodically wake up. a picture of this application, being powered by the pmu, is shown in fig. 17. fig. 17. electronic application: powering a mcu that flashes a led. 0 20 40 60 80 100 120 140 160 0 0.5 1 1.5 2 2.5 3 3.5 time vs capacitance (h/f) v c ( v ) 0 1 f 1 f 3 f 5 f the optical generator provides a laser with an incident power of approximately -7 dbm (actually, -6.85 dbm), which is converted to electrical power by the photodiode and is delivered to the pmu that manages the charge of a capacitor and the power supply for the application. in fig. 18, it is shown a cold start of the pmu, where it can be identified the various stages that the supplying voltage goes through until it starts powering up the mcu. vsys vbat fast charging main boost in asynchronous mode main boost in synchronous mode mcu power up fig. 18. phases of operation, from cold start until the mcu (load) powers up. as it can be seen fig. 18, this takes about 72 s. note that when vbat reaches about 2.5 v, i.e. about 60 s after the cold start, the dc-dc converter enters the synchronous mode, as already mentioned. vbat is the name adopted by the datasheet of the adp5091 to identify the voltage of the storage device. however, this voltage has been introduced before in this paper and is named vc. vsys is the output voltage of the pmu, which directly interfaces with the load to supply it. the adp5091 is configured to swing the charging and discharging voltages of the storage element between vcmax = 3.2 v and vcmin = 2.5 v, respectively. let us consider the scenario where the load, when active, requires pla = 2.5 mw during ton = 211 ms and pli = 35.5 μw, when idle. according to (10) and (12), the value of the storage capacitance will be c = 259 μf and the duty-cycle δ ≈ 0.5%. using the e6 series, the nearest capacitance value is 330 μf. the cycling period is t = ton / δ = 42.2 s. since the capacitor is small, when there is a demand for energy, it rapidly drains out. this is the reason why there is a drop in the value of vsys and vbat when the mcu is turned on, because the first thing it does is to turn the led on, thus demanding a reasonable amount of current. the current consumption when the mcu is in sleep mode is 12.5 µa, as measured using a digital ammeter. when the mcu wakes up, it turns the led on for approximately 211 ms increasing the supply current to 880 μa. in addition, it can also be noted that the decreasing variation in vbat is approximately 500 mv, as shown in fig. 19. to have a broader perspective about how the voltages in fig. 19 evolve over time, one can observe fig. 20. from the above results we conclude that the minimum usable power that can be extracted from a gpon is 48 µw, for -7 dbm (200 µw) of optical power, assuming an efficiency of 30% for the photodiode and 80% for the dc-dc converter. if a more power demanding application is to be supplied, a longer interval must be allowed for the storage device to charge, in addition to having a device with a higher storage capacity. if the purpose is to supply a ont module for data transmission, encompassing the phases that go from powering up the module, obtaining the ip address and then communicating the intended data, enough energy must be harvested. after the characterization of the various supercapacitors, it is possible to estimate the energy needed to power a gpon small form-factor pluggable optical network terminal (gpon-sfp-ont). these have a lower power consumption when compared to the power that other onts put available to customers through service providers. fig. 19. detail showing the current consumption and related waveforms. fig. 20. variation of the working voltages during long term operation. in table vi, the consumption of some gpon-sfp-ont modules is shown, to serve as a guide to size the features of the supplying system. table vi sfp gpon-ont modules manufacturer / model current [ma] power [w] finisar ftgn2117p2xxn 450 1.418 wtd rtxm167-522 200 0.627 microtik fg1537twgpa04t8 600 1.884 prolabs gpon-sfp-olt-b+-nc 500 1.565 delta electronics opgp-34-a4b3sl-b 400 1.252 let us consider the use of the less power hungry gpon-sfpont in table vi, the wtd rtxm167-522 [37], which has a bit rate of 2488 mbps for upstream and 1244 mbps on downstream. if this device is used for 10 seconds, to periodically transmit data collected over a relatively long period (e.g. billing telemetry), system sizing can be carried out as follows. for the same voltage swing in vbat (vc) as before, if a constant power of 627 mw and an on time of 10 s is considered, according to equations (10) and (12), the value of the storage capacitance is c = 3.2 f and the duty-cycle is δ ≈ 19.94×10-6, i.e t = 5.8 days. note that this result does not take into consideration the leakage current of the storage device, which increases with larger capacitance values. for a capacitance of 3 f, the leakage current cannot be neglected, as shown in fig. 16 and table iv. moreover, considering the leakage and temperature variation, one obtains (from fig. 16) a charging time of 3.17 days, i.e. 3 days and 4 hours. however, note that the charging times in fig. 16 are taken with no load, i.e. pli = 0 µw, which explains the shorter charging time, compared to theoretical expected results. v. conclusions this paper presented a feasibility study about harvesting light energy flowing in gpon, store it, and use it to supply low-power nodes for the iot. the harvested light comes from the rf video-overlay wavelength in 1550 nm. the instantaneous energy extracted from the fiber is small and, in the worst case, it is 60 µw. storing it into a supercapacitor over time, allows for periodically powering an application with more demanding requirements. experimental results show that it is possible to power low-power nodes from a gpon with low duty-cycle activity and extremely low power when devices are in idle or sleeping. in the worst case, it should be less than 48 µw. the theory was validated by a prototype that periodically powers a 2.5 mw load during 211 ms, with a period of about 44.8 seconds. with an incident power of -6.85 dbm, the present study can serve as a worst-case scenario that can be obtained from the gpon. a theoretical scenario to power a gpon onu with a 627 mw during 10 seconds with a periodicity of 6 days was presented. theory shows that it is possible to power a iot node from the gpon and use the gpon to send the collected data to a remote server. a design framework was derived, so that the designer can conveniently determine both the values of the storage capacitance and the operating duty-cycle. in the approach that was followed, a working methodology was also established to characterize the harvester photodiode using a numerical least squares approach and establish its electrical model. in addition, several supercapacitors were studied, by letting them charge over a large period, encompassing several days, as it would happen in a real low duty-cycle situation. from this study, the effects of the leakage current were observed. a commercial pmu was selected, to manage the charging process of the storage device and to serve as the power supply to the load. the circuitry around the pmu was configured to be in accordance to the mpp of the harvester, thus enabling mppt. acknowledgements the authors wish to thank to fábio martinho and rui lopes for the measurements of the supercapacitors charging voltages. also, the authors wish to thank to sérgio andré for his helpful work in disassembling the optical module and for several helpful discussions about this work. references [1] b.c. deloach, r.c. miller and s. kaufman. sound alerter powered over an optical fiber. bell syst. tech. journal, vol. 57, pp. 3303-3316, 1978 [2] a. basanskaya. electricity over glass [fiber optic to transfer electric power]. ieee spectrum, vol. 42, no. 10, pp. 18, 2005. [3] m. dumke, et. al. power transmission by optical fibers for component inherent communication. journal of systemics, cybernetics and informatics, vol. 8, no.1, pp. 55-60, 2010. [4] t. c. banwell, r. c. estes, l. a. reith, p. w. shumate and e. m. vogel, powering the fiber loop optically-a cost analysis. journal of lightwave technology, vol. 11, no. 3, pp. 481-494, march 1993. [5] j.g. werthen, a.g. andersson, h.o. björklund and s.t. weiss. current measurements using optical power, in proc. transmission and distribution conference, pages 213-218, 1996. [6] s. zadvornov and a. sokolovsky. an electro-optic hybrid multifunctional instrument for 3-phase current measurements, in proc. conference on instrumentation and measurement technology, 2008. [7] h. ramanitra, p. chanclou, j. etrillard, y. anma, h. nakada and h. ono. optical access network using a selflatching variable splitter remotely powered through an optical fiber link. optical engineering, vol. 46, no. 4, 2007. [8] j.h. lee, k.-m. choi and c.-h. lee. a remotely reconfigurable remote node for next-generation access networks. ieee photonic technology letters, vol. 20, no. 11, pp. 915-917, 2008. [9] a.p. goutzoulis, j.m. zomp, and a.h. johnson. development and antenna range demonstration of an eight-element optically powered directly modulated receive uhf fiberoptic manifold. ieee journal of lightwave technology, vol. 14, no. 11, pp. 2499-2505, 1996. [10] j. lim, p.r. jackson, b.e. jones, k.f. hale and q.p. yang. an intrinsically safe optically powered hydraulic valve, in proc. 7th international conference on new actuators, 2000, pages 216-219. [11] d. kuhn, e. lo and t. robbins. powering issues in an optical fiber customer access network, in proc. 13th international telecommunications energy conference, 1991, pages 51-58. [12] s. al-chalabi, optically powered telephone system over optical fiber with high service availability and low risk of investment in ftth infrastructure. ieee communications magazine, vol. 50, no. 8, pp. 102-109, august 2012. [13] j.-g. werthen. powering next generation networks by laser light over fiber, in proc. optical fiber communication/national fiber optic engineers conference, 2008, pages 1-3. [14] m. valentine. power over fiber shines at voltage isolation. power electronics technology, 2007. [15] m. matsuura and j. sato. power-over-fiber using doubleclad fibers for radio-over-fiber systems, in proc. 19th european conference on networks and optical communications (noc), milano, italy, june 2014, pages 126-131. [16] t. yasui, j. ohwaki, m. mino and t. sakai. a stable 2-w supply optical-powering system, in proc. 28th photovoltaic specialists conference, 2000, pages 16141617. [17] l.j. cashdollar and k.p. chen. fiber bragg grating flow sensors powered by in-fiber light. ieee sensors journal, vol. 5, no. 6, pp. 1327-1331, 2005. [18] h. miyakawa, e. herawaty, m. yoshimoto, y. tanaka, and t. kurokawa. power-over-optical local area network systems, in proc. 3rd world conference on photovoltaic energy conversion, 2003, pages 24662469. [19] r. peña, c. algora, i.r. matias, and m. loper-amo. fiberbased 205-mw (27% efficiency) power-delivery system for an all-fiber with optoelectronic sensor units. applied optics, vol. 38, no. 12, pages 2463-2466, 1999. [20] t. nango, t. kawashima, j. ohwaki and m. tokuda. new imitated equipment with optical powering system for evaluating anechoic chamber characteristics, in proc. international symposium on electromagnetic compatibility, 2001, pages 274-279. [21] g. böttger et al. an optically powered video camera link. ieee photonics technology letters, vol. 20, no. 1, pp. 39-41, jan 2008. [22] j. jeong, x. jiangand and d. culler. design and analysis of micro-solar power systems for wireless sensor networks, in proc. of the 5th international conference on networked sensing systems (inss 2008), 17-19 june 2008, pages 181-188 [23] p. kamalinejad, c. mahapatra, z. sheng, s. mirabbasi, v.c.m. leung and y.l. guanm. wireless energy harvesting for the internet of things. ieee communication magazine, vol. 53, no. 6, pp. 102-108, june 2015. [24] m. gorlatova, j. sarik, g. grebla, m. cong, i. kymissis and g. zussman. movers and shakers: kinetic energy harvesting for the internet of things. ieee journal on selected areas in communications, vol. 33, no. 8, pp. 1624-1639, aug. 2015. [25] j. wang et al. power-over-fiber technique based sensing systems for internet of things, in proc. 15th international conference on optical communications and network (icocn), hangzhou, 2016, pages 1-3. [26] h. ujikawa, t. yamada, k. i. suzuki, a. otaka, h. nishiyama and n. kato, stand-alone and cooperative deep sleep for battery-driven optical network unit, in ieee internet of things journal, vol. 3, no. 4, pp. 494502, aug. 2016. [27] t. esram and p.l. chapman. comparison of photovoltaic array maximum power point tracking techniques. ieee transactions on energy conversion, vol. 22, no.2, pp. 439-449, june 2007. [28] s. lineykin, m. averbukh and a. kuperman. five— parameter model of photovoltaic cell based on stc data and dimensionless, in proc. ieee 27th convention of electrical and electronics engineers in israel, eilat, israel, 2012, pages 1-5. [29] analog devices. ultralow power energy harvester pmus with mppt and charge management ad50915092 datasheet, 2016. [30] analog devices. adp5091-2-evalz user guide ug927, 2016. [31] p.h. chou and c. park. energy-efficient platform designs for real-world wireless sensing applications, in proc. ieee acm international conference on computer-aided design (iccad-2005), 6-10 november 2005, pages 913-920. [32] c. carvalho and n. paulino. indoor light energy harvesting system for wireless sensing applications. isbn: 978-3-319-21616-4, springer international publishing, 2016. [33] f. simjee and p.h. chou. everlast: long-life, supercapacitor-operated wireless sensor node, in proc. 2006 international symposium on low power electronics and design (islped'06), 4-6 october 2006, pages 197-202. [34] x. jiang, j. polastre and d. culler. perpetual environmentally powered sensor networks, in proc. fourth international symposium on information processing in sensor networks (ipsn 2005), 15 april 2005, pages 463-468. [35] eaton, powerstor pb family 5.0 volt cylindrical supercapacitors technical data 4393, march 2015. [36] eaton, phb supercapacitors cylindrical pack technical data 4402, june 2017. [37] wuhan telecommunication devices co., rtxm167522, wuhan telecommunication devices co. ltd., 2012. passive optical communications module for the internet of things passive optical communications module for the internet of things ribeiro,t.a a department of electrical and computer engineering, instituto superior técnico tiagorib93@gmail.com abstract the internet of things (iot) promotes interconnectivity between devices and these keep appearing in larger quantities throughout the years, with the evolution of communication technologies. however, scalability comes with a price, since for a higher quantity of devices comes the need for better transmission channels, with higher reach, availability and improved security capabilities. wireless technologies already provide a way to solve these issues, but the use of optical fibers would give to the iot their own unique features. but iot devices should not be power hungry nor have costly electrical-to-optical conversions, so a passive optical communications module based of fiber bragg should be implemented. this module would be integrated in the iot ecosystem by connecting it to the many existent dark fibers all over the world. a simulator of this module was implemented, capable of reproducing its characteristics for the transmission of information modulated in frequency-shift keying and on-off keying modulation schemes. keywords: internet of things, fiber bragg gratings, acoustooptic modulator, frequency-shift keying, on-off keying i. introduction the increase in internet traffic over these past decades was only made possible with the deployment of a worldwide optical network. the long-haul fiber networks have evolved into a complex web with mesh connectivity in metropolitan areas, pushing the fiber into the edge of the network, thus forming a fiber-to-the-home (ftth) network. therefore, the fiber is no longer used solely for intercity links, but also supports metro and last-mile connectivity. it can also be found in industrial plants, alongside highways, remote rural roads, power transmission lines and railways. of course, with this large presence of optical fiber cables, there are some spare fibers available, known as dark fibers, that can be explored for broadband capability and for providing basic connectivity with newly developed technologies [1]. however, the human ability to consume information is no longer the only drive for setting the limits to the required network bandwidth, but the by now dominant amount of machine-to-machine (m2m) traffic that rises from data-centric applications, sensor networks and the growing penetration of the internet of things (iot). the iot can and will impact manufacturing and supply chains from industries around the world, whose main goal is to increase efficiency by controlling their machines remotely, where environmental conditions may be unfavorable to allow wireless networks to operate. for a certain industry to achieve this objective, it must use fiber optic cables which, due to the previously mentioned characteristics, are the best medium for the handling of information that is transmitted throughout the facility. this, of course, will lead to a growth of in-building fiber deployments not only for industrial facilities, but also for residential buildings which cannot be reached wirelessly [2]. the iot can be considered as a global network which provides communication between human-to-human, human-to-things and things-to-things, each one with their own unique identity. however this requires a global infrastructure of networked physical objects that enables anytime, anyplace connectivity for anything, not only for any one [3]. a. motivation and main objectives in 2016, at the 21st optoelectronics and communications conference, it was stated that, by the year of 2019, 83% of global data traffic is expected to come from cloud services and applications and it will gather a total data of 10.4 zettabytes per year [4]. by the year of 2008, the total number of interconnected devices over the internet exceeded the world’s population for the first time, and the trend of attaching more “things” of daily use to this network is accelerating. it is estimated that by the year of 2020, the number of devices connected with each other over the internet, the so called internet of things, is expected to be around 50 billion, as depicted in figure 1, with the left axis referring to the world’s population and internet-connected devices and the right one, the internet-connected devices per person. fig. 1: bar chart of the number of connected devices compared to the world’s population [1]. the iot limits are mainly based on the economic value that its featured services provide to society. the deployment of computational and storage capacities to support these services depends on the potential profits and gains obtained from building an infrastructure. if this network continues to be increasingly more ubiquitous, the investment in these infrastructures will still increase and the doubt remains if the same number of large companies will keep paying for it. thus innovation needs to happen in a way in which the cost of new i-etc: isel academic journal of electronics, telecommunications and computers iot-2018 issue, vol. 4, n. 1 (2018) id-5 http://journals.isel.pt mailto:tiagorib93@gmail.com file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_1 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_2 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_3 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_4 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_1 components designed for optical fiber systems are reduced considerably [5]. for the iot’s transmission layer, the aim is to transfer data over long distances or large areas. several wireless standards for iot applications such as zigbee and wi-fi are capable of successfully transmitting data to a given end-user, although, for an industrial application, a wired solution is more feasible. wireless communications for wide areas have several issues such as spectrum usage, causing loss of the quality of the signal, and the environment topography, from which comes the degradation of the transmitted signals due to the distance getting bigger and because of the multipath fading effects that may also induce signal-to-noise fluctuations that may cause unreliability towards the requirements for providing connectivity for critical iot applications. adding to these issues, the broadcasting nature of wireless connectivity has some evidently fundamental privacy and security vulnerabilities. an optical wired communication system can be adopted to contradict these issues, since the optical fiber offers a huge bandwidth with very low attenuation, typically 0.2 db/km for a standard single mode fiber (smf), allowing the transmission of data over tens of kilometers. the concept of iot over fiber (iotof) uses fiber bragg gratings (fbgs) to modulate the information coming from iot devices in the optical carrier and transmit it over a large distance of over tens or maybe thousands of kilometers, when optical amplification schemes are used. the connection of this novel technology with the already existent optical fiber pipeline is done by using the dark fibers available, which are key for the iot concept. to provide the direct physical connectivity (phy) to iot devices themselves might seem an overkill solution, but this would be viable and applicable when availability, reliability, security and other limiting factors handicap the physical connectivity solution already based on wireless and other wireless technologies. the iotof system can be used in the monitorization of urban areas with a crowded and/or polluted spectrum, power lines structure and environment monitoring, collecting data in the gas and oil industries. also, it can be used in wildfire monitoring in extremely remote locations, since the dense forests in hilly regions can create very unfavorable wireless propagation environments. b. state of the art several wireless standards for iot applications have been proposed during the last two decades, with table 1 depicting the comparison between the wireless iot technologies and the novel iotof introduced in this study. table i wireless iot technologies and the iotof concept technology data throughput (bps) range (km) bandwidth (mhz) spectral efficiency (bps/hz) lte emtc 1×10 6 11 1.4 0.714 wi-fi ieee 802.11ac 500×10 6 0 1 via mesh 80 6.25 bluetooth 2×10 6 0.75 2 1 thread 250×10 3 0 1 5 0.05 zigbee 250×10 3 0.13 los 2 0.125 z-wave 100×10 3 0.03 0.2 0.5 lora 22×10 3 30 (water) 15 (ground) 0.250 0.088 sigfox 100 30 0.2 0.0005 iotof 300 >30 0.0008 0.375 the continuous increase of traffic calls for the deployment of a worldwide optical network in which long-haul fiber networks form a complex web of mesh connectivity in metropolitan areas. also, with wavelength division multiplexing (wdm) techniques, these optical fibers can support several bandwidthhungry applications [1]. despite also enabling the iot ecosystem over optical fibers, the iotof concept presented in this study is a niche solution to fill the gaps from an iot arena dominated by extremely low cost and widespread wireless solutions, therefore one of the main challenges in this novel developed system is the ease of capital and operational costs (capex and opex, respectively). however, optical solutions are costly, not because of the optical fiber itself, but because they involve electrical-to-optical (e/o) and optical-to-electrical (o/e) interfaces. such interfaces are usually power-hungry circuits, whereas optical fibers are electrically passive elements. thus, energy presents itself as a major constraint for iot solutions and these conversions need to be addressed in iotof phy architectures before benefiting from long-reach with reliability, privacy and other appealing features that optical fibers have, contrasting with the wireless ones. table 2 depicts some matching factors between conventional optical communication connectivity and the proposed iotof phy architecture. the data throughput required by individual iot devices are usually several orders of magnitude below what the optical interfaces can offer. long reach reliability from optical fibers is what iotof aims to achieve in the iot ecosystem, thus these requirements cannot be compromised by low cost optical solutions, i.e., based on leds, for example. installation and maintenance in conventional optical systems is costly and time-consuming at e/o and o/e interfaces, since these involve specialized splicing equipment and connectors, whereas the wireless counterpart operation is virtually effortless and costless. with this in mind, iotof should be aimed at extremely simple operations. table ii iot requirements and conventional optical and iotof phys [6-8]. iot requirements conventional optical phy @ onu (per subscriber) iotof optical phy throughput up to 2.5 gbps at least 100 bps range up to 20 km at least 20 km power consumption around 1 w 0 w (passive) also, conventional optical communication e/o and o/e conversions are power consuming due to the need of biasing and cooling circuitry, while iotof would aim, if possible, at optical passive solutions at iot devices. spectral efficiency is not a major concern in optical systems seen in the widespread use of ook and direct detection to reduce cost. in contrast, the extremely low-cost requirement from the iot systems will push iotof solutions with limited end-to-end bandwidth and, therefore, spectral efficiency will become an issue to be addressed [9]. figure 2 depicts the comparison that can be made between the discussed technologies both in terms of bandwidth, which is related to the data throughput through the spectral efficiency, and the transmission distance. these parameters lead to a tradet. ribeiro | i-etc iot 2018, vol. 4, n. 1 (2018) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_5 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_1 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_6 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_9 off correlated to the frequency bands, i.e., higher frequency bands have more channels and more bandwidth, which allows for more data throughput, however, the transmission range is much lower. on the opposite hand, lower frequency radio waves are less vulnerable to propagation disturbances than the higher ones, presenting a higher operation range, but the data throughput is reduced [10]. fig. 2. relation between throughput and range for iot standards, with the inclusion of the novel iotof concept. in sum, iotof should aim at ranges beyond conventional wireless solutions, but without compromising minimal throughputs achieved by them. this novel technology is therefore expected to exceed a fundamental wireless limit, i.e., the radio horizon range (rhr), which is the furthest reach imposed by the earth’s curvature line-of-sight (los) propagation, including the atmosphere’s refraction effects. the rhr for a 30 m height gateway antenna and a base station at 1 m is around 20 km, which is proved experimentally as an achievable range by lora and sigfox. this is also g-pon’s basic reach, thus 20 km provides for an important optical and wireless landmark for iotof to surpass and find its own niche [7, 11, 12]. ii. theoretical foundation a. fiber bragg grating technology during the development of fiber optics technology, fbgs were applied in many photonic devices. an fbg is an optical filter inscribed in a short segment of an optical fiber, so that it can reflect specific wavelengths coming from a certain optical source and transmit all others. it can also be regarded as a fiber device with periodical variation of the refraction index of the core along the fiber [13]. due to this successful implementation, fbgs proved their utility in a wide set of communication applications in the optical realm. they can be used as add/drop multiplexers for wdm resulting in an increased capacity of optical networks, as mode converters, as dispersion compensators and as pulse compressors [14]. also, by studying the displacement of a given reference wavelength it is possible to measure several physical parameters, such as temperature, pressure, strain, etc. thus, fbgs can be used as sensors and have many advantages, like the lack of los requirement, resistance to corrosion, immunity to electromagnetic interference and the easiness of implementation in miniaturization. also very important, is the fact that it facilitates the remote control and processing of the information, simply because both sensing and propagation are rolled into one [4, 13, 15]. cladding co e input reflected ransmitted λ δn r e fr a c ti o n in d e x position in the fiber fig.3. schematic of the fbg with the spectral input and outputs. from the first order bragg condition results the reflected wavelength, also called the bragg wavelength and it is given by the expression in equation 2.1: 𝜆𝐵 = 2λ𝑛𝑒𝑓𝑓 (2.1) where 𝜆𝐵 is the central bragg wavelength of the reflected signal, 𝑛𝑒𝑓𝑓 is the effective refraction index of the optical fiber, which is an average of the refraction index of the core and of the cladding, and λ is the period of the modulation of the refraction index in the core of the fiber. the periodic perturbation of the refraction index (𝛿𝑛𝑒𝑓𝑓 ) along the propagation axis 𝑧 can be described by equation 2.2: 𝛿𝑛𝑒𝑓𝑓 (𝑧) = 𝛿𝑛𝑒𝑓𝑓̅̅ ̅̅ ̅̅ ̅(𝑧) {1 + 𝜉(𝑧) cos [ 2𝜋 λ 𝑧 + 𝜙(𝑧)]} (2.2) where 𝛿𝑛𝑒𝑓𝑓̅̅ ̅̅ ̅̅ ̅ is the variation of the mean value of the modulation, 𝜉 is the fringe visibility of the index change varying from 0 to 1, and 𝜙(𝑧) describes grating chirp for aperiodic gratings [16, 17]. for a periodic single-mode bragg reflection grating the relations depicted in equations 2.3 and 2.4 are known: 𝛿 = 2𝜋𝑛𝑒𝑓𝑓 ( 1 𝜆 − 1 𝜆𝐵 ) (2.3) к = 𝛿𝑛𝑒𝑓𝑓 𝜋 𝜆 (2.4) from equation 2.3, comes the value called the detuning, which is independent of 𝑧 for all gratings. to simplify the upcoming equation 2.6, it is useful to define the following variable of equation 2.5: 𝛾 = √к2 − 𝛿 2 (2.5) using the previous equations, the reflectivity of a grating with constant modulation amplitude and period is obtained as function of the wavelength 𝜆 and the length 𝐿 of the grating, as shown in equation 2.6: 𝑅𝜆,𝐿 = sinh2(𝛾𝐿) cosh2(𝛾𝐿) − 𝛿 2 к2 (2.6) b. fbg’s manufacture techniques there are many grating manufacture techniques, but in this study only the phase mask technique and the interferometer with phase mask technique will be considered. the optical element known as the phase mask is made from t. ribeiro | i-etc iot 2018, vol. 4, n. 1 (2018) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_10 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_7 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_11 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_12 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_13 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_14 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_4 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_13 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_15 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_16 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_17 a flat slab of silica glass, which is transparent to uv light, and it features a periodic pattern shape that approximates a square wave in profile. the uv light, which is incident normally to the phase mask, passes through it and is diffracted by the periodic corrugations of the mask. these corrugations are almost in contact with the optical fiber. usually, most of the diffracted light is contained in the 0, +1 and -1 diffracted orders, however, the phase mask is designed in such a way that it is able to suppress the diffraction into the zero-order to less than 5% and can divide 40% of the total light intensity equally in the 1+ orders, approximately. in other words, the zero-order diffraction is minimized and the +1 and -1 orders are maximized. the interference of these maximized order beams produce a periodic pattern that photoimprints a corresponding grating in the fiber. the obtained period of the grating pattern in the core of the fiber is λ = λpm/2, where λpm is phase mask period [17, 18]. the phase mask technique can also be used to manufacture gratings with controlled spectral response characteristics. the typical spectral response of a finite length grating with a uniform index modulation along the fiber length has a secondary maximum on both sides of the main reflection peak. in wdm applications this kind of response is not desirable, so in order to avoid secondary maxima, a bell-like functional shape is given to the fbg. this process is called apodisation, which through the years have achieved suppressions of the sidelobes of 30 db to 40 db, using the phase mask technique [18]. the apodisation technique also extends to the manufacture of chirped fbgs (cfbgs), also known as aperiodic fiber gratings, which are used for making dispersion compensators. the chirping aspect means varying the grating period along the length of the fiber in order to broaden its spectral response [18]. in the context of “amplitude-splitting interferometer” [17], the fringe grating period can me altered by varying the incidence angle, 𝛼, or by modifying the wavelength of the incident radiation, 𝜆𝑈𝑉 . the choice of this wavelength is limited to the uv photosensitivity region of the fiber, however, there is no restriction for the choice of the angle 𝛼. the amplitude-splitting interferometer offers the ability to inscribe bragg gratings of various characteristics, but it is susceptible to mechanical vibrations. this disadvantage originates from sub-micron displacements in the position of the mirrors, the beam splitter, or other optical mounts in the interferometer during uv irradiation, causing the fringe pattern to drift [19]. in the interferometer with phase mask technique, the uv beam splitting is done by a phase mask. the diffracted orders are reflected on lateral mirrors and are recombined in the fiber where they interfere with a certain pattern. a phase mask, instead of an amplitude splitter, is used because, besides its economical factor, the alignment process is simplified. the wavelength of the designed grating is controlled by the incidence angle of the two diffracted orders on the fiber, which can be controlled by the lateral mirrors [17]. c. fbg’s behavior due to external disturbances the effective refraction index and the period modulation, both mentioned in equation (2.1), change with temperature and strain, thus changing the bragg wavelength. the change in the reflected wavelength when exposed to variations in temperature, ∆𝑇, and/or mechanical deformations, ∆𝑙, is given by equation 2.7 [13]: ∆𝜆𝐵 = 2 (λ 𝜕𝑛𝑒𝑓𝑓 𝜕𝑇 + 𝑛𝑒𝑓𝑓 𝜕λ 𝜕𝑇 ) ∆𝑇 + 2 (λ 𝜕𝑛𝑒𝑓𝑓 𝜕𝑙 + 𝑛𝑒𝑓𝑓 𝜕λ 𝜕𝑙 ) ∆𝑙 (2.7) an applied longitudinal deformation changes the λ parameter due to the increasing pitch of grating and changes the 𝑛𝑒𝑓𝑓 parameter, because of the photoelastic effect. the latter comes from an observation made when compressing a transparent material, in which two effects can be observed, one is the increase of the refractive index due to the increase of density of the material and the other is the mentioned photoelastic effect, which produces the opposite effect. in an equal manner, both parameters from the fundamental bragg condition can be changed due to a variation in temperature, which can occur via thermal dilation, which influence the λ parameter, and via thermo-optic effect, which influence the 𝑛𝑒𝑓𝑓 parameter. the first term of equation 2.7 shows the effect of temperature in the reflected bragg wavelength and the second one denotes the effect of the mechanical strain. so, if only change in temperature is considered (∆𝑙 = 0), one gets [17]: ∆𝜆𝐵 = 𝑆𝑇∆𝑇 = 𝜆𝐵 (𝛼λ + 𝛼n) (2.8) where 𝑆𝑇 is the thermal sensitivity of the fbg. the 𝛼λ parameter is the thermal expansion coefficient of the fiber (~0.55 × 10−6℃−1 for silica), and 𝛼n is the thermo-optic coefficient (~8.6 × 10−6℃−1 for germanium doped silicacore fiber). a thermal sensitivity approximately equal to 13 pm/ºc is expected for fbgs working at the spectral region of 1550 nm [13, 17]. when neglecting the thermal disturbances (∆𝑇 = 0), the effect of mechanical disturbances in the bragg wavelength is described by equation 2.9 [17]: ∆𝜆𝐵 = 𝑆∆𝑙 𝑧 = 𝜆𝐵 (1 − 𝑝𝑒 ) 𝑧 (2.9) where 𝑆∆𝑙 is the sensitivity to the strain felt over the longitudinal axis, and 𝑧 is the relative strain over the longitudinal axis. when the fbg suffers a contraction the 𝑧 value is negative and when it is expanded this value turns positive. by integrating the intrinsic interference between the reflected spectra of the fbg and the fabry-perot interferometer (fpi) with respect to the wavelength at each time instant, one gets an optical power variation. this process is called edge filtering, where the fpi acts an optical power discriminator. d. modulation techniques for the proposed aom system, several modulation techniques can be tested as candidates for a successful transmission of the encoded data that comes from the iot devices, namely ook, amplitude-shift keying (ask) and frequency-shift keying (fsk). phase related modulations were not considered since the existent system lacks the phase stability required to perform carrier-phase estimation in order to properly modulate the data [20]. the actual tested modulation techniques were the ook and the fsk. in the context of this system, the ook modulation technique is used to modulate the optical subcarrier with respect to the frequency, whereas the fsk method is used to modulate an electrical subcarrier with respect to its frequency, which therefore modulates the optical subcarrier in the t. ribeiro | i-etc iot 2018, vol. 4, n. 1 (2018) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_17 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_18 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_18 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_18 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_17 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_19 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_17 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_13 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_17 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_13 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_17 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_17 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_20 frequency domain. e. on-off keying modulation the transmission of a binary sequence can be done with use of on-off signals, i.e., when the transmitted bit is 1, the transmitted signal waveform is 𝑠1(𝑡) = 𝑠(𝑡), and when the transmitted bit is 0, the signal waveform becomes 𝑠0(𝑡) = −𝑠(𝑡), where 𝑠(𝑡) = 𝐴, 0 ≤ 𝑡 ≤ 𝑇𝑏 . these types of signals are known as antipodal signals, since one signal waveform is the negative of the other. thus, the received signal waveform, after going through a noisy channel, may be represented as: 𝑟(𝑡) = 𝑠(𝑡) + 𝑛(𝑡)− + , 0 ≤ 𝑡 ≤ 𝑇𝑏 (2.10) where 𝑛(𝑡) represents the additive white gaussian noise. the optimum receiver consists of a correlator whose output is sampled at 𝑡 = 𝑇𝑏 , and followed by a detector that compares the sampled output with a certain threshold valued as 𝛼. for signal waveforms with equal probabilities, the optimum detector compares 𝑟 with a threshold 𝛼 = 0. if 𝑟 > 0 the detector decides that 𝑠(𝑡) was transmitted, whereas when 𝑟 < 0, the decision is made that −𝑠(𝑡) was transmitted [20]. f.frequency-shift keying with m-ary fsk, it is possible to transmit 𝑘 = log2 𝑀 bits per symbol with 𝑀 signal waveforms which can be expressed in equation 2.11: �̂�𝑚(𝑡) = √ 2𝐸𝑠 𝑇 cos(2𝜋(𝑓𝑐 + 𝑚𝛥𝑓)𝑡) (2.11) for 𝑚 = 0, 1, … , 𝑀 − 1 and 0 ≤ 𝑡 ≤ 𝑇, where 𝐸𝑠 = 𝑘𝐸𝑏 is the energy per symbol (being 𝐸𝑏 the signal energy per bit), 𝑇 = 𝑘𝑇𝑏 is the symbol interval (with 𝑇𝑏 corresponding to the duration of the bit interval), 𝛥𝑓 is the frequency separation and 𝑓𝑐 is the carrier frequency [20]. in real life conditions, the received signal is different from the transmitted one, i.e., when the modulated signal goes through a transmission channel, it gets delayed and it is affected by noise, therefore 𝑟𝑚 (𝑡) is expressed by: 𝑟𝑚(𝑡) = √ 2𝐸𝑠 𝑇 cos(2𝜋(𝑓𝑐 + 𝑚𝛥𝑓)𝑡 + ∅(𝑡)) + 𝑛(𝑡) (2.12) where, ∅(𝑡) represents the phase shift of the mth signal and 𝑛(𝑡) is the additive bandpass noise. when the signals {𝑟𝑚𝑐 , 𝑟𝑚𝑠 }𝑚=0 𝑀−1 have the same probability, the signal envelopes can be computed by the square law detector 𝑟𝑚 2: 𝑟𝑚 2 = 𝑟𝑚𝑐 2 + 𝑟𝑚𝑠 2 (2.13) which will then select the signal corresponding to the largest {𝑟𝑚 2 } value. figure 4 depicts the schematic of the discussed process of detection and demodulation for an m-ary fsk received signal, where the basis functions of the correlators feature the known phase shift for each symbol when equalization is not used. this was the approach used in the simulations described in section 5. detector received signal sample at . . . output decision sample at sample at sample at sample at sample at fig.4. schematic of the m-ary fsk detection and demodulation process (adapted from [20]). g.probability of error for m-ary orthogonal signals for m-ary orthogonal signals, the probability of error, assuming the symmetry of the signal space and equal probabilities for all 𝑀 symbols, can be simplified so that only two parameters affect the overall probability, namely 𝑀, which is the number of symbols, and 𝐸/𝑁0, which is the signal-tonoise ratio. 𝑃𝑒 = 1 √2𝜋 ∫ {1 − [1 − 𝑄(𝑦)]𝑀−1}𝑒 −(𝑦−√2𝐸/𝑁0) 2 /2𝑑𝑦 ∞ −∞ (2.14) with 𝑄(𝑦) = 1 2𝑒𝑟𝑓𝑐(𝑦/√2)⁄ , 𝑀 representing the total of symbols for a certain m-ary fsk modulation schemes, 𝐸 being the symbol energy and 𝑁0 = 𝜎 2/2, with 𝜎 2 being the noise variance [20]. h.estimation of the ber from the evm error vector magnitude (evm) is proven to be an appropriate metric for optical channels limited by awgn. it can be described as the effective distance of the received complex symbol from its ideal position in the constellation diagram. in mathematical terms, a received signal vector 𝐸𝑟 deviates from an ideal transmitted vector 𝐸𝑡 by an error vector 𝐸𝑒𝑟𝑟 . 𝐸𝑉𝑀𝑎 2 = 𝑀 𝐼 ∑ |𝐸𝑟,𝑖 − 𝐸𝑡,𝑖| 2𝐼 𝑖=1 ∑ |𝐸𝑡,𝑖| 2𝑀 𝑖=1 (2.15) this last parameter is used for the estimation of the ber, whose approximate mathematical expression is shown in equation 2.15, with 𝐿 being defined as the number of signal levels in each dimension of the constellation, which, in the case of the m-ary fsk modulation, is equal to the number of the m symbols. i is the number of randomly transmitted data [22]. 𝐵𝐸𝑅 ≈ (1 − 𝐿−1) log2 𝐿 𝑒𝑟𝑓𝑐 [√ 3 log2 𝐿 (𝐿2 − 1)𝐸𝑉𝑀𝑎 2 log2 𝑀 ] (2.16) iii. aom for the iot a detailed schematic of the iotof concept can be seen in figure 5, featuring the five main blocks or stages: the receiver stage, the demodulation stage, the iot devices, the codification-modulation stage and the aoms. the multiplexing capability of the fbgs and the inherent low optical fiber attenuation, provides an advantage in a sense that it is possible to distribute several aoms along the same optical t. ribeiro | i-etc iot 2018, vol. 4, n. 1 (2018) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_20 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_20 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_20 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_20 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_22 fiber, with each aom having its own bragg wavelength associated with (𝜆1, 𝜆2, …, 𝜆𝑛). the bbs feeds the aoms, since the fbgs included in these modulators are completely passive elements, which are mechanically driven by acoustic signals containing the iot information. then, an optical circulator redirects the reflected spectra of the fbgs to the several fpis, located at the receiver stage, which act as linear edge filters. this allows for the straightforward translation from the spectral shifts from the several existent fbgs, into optical power variations, which are then converted into the electrical domain by the photodetectors, which are also located at the same receiver stage. internet n channels adc broadband source circulator cmd cmd cmd optical fiber optical fiber optical fiber aom 1 aom 2 aom n linear edge filter pd1 pd2 pdnop-amp usb gateway modulated signals n demux optical excitation stage receiver stage codification modulation stage iot devices optical path electrical path op-amp op-amp data receiver fig.5. schematic of the iot2of approach. the electrical signals are amplified by the operational amplifiers (op-amps) in transimpedance configuration and posteriorly acquired by an adc module. the data receiver is able to process the received signals and demodulate them, which makes possible the sharing of this information with the internet. however, for this whole process to occur, it is necessary the use of codification-modulation devices (cmds) which are responsible for the digital modulation and generation of the electrical signals that excite the aoms, since the aoms are not able to read the information coming from the iot devices directly. also, the previous schematic can be translated in terms of a phy functional sub-layers scheme, as depicted in figure 6, with the cmd stage corresponding to the information-toelectrical (i/e) phase, the electrical-to-acoustic (e/a) phase conversion being the induced mechanical vibrations on the fbg from the speaker, followed by an acoustic-to-optical (a/o) conversion, in order for the optical channel to transmit the modulated information through long distances to finally reach the optical receiver, where the received signal undergoes an o/e conversion, done by the photodetectors. finally, there is the electrical-to-digital (e/d) conversion performed by the adc and the digital-to-information (d/i) phase, which contains the demodulation and decision processes. i e e a a o i d d e e o optical acoustic electrical information phy sublayers optical channel acoustic channel cmd iot (uplink) input bits gateway output bits figure 6: flowchart of the phy sublayers communication. a. simulator implementation figure 7 depicts the block diagram of the implemented simulator that was done with matlab. its working principle can be explained in different phases, with each one having its own functional block: input electrical signal – the signal that is injected in the system, i.e., the ook modulated signal or the m-ary fsk modulated signal that contains the information from the first phy sublayer, depicted in figure 5, in which is applied gray codification (cmd stage). this type of codification was used in order to reduce the impact of symbol error rate over bit error statistics. the electrical signals applied are converted into acoustic signals because of the transducer of the speaker. max. wl shift – contains the maximum wavelength shift induced in the fbg, which in this case is around 80 pm. this value will be explained in section 4. system response filter – contains the transfer function of the aom system, with a bandwidth of 1 khz and a similar shape to the one from figure 9 (a). initially there is an imposed delay of around 57 ms in the received signal, which is due to the fact that a finite impulse response (fir) digital filter was implemented, with order equal to 5000, using the leastsquares approach, for a sampling rate of 44 khz, so that a considerably large number of samples per symbol could be acquired. this type of filter has a linear phase response which in turn preserves the waveshape of the input signal, to the extent that is possible, since some frequencies will be changed in amplitude by the action of the filter. the system’s phase response was applied to the output signal of the system response filter block, using the fast fourier transform (fft) [23]. volt to wl – the amplitude of the electrical signal causes the excursion movement of the speaker’s diaphragm which will extend or contract the fbg, which in turn causes the bragg wavelength to change in each time instant. in other words, it is here where the a/o conversion takes place. wl to fbg – here the bragg wavelength obtained in the previous block is substituted in equation 2.6, with the other parameters being 𝑛𝑒𝑓𝑓 = 1.458, 𝐿 = 10 𝑚𝑚 and 𝛿𝑛𝑒𝑓𝑓 = 10−4. this block calculates the power spectral density spectrum reflected by the fbg for each time instant. attenuator – the attenuations in the optical channel from both the optical circulator and the voa are applied. fpi – contains the transfer function depicted in the upcoming figure 4.9 (iii). for each time instant, this transfer function will apply an optical power variation to the fbg’s reflected spectrum throughout time. spectral density to power – in this block, the integral of the file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_23 interference between the reflected fbg and fpi spectra, with respect to the wavelength, is applied, i.e., the resulting power signal of the interference between the fbg and the fpi spectra is computed. awgn channel – this block serves as way to introduce noise originated from all the optical elements connected in the montage of the system, thus allowing for a study of its performance. gain – contains the responsivity of the photodetector, the gain from the op-amp and the resistance value of the adc. these are responsible for the o/e conversion. input electrical signal max. wl sh ft volt to wl attenuator spectral density to power output optical signal system response f lter wl to fbg awgn channel gain output electrical signal receiver fpi fig. 7. block diagram of the implemented simulator. the signal received in the output electrical signal block, which can be an ook or m-ary fsk modulated signal with or without equalization, is demodulated, i.e., the e/d conversion happens. finally, the retrieved digital signal is converted into its corresponding gray code. ii. system implementation a. fbg manufacture given the other acousto-optic modulation systems, in this study a low-cost aom system based on the usage of an fbg and a commercial speaker with a membrane diameter of 48 mm is proposed. a photosensitive smf (thorlabs gf1b) was used to record the fbg with the phase mask technique by using an excimer laser emitting at 248 nm in an fbg inducing system built at instituto de telecomunicações – pólo de aveiro (it). the operation principle of the proposed aom is based on the longitudinal movement of the speaker diaphragm (excursion), as depicted in figure 8, which imposes a contraction ( < 0) and an expansion ( > 0) in the fbg when the speaker is exposed to a certain applied electrical signal. when launching an optical signal from a bbs in the fiber, a blueshift or redshift is observed in the fbg reflected spectrum. the bragg wavelength is thus changed as a function of the external electrical signal that is fed to the speaker. fig. 8. schematic of the fbg operation principle. b. system characterization for the characterization of both frequency and temporal responses of the proposed system, two analyses were performed. the system response was analyzed by using a 𝐴sin(𝜔𝑡) signal generated by the awg, where 𝜔 = 2𝜋𝑓𝑐. the amplitude (𝐴) of the signal was maintained constant ( 1− + v) as well as the volume in the speaker. then, the frequency (𝑓𝑐) was swept from 10 hz to 1000 hz in steps of 10 hz, with each step lasting a total of 2 seconds, approximately. after obtaining the information regarding the amplitude and phase delay between the reference and the modulated signal, it is possible to determine the transfer function of the system and it is depicted in figure 9. figure 9 (a) shows the amplitude transfer function of the system, in which the maximum optical amplitude variation is set between 550 hz and 650 hz, where the resonance frequency is evidenced around 570 hz. it can also be observed that for extreme frequencies, especially for values lower than 50 hz and higher than 900 hz, the amplitude is attenuated in more than, approximately, 45 db, therefore, the effective bandwidth of this solution is around 800 hz. (a) fig. 9. amplitude response of the aom system. figure 9 (b) depicts the phase response of the opticalacoustic modulation system, again with the points representing the measured phase delays and line being the estimated phase delays for the in-between frequencies. the minimum phase delay is achieved at around 200 hz, with an observed linear phase increment from this same frequency up to 500 hz. finally, the delay is drastically increased for frequencies higher than 500 hz, i.e., near the resonance frequency. figure 10 depicts the resulting electrical amplitude of the output signal when the input is a sinusoidal of a certain frequency and unitary amplitude. using the mentioned spectrometer, the bragg wavelength shift (red solid line) is obtained for every obtained electrical signal (dashed blue line), when the input voltage signal has a frequency that ranges from 10 hz to around 300 hz. fig. 10. wavelength shift versus output amplitude. by subtracting the obtained wavelengths for each frequency values from the pre-stressed fbg bragg wavelength (𝜆𝐵 = 1547.5 nm) and multiplying them by a factor of 2, the resultant shifts from the lowest to the highest bragg wavelength are obtained. from figure 10, one can already deduce, assuming a linear relation between both curves, a relation of around 2 pmpp/mvpp, i.e., there is linear relation between wavelength displacement and the amplitude of the electrical signal received. knowing that the maximum output voltage, for this implementation, is around 80 mvpp, a maximum wavelength shift of 160 pmpp is obtained, i.e., 80 pm from the central bragg wavelength to the maximum displacement. another characteristic of the system that was analyzed was the pulse response, in which it is possible to observe that for pulse periods lower than 2.5 ms, approximately, there is an increase in the underdamped response, thus the resonance frequency is predominant in the underdamped temporal response. iv. simulation results mainly two modulation approaches will be used in the simulations. the first one is the direct fsk modulation, in which the optical carrier is modulated directly, i.e., when the amplitude of the signal is highest, so is the wavelength displacement in the fiber. the second one is the most enticing one, which is to have an fsk subcarrier to modulate the optical one, which will allow for the imprinting of mechanical displacement into frequency modulation. this last was proven to be the most effective one in the experimental work done and it will serve as base for the dimensioning of the implemented simulator’s awgn’s noise variance. in order to perform the simulations, the noise variance was estimated so that, for a transmission line distance of 30 km, a bit error rate (ber) around 0.2% would be measured for the 8-fsk modulation scheme with complex equalization for a total of 10.500 bits. an optical attenuation ranging from 0 to 20 db, for a 2 db iteration, was applied, using the bit comparison method. figure 5.1 depicts the 8, 4 and 2-fsk ber curves for a dimensioned noise variance equal to, approximately, 3.31 × 10-20 w. v. performance assessment for ook modulation with the dimensioned noise variance for the awgn channel, now it is possible to perform simulations for the assessment of the system’s performance. first, the simplest modulation was applied, the ook one. a binary sequence serves as the input signal, with the -1 v amplitude being encoded with the bit 0 and the +1 v amplitude with the bit 1, so that the rejection in the output signal between the zero level and the one level is as high as possible. a total of 10.500 bits were considered, with the signal duration varying according to the bit rates. the non-equalized received electrical signal is depicted in figure 11 (a), which was obtained by simply injecting the system with, in this case, the ook modulated signal. it can be observed that the isi is quite predominant, since the eye diagram is quite obstructed. the equalization was done by applying to the modulation carrier the inverse transfer function of the complex transfer function of the system, which is depicted in figure 8. it serves as much more promising modulation for this system, since it reduces the isi significantly, thus the eye is more opened (figure 10 (b)). fig. 11. eye diagram for the equalized 1000 bps system. complex equalization for the ook modulation scheme can be assessed in relation to its performance in an awgn channel with the ber estimation. given the fact that a sequence of randomly transmitted data is generated, i.e., a stochastic process occurs in each simulation that is done, several simulations were done to get a better understanding of the system’s performance. figure 12 shows the mean ber curves for three simulations done for each bit rate. by applying the previously mentioned process of computing the distances between the detector output values and their mean values for each optical power at the receiver’s input. the dashed black line is the maximum tolerable pre-forward (a) (b) (b) error correction (fec) ber, also known as the “fecthreshold”, which, in the context of optical transport networks, is the associated limit ber value for which, when the system performs at a ber lower than this, it is considered as error-free [24]. fig. 12. eye diagram for the equalized 1000 bps system. one can observe that, for example, for a power of -49.24 dbm, only the 200 bps equalized ook modulation can be considered as error-free, with a mean ber value between 0.2% to 0.3%. the other three cases, 800, 1000 and 400 bps, have higher ber values than the pre-fec ber threshold, with mean ber values of approximately 0.4%, 0.5% and 0.7%, respectively. less samples are taken for higher bit rates, with the lowest number of samples being for a bit rate of 1000 bps, whereas in the case of 200 bps, a total of 220 samples are taken. this might influence the obtained results because of the aliasing effect, since there can be an insufficient number of samples to represent each bit. thus, for higher bit rates, the results are less reliable. for optical powers higher than -48 dbm, approximately, all the tested bit rates were proven to be successful in the transmission of information and offer data throughputs that are much higher than the one achieved by the sigfox protocol. vi. performance assessment for m-ary fsk modulation also, taking the equations for the ber estimative from the evm into account, a comparison between the equalized m-ary fsk modulated schemes can be made. the 8-fsk modulated system has the highest ber for the same received optical power, and the ber for 4-fsk is lower than the previous one but higher than the one from 2-fsk modulated system. these results were obtained for simulation using a total of 10.500 bits over different durations, varying according to the different bit rates, with the ber curves being the mean values for three simulations, in the case of 2-fsk, another three simulations for 4-fsk, and finally, seven simulations for 8fsk. while, for a received optical power of approximately -49.24 dbm, the 2-fsk and 4-fsk systems have a much lower ber than the one from 8-fsk, which is around 4%, i.e., it is higher than the pre-fec ber value. thus, one can deduce that for the same received optical power at the receiver’s input, the 8-fsk equalized system is much more affected in the transmission link than the others. next, it will be discussed how the equalized system performs in comparison with the nonequalized one, with ber estimation from evm method for the same received optical power. the results for this particular case are unexpected, since the symbol probabilities are equal to one another and this is a case of orthogonal symbols, one can use the probability of error expression, as depicted in equation 2.14, to study this case. it is important to notice that for different energy symbols there is no symmetry of the signal space, but for the stake of simplicity and since this is a case of study that is not in the scope of this study, one can make this assumption. the fact that the 2-fsk equalized system has an overall higher ber than the non-equalized one, may be due to the fact that the latter has a higher 𝐸/𝑁0 value than the former. also, for lower noise variance, i.e., for a lower 𝑁0 a result where the equalized system performs in a better way could have been achieved. for a measured received optical power of -53.24 dbm, one estimates a ber of 0.3% for the equalized system and 0.06% for the non-equalized one, which is lower. but at the same time, both cases can be considered error-free for that same received power. whereas for lower powers both cases have ber values higher than pre-fec ber threshold. in the case of the 4-fsk, the same unexpected conclusion of the 2-fsk modulated system was obtained, although with a much more significant difference. for a received optical power of -51.24 dbm, the ber for the equalized system is around 2%, whereas for the non-equalized on the ber is around 0.06%, which can be considered errorfree and it has significantly better performance overall than the equalized scenario. again, the fact that the 4-fsk equalized system has an overall higher ber than the non-equalized one, can be explained with the same analogy used in the 2-fsk scenario. the higher 𝐸/𝑁0 value is due to symbol 3 having the highest energy of all four symbols. regarding fsk modulation results, the most promising modulation scheme is the 8-fsk, providing a bit rate of 300 bps, and where a significant difference can be seen when comparing the equalized and non-equalized solutions. a discrepancy between an equalized and a non-equalized 8-fsk system is not unexpected, since the non-equalized solution will induce a significant isi, especially in the transitions from a high frequency to a low one, as it can be seen in figure 13 (b), being the most evident one the transition from 700 hz to 100 hz, in the 0.32 to 0.33 s time interval (marked with a red circle). in the transition from 700hz to 100 hz, the 100 hz wave has a superimposed harmonic component of the natural frequency. for strong oscillation changes, the underdamped response requires a period of around 12 ms for it to stabilize, which is higher than the minimum period of symbol. this can be considered as isi and, in case of a transition from a lower to a higher frequency, i.e., neighbor subcarriers, the wave shape is not going to be affected. when complex equalization is applied, this isi is removed by applying the inverse transfer function of the system’s complex response, thus the wave shape of the signal will not be affected overall. for the example depicted in figure 14 (b), this fact be observed, where there is no longer wave shape distortion in the (a) file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_24 800 hz to 100 hz transition (marked with a green circle). fig. 13. non-equalized 8-fsk received signal (b) as a function of the transmitted signal (a). fig. 14. equalized 8-fsk received signal (b) as a function of the transmitted signal (a). figure 15 depicts the spectra for both non-equalized, on the left side, and equalized, on the right side, 8-fsk systems, where, as expected, the equalized one has the frequencies nearly at the same amplitude, whereas in the case of nonequalization, the 600 hz, which is near the resonance one, has the biggest amplitude. figure 15: spectra for both non-equalized (a) and the equalized (b) 8-fsk system. but confirmation on the advantages of equalization for the 8fsk solution must obtained with the evm measure technique. figure 16 depicts the comparison between the use of equalization and non-equalization, zoomed in for ber values over 0.1%, because the equalized system has much lower ber values for received optical powers between -47 dbm and 39.24 dbm, approximately. the red curve is the mean ber curve for the non-equalized system, for a total of 30 simulations. fig. 16. comparison between equalized and non-equalized 8-fsk. contrary to previous simulations, the measured ber values showed significant deviations from the mean values, so a greater number of simulations needed to be executed, in order to have a better understanding of the overall system performance. thus, in this case, the error bars were placed in the measured points. when inspecting the detector output values for a signal power equal to -45.24 dbm, approximately, for the non-equalized situation, symbol 5 has the greatest mean value, which leads to serious consequences for the system’s performance. by thoroughly analyzing all the results regarding m-ary fsk modulation schemes, equalized 8-fsk seemed like the most promising one for the iotof, with an estimated ber always lower than the pre-fec threshold until an optical power around -47 dbm, whereas the non-equalized modulation shows much higher ber values for higher optical powers. with this solution, the highest throughput in m-ary fsk modulation was achieved, being also higher than the bit rate of the sigfox standard and, also, with a higher range than lora. the ook modulation scheme with complex equalization can also be used for this application, although for an optical power received between -50 dbm and -48 dbm, for bit rates of 400 bps, 800 bps and 1000 bps, this solution does not seem feasible. the best results can be observed when a bit rate of 200 bps is used, which does improve the data throughput that the sigfox protocol achieves, but has a worse performance than the 4-fsk modulated system, which already is capable of transmitting data at 200 bps. isi has a significant impact in this system for this scheme, causing the non-equalized ook solution to be far from ideal. vii. laboratorial results in collaboration with instituto de telecomunicações – pólo de aveiro, a laboratorial work was done for an arbitrary binary sequence of 3000 bits, with gray codification, with preemphasis equalization for 2, 4 and 8-fsk modulation schemes. the laboratorial setup of the work done can be viewed in figure 17, which is depicted below. (b) (b) (a) (b) (a) (a) fig. 17. photograph of the laboratorial setup. the system performance was tested as a function of the attenuation applied (0.2 db/km), with each iteration representing a transmission distance of 5 km. the ber as a function of the transmission distance was analyzed for the several modulation schemes used, with the results being depicted in figure 18. for the highest bit rate (300 bps), the system is less tolerant to the link attenuation. a comparison was made with a receiver sensitivity of 1%, considered in narrowband wireless systems. for this specific condition, this system offers a maximum transmission distance of around 30 km and an additional distance of 5 km can be added to this for the 4-fsk (200 bps) and 2-fsk (100 bps) [25, 26]. fig. 18. ber as a function of the transmission distance. viii. conclusions and future work for ook modulation, the system already reaches a data throughput higher than the sigfox standard, but has an overall worse performance than the 4-fsk modulated system, in the case of the 200 bps bit rate. regarding the m-fsk modulation schemes, the optimum solution found in the performance assessment is the equalized 8-fsk system, reaching a data throughput of 300 bps, which is higher than sigfox, and being error-free until a received optical power of around -47 dbm. it is important to notice that the acoustic channel can be improved upon by using more elaborate e/a structures to enhance the system’s response. the use of a forward error correction and a more powerful bbs would extend the reach of this system beyond what was proven to be possible. other modulation techniques, such as quadrature amplitude modulation (qam) and coding strategies might also improve the aom system’s spectral efficiency, thus improving the enduser bit rate. the proposed iotof technology will need a proper short framing link layer solution over iotof phy. also, another important aspect is the aom’s power consumption, which will consequently influence the iotof’s device autonomy. distributed raman amplification can be used to extend the transmission’s reach, while its pump power recycling is used for power transmission over the optical fiber, which could be used for feeding iotof devices [27]. references 1. kilper, d., et al., optical networks come of age, optics and photonics news, vol. 25, no. 9, pp. 50-57, 2014. 2. winzer, p.j., scaling optical fiber networks: challenges and solutions. optics and photonics news, vol. 26, no. 3, pp. 28-35, 2015. 3. madakam, s., r. ramaswamy, and s. tripathi, internet of things (iot): a literature review, journal of computer and communications,vol. 3, no.5, pp. 164, 2015. 4. ji, p.n. and w. ting. internet of things with optical connectivity, networking, and beyond, in 2016 21st optoelectronics and communications conference (oecc) held jointly with 2016 international conference on photonics in switching (ps), niigata, jul. 2016. 5. alkhatib, h., et al., ieee computer society 2022 report. recovered by: http://www.computer.org/cms/computer.org/computingnow/ 2022report.pdf, 2014. 6. vetter, p., et al., energy-efficiency improvements for optical access. ieee communications magazine, vol. 52, no. 4, pp. 136-144, 2014. 7. larsen, c.p., a. gavler, and k. wang. comparison of active and passive optical access networks. in telecommunications internet and media techno economics (ctte), 9th conference, ghent, belgium: ieee, jun. 2010. 8. grobe, k., et al., cost and energy consumption analysis of advanced wdm-pons. ieee communications magazine, vol. 49, no. 2, 2011. 9. senior, j.m. and m.y. jamro, optical fiber communications: principles and practice. pearson education. 2009 10. reiter, g., wireless connectivity for the internet of things. europe, vol.433, pp. 868mhz. 2014. 11. winder, s. and j. carr, newnes radio and rf engineering pocket book, newnes, 2002. 12. skubic, b., et al., energy-efficient next-generation optical access networks. ieee communications magazine, vol. 50, no. 1, 2012. 13. chen, q. and p. lu, fiber bragg gratings and their applications as temperature and humidity sensors, atomic, molecular and optical physics, pp. 235-260, 2008 14. torres, p., l.c.g. valente, and m.c.r. carvalho, security system for optical communication signals with fiber bragg gratings, ieee transactions on microwave theory and techniques, vol. 50, no. 1, pp. 13-16, 2002. 15. zhang, h. and t. dong. fiber bragg grating sensor system in the application of guarding against burglary, in 2012 second international conference on intelligent system design and engineering application. 6-7 january, sanya, hainan, china, 2012. 16. erdogan, t., fiber grating spectra. journal of lightwave technology, vol. 15, no. 8, pp. 1277-1294, 1997. 17. antunes, p.f.d.c., sensores ópticos para monitorização dinâmica de estruturas, in departamento de física., universidade de aveiro, 2011. 18. hill, k.o. and g. meltz, fiber bragg grating technology fundamentals and overview. journal of lightwave technology, vol. 15, no. 8, pp. 1263-1276, 1997. file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_25 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_26 file:///c:/users/ana%20teresa%20branco/dropbox/i%20love%20you.%20i%20know/extended%20abstract.docx%23_enref_24 http://www.computer.org/cms/computer.org/computingnow/2022report.pdf http://www.computer.org/cms/computer.org/computingnow/2022report.pdf 19. othonos, a., et al., fibre bragg gratings, in wavelength filters in fibre optics, springer. pp. 189-269. 2006 20. proakis, j., m. salehi, and g. bauch, contemporary communication systems using matlab. nelson education, 2012 21. hughes, l.w., a simple upper bound on the error probability for orthogonal signals in white noise. ieee transactions on communications, vol. 40, no. 4, pp. 670, 1992. 22. schmogrow, r., et al., error vector magnitude as a performance measure for advanced modulation formats. ieee photonics technology letters, vol. 24, no. 1, pp. 61-63, 2012. 23. olson, j.d. finite impulse response filters. in biomedical digital signal processing, prentice-hall, inc, 1993. 24. cho, j., c. xie, and p.j. winzer, analysis of softdecision fec on non-awgn channels, optics express, vol. 20, no. 7, pp. 7915-7928, 2012. 25. texas instruments, cc1120 high-performance rf transceiver for narrowband systems. datasheet available online at: http://www.ti.com/lit/ds/symlink/cc1120.pdf (accessed on june 23 2012), 2015. 26. hayashi, g., et al. a 10.8 ma single chip transceiver for 430mhz narrowband systems in 0.15/spl mu/m cmos. in solid-state circuits conference, 2006. digest of technical papers. ieee international. 6-9 february, 2006 ieee. 27. andré, p., et al., raman amplified access networks with pump signal recycling for electrical power conversion. microwave and optical technology letters, vol. 54, no. 1, pp. 116-119, 2012. http://www.ti.com/lit/ds/symlink/cc1120.pdf design of the alamouti scheme for a mimo receiver and its implementation on an fpga design of the alamouti scheme for a mimo receiver and its implementation on an fpga jorge silva instituto superior de engenharia de lisboa (isel), área departamental de engenharia electrónica e telecomunicações e de computadores (adeetc), lisboa, portugal 29482@alunos.isel.pt pedro pinho instituto superior de engenharia de lisboa (isel), área departamental de engenharia electrónica e telecomunicações e de computadores (adeetc), lisboa, portugal instituto de telecomunicações campus de santiago, aveiro, portugal ppinho@deetc.isel.pt mário véstias instituto superior de engenharia de lisboa (isel), área departamental de engenharia electrónica e telecomunicações e de computadores (adeetc), lisboa, portugal inesc-id, libsoa, portugal mvestias@deetc.isel.pt keywords: alamouti, mimo receiver, fpga. abstract: this paper analyses the alamouti scheme for different antenna configurations and different modulation types, namely bpsk, qpsk and qam. all configurations were modeled and simulated in matlab. a mimo receiver for a 2×1 antenna configuration and bpsk modulation was implemented in a fpga. the fpga results indicate that the alamouti scheme is a good design option for hardware implementation of a mimo receiver. the receiver uses only about 10% of the resources of a medium-sized fpga and achieves almost 300 msymbols per second. 1 introduction currently, there is an increasing offer of services based on wireless communications, which require high speeds to ensure the quality of service without compromising the transmission power of the system or the bandwidth. mimo (multiple input multiple output) technology offers increased capacity to these systems without requiring an increase in bandwidth or transmitted power. this technology is based on the channel space exploration using multiple antennas at both ends of the communications system, offering various gains: beam forming, array gain, spatial diversity gain, spatial multiplexing gain and interference reduction. the spatial diversity technique is the most efficient and effective to combat the effects of multipath fading, achieving in this way an improved system reliability. spatial diversity gain arises from the use of space-time coding of the transmitted/received signals, where time, which is the natural dimension of digital communication systems, is complemented with the spatial dimension, from the use of multiple antennas on the part of the system. there are several space-time encoding schemes [2], as for example: space-time block codes (stbc), space-time trellis codes (sttc), space-time turbo trellis codes (stttc) and layered space-time codes (lstc). in stbc it is possible to reach a maximum order of spatial diversity (equals the number of transmit antennas, mt ) using a simple and linear processing on the decoding. a well-known type of stbc scheme is the alamouti scheme [1]. the paper is organized as follows. in section 2, the alamouti scheme is briefly described. in section 3, we describe the implementation of a mimo receiver based on the alamouti scheme on an fpga. section 4 reports the results of the implementation in terms of area and performance. finally, section 5 concludes the paper and proposes some future improvements. i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-6 http://journals.isel.pt/index.php/iajetc 2 alamouti scheme the original alamouti scheme [1] is a transmit diversity technique using two transmitting antennas and one receiving antenna, and can be divided into three parts: (1) the encoding and transmission sequence of information symbols at the transmitter, where the bits of information are encoded into symbols in the domains of space and time and then transmitted, one for each transmit antenna; (2) the scheme for combining the received signals at the receiver, which prepares the symbols for the maximum likelihood detector and (3) the decision rule for maximum likelihood detection for the sent symbols. figure 1 shows the alamouti scheme for the 2×1 antenna configuration. in this transmitting scheme, two signals are simultaneously transmitted from both antennas (antenna one and two) in two steps. in the first step, signals s1 and s2 are transmitted from antennas one and two, respectively. then, in the second step antenna one transmits −s2∗ and antenna two transmits s∗1 (the mark * stands for the conjugate of the symbol) (see figure 1). at the receiver side, two signals are received, r0 and r1, from antenna one and antenna two, respectivelly. these signals are expressed as: r0 = h0s0 + h1s1 + n0 (1) r1 =−h0s∗1 + h1s ∗ 0 + n1 (2) where h0 and h1 are the channel responses between each of the transmiting antenna and the receiving antenna. n0 and n1 represent complex noise and interference of the channel. estimations of the sent signals, s̃0 and s̃1, are then determined from the received signals and also using the channel responses, h0 and h1, as follows: s̃0 = (|h0|2 +|h1|2)s0 + h∗0n0 + h ∗ 1n1 (3) s̃1 = (|h0|2 +|h1|2)s1 −h0n∗1 + h ∗ 1n0 (4) this operation is implemented at the receiver by a block designated combiner. finally, the sent bits are obtained from the received symbols using a demodulator (the demodulator block in figure 1). the alamouti scheme can be generalized to the case of two transmitting antennas and nr receiving antennas, achieving in this way a diversity gain of 2× nr. for example, in the 2×2 antenna configuration the received signals, r0, r1, r2 and r3, are expressed as: r0 = h0s0 + h1s1 + n0 (5) r1 =−h0s∗1 + h1s ∗ 0 + n1 (6) r2 = h2s0 + h3s1 + n2 (7) r3 =−h2s∗1 + h3s ∗ 0 + n3 (8) where h0, h1, h2 and h3 are the channel responses between each of the transmiting antennas and receiving antennas. the combiner would then calculate an estimative of the sent symbols from these signals as follows: s̃0 = (|h0|2 +|h1|2 +|h2|2 +|h3|2)s0 +h∗0n0 + h1n ∗ 1 + h ∗ 2n2 + h3n ∗ 3 s̃1 = (|h0|2 +|h1|2 +|h2|2 +|h3|2)s1 −h0n∗1 + h ∗ 1n0 −h2n ∗ 3 + h ∗ 3n2 (9) the alamouti scheme for different antenna configurations was implemented in matlab in order to validate their theoretical concepts. figure 2 presents the alamouti scheme for multiple antenna configurations for bpsk modulation.   figure 2: alamouti scheme for various antenna configurations using bpsk. an important aspect is the difference in power between the configuration of one transmitting antenna and two receiving antennas (1×2) to two transmitting antennas and one receiving antenna (2×1). considering that the transmitted power is normalized to 1, the power transmitted by each transmit antenna in the case of the 2×1 is half the power transmitted by the antenna of the 1×2 case. this means that there is then a 3 db difference between the 1×2 and 2×1 configurations. the increase in the number of receiving antennas also makes the order of diversity increases, 2×nr. the alamouti scheme was also implemented with modulations qpsk and 16 qam. in figure 3 it can j. silva et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc bits [ ]10 ss space-time coder modulator [ ]*10 ss − [ ]*01 ss + 1 0 n n combiner r 0h 1 h ~ 0s ~ 1s demodulator 0h 1h alamouti transmitter bits alamouti receiver figure 1: alamouti scheme. be seen that modulations with higher levels require a better snr ratio, that is, for the same ber, a higher modulation has a higher snr.  figure 3: alamouti scheme with bpsk, qpsk and qam modulations. 3 fpga receiver for a 2×1 antenna configuration with bpsk in this section we desribe the fpga implementation of the alamouti receiver for a 2 × 1 antenna configuration and bpsk modulation (see figure 4). rx 1 0 n n combiner r 0h 1h ~ 0s ~ 1s demodulator bits [ ]10 , rr [ ]10 , ss figure 4: alamouti receiver scheme. from section 2 we know that the received signals in figure 4 are given by r0 = h0s0 + h1s1 + n0 (10) r1 =−h0s∗1 + h1s ∗ 0 + n1 (11) and the estimated symbols at the receiver are given by s̃0 = (|h0|2 +|h1|2)s0 + h∗0n0 + h ∗ 1n1 (12) s̃1 = (|h0|2 +|h1|2)s1 −h0n∗1 + h ∗ 1n0 (13) from equations (3) and (4), we see that s̃0 and s̃1 are the result of mathematical operations between complex numbers. the complex multiplication between two complex numbers, x and y, is given by: x×y = (a + bi)×(c + di) = (a×c−b×d)+(a×d + b×c)i based on equation (5) the block diagram of the complex multiplier is given in figure 5. reala imga realb imgb realout imgout figure 5: block diagram of a complex multiplier. this complex multiplier will serve as the basis for the construction of the combiner block (see figure 6). to recover the original transmitted bits from s̃0 and s̃1 is then necessary to carry out the demodulation of the received symbols. for the demodulation method we have used hard decision. using the hard decision process to perform the bpsk demodulation, it is enough to consider only the most significant bit (msb) of the real part of the complex symbol, that is, the signal bit. if the real part of the estimated symbol is positive, then the most significant bit of the symbol is a ’0’. on the other side, if the real part of the estimated symbol is negative, then the most significant j. silva et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc size lut dsp freq throughput 16 bits 2058 (6%) 8 (4%) 29 mhz 29 msymbols/s 23 bits 2986 (9%) 8 (4%) 24 mhz 24 msymbols/s table 1: fpga results for a 2×1 antenna configuration with bpsk without pipeline. size lut dsp latency freq throughput 16 bits 2268 (7%) 8(4%) 11 285 mhz 285 msymbols/s 23 bits 3230 (10%) 8(4%) 17 263 mhz 263 msymbols/s table 2: fpga results for a 2×1 antenna configuration with bpsk with pipeline. real r0 img r0 real r1 img r1 real h 0 img h0 real h1 img h1 complex multiplication complex multiplication complex multiplication complex multiplication realout realout realout realout realout realout 0 ~ s 1 ~ s 0 * 0 rh × 0 * 1 rh × * 11 rh × * 10 rh × figure 6: block diagram of the alamouti receber using bpsk modulation. bit is a ’1’. therefore, if the msb is ’0’, then the recovered bit is ’0’, otherwise is ’1’. in this implementation, all numbers were represented in two different custom floating-point formats: 16 bits (5 bits for the exponent and 11 bits for the mantissa) and 23 bits (6 bits for the exponent and 17 bits for the mantissa). 4 results the system was implemented in a virtex4-sx35 fpga with speed grade -12. two architectural implementatiosn were considered: with and without pipeline (see tables 1 and 2). the results show that the receiver uses less than 10% of the available resources of a medium sized fpga. wihout pipeline the receiver decodes upto 29 msymbols per second. the pipelined version increases the decode ratio to about 10 times more with only a marginal increase in the utilized resources. 5 conclusion in this work we have modulated and simulated in matlab the alamouti scheme for various antenna configurations and different modulations. the results are as expected and show the advantages of using this stbc scheme. the good throughputs achieved when implemented in an fpga even using floating point arithmetic and multiple antennas show that the method is a good design option for a mimo implementation using an stbc scheme. in the future, we will do a detailed study about the precision of the used arithmetic in order to improve the area and throughput of the hardware implementation. references [1] s. m. alamouti. a simple transmit diversity technique for wireless communications. ieee journal on selected areas in communications, 16:1451–1458, 1998. [2] m. jankiraman. space-time codes and mimo systems. artech house, boston-london: uk, 2004. j. silva et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc electroges a household iot energy management system electroges a household iot energy management system mariana bentoa, gonçalo rodriguesa, vitor vaz da silvaab, adepartment of engineering of electronics, telecommunication and computers, isel/ipl instituto superior de engenharia de lisboa, portugal bcts-uninova, fct-unl, portugal vsilva @ deetc.isel.ipl.pt abstract 1 — domestic appliances plugged to the electrical mains power may, at some point, exceed the maximum contracted power supplied from the electrical company. to avoid a power cut-off on these cases, an energy management system is needed so that some appliances are automatically switched off before others are switched on. in order to do this, appliances have to be described in a privileged structure so that power is always available for some of them while other share leftovers. elecroges is a system that provides the house owner with a means to define the priorities of the connected appliances and manage them in an autonomous way; a working implementation is provided. keywords: iot, energy management, home, domotics i. introduction the appliances found in a home are mainly electronic devices aimed at domestic usage and most of them are connected to the national, or private, energy supply network through an electrical plug. these devices present a consumption of energy over time defined as the power usage of the device while operational. in portugal, the electrical installations at each home, and the contracts with the electrical supply companies, limit the total domestic power consumption to a given value, which can lead to an energy cut, in the eventuality of all the connected devices consumption exceeds that limit [1]. in the portuguese market, it is already possible to have access to several energy control systems such as the smart plug ‘edimax sp-20101w’ with which it is possible to manage the energy consumption of each device trough an iphone or android mobile device. another example, specific for the portuguese market, is the ‘edp re:dy plug’ that transforms the connected appliances into smart ones and allows their management through a mobile device. as a final example, there is also a controller called ‘fator potência pfw01-m12’ which monitorizes eletrical measurements and can reduce or eliminate losses in an electrical system. the electroges system differs from the previous examples primarily because it joins the electrical consumption statistics with a configurable autonomous energy management system, without overloading the user. the objective of this project is to present a system which can reduce domestic power consumptions and avoid electrical cuts or failures due to excess, it also helps reducing the ecological impact and the financial cost of energy usage at home [2]. even though electroges presents a slow initial configuration process, once that is finished, the system does not require more user intervention, except to add new appliances or to change the desired settings. ii. architecture the electroges system is characterized by three main elements, these being the controller, the eg-plug (electroges plug) and the configuration app, all connected through the local wi-fi network [3][4] as shown in figure 1. figure 1 example diagram of the electroges system the controller is the central processing and coordination element in the system. it reads the consumptions and manages all the connected eg-plugs. the eg-plug is the link between the electrical network and the connected appliance. finally, the app is the tool used to make the initial configurations of each new eg-plug introduced and allows its’ connection to the system. i-etc: isel academic journal of electronics, telecommunications and computers vol. 5 , n. 1 (2019) id-1 http://journals.isel.pt both the controller and the eg-plug have an hardware and a software component as follows. a. hardware figure 2 hardware diagram for the eg-plug the eg-plug, shown in figure 2, is essentially composed of a sensor block, which reads the amount of current being consumed using an hall effect sensor (acs712), a 220[v] ac to 5[v] dc transformer block, to power all the required components and a cpu block, where the entire processing is executed. to implement the processor, the microcontroller nodemcu esp8266 was chosen due to the embedded integration of the wi-fi module esp-12e [5]. to control the electrical current for the appliance, the plug makes usage of an ssr (solid-state relay) but can also be used with an emr (electromechanical relay). the final aspect of the plug’s hardware is its user interface, with three leds, informing of the state of the plug, and a push button that allows the user to manually override the module. figure 3 hardware diagram for the controller the controller, shown in figure 3, is simpler than the egplug, presenting only a cpu block, which is equal to the cpu block used before, making use of the same microcontroller, and a user interface, implemented through the use of a small lcd screen with a resistive touch sensor and the respective driver. in the prototype version of the controller, the microcontroller is powered by the usb port, found on the esp8266 board, and a regular smartphone charger. the communication between the user interface and the microcontroller is made using two spi (serial peripheral interface) slave connections where the first channel communicates with the lcd screen and the second communicates with the screen’s touch sensor driver, which converts the analog signals from the sensor, into spi data frames. b. software in structural terms, the controller and eg-plug have similar software, since it was developed in a hierarchic block structure. each block is responsible for managing a corresponding hardware component, with a top layer main block which coordinates the remaining ones as a whole. the main differences are visible in the blocks themselves as these differ from the eg-plug to the controller since they have distinct objectives and functionalities. for example, the block that has to control the ssr state only exists in the egplug. as mentioned above, the system communicates through the local wi-fi network, making use of the mqtt (message queueing telemetry transport) [6] and tcp/ip protocols to effectively transmit its data. the data is all formatted in a simple protocol, with specific fields for information and action parameters, created for the system and called egc (electroges command). iii. application & discussion the boot up processes of both the controller and the egplug are very distinct. during the plug’s power up, a software ap (access point) is started creating a private network to allow for a device to connect and send the configuration data through the external app. after the information reaches the plug, it turns off the ap mode and attempts to connect to the specified wi-fi network from which it enters its normal operation mode, then being able to receive and send data to the connected controller and communicate through the user interface. figure 4 frontal panel and user interface of the eg-plug as mentioned in a previous section, the user interface has three informative leds and a push button, identified as ‘user’ on the plug’s panel as shown in figure 4. two more buttons exist in the prototype’s interface, and not shown in figure 2 to assist in debug and development; these are identified as ‘rst’ for resetting the hardware and ‘fsh’ to enable the program to be written to the flash memory. m. bento et al. | i-etc, vol. 5, n. 1 (2019) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt the red led informs the state of the plug, if it is on or off. the blue led in the middle informs the user if the plug is connected to the wi-fi or not, depending whether it is on or off, and blinks if the plug is receiving data. inside the system, each plug has a priority level from 0 to 2 which is meant for situations where the system is required to shut down devices, and has to decide which ones to close. this way it can search for the lower priority (0) ones first, and only if no lower priority devices are active, the highest are going to be deactivated. when all devices are on the same priority level, the system turns off the ones which consume the most, considering that the level of priority 2, means that the device is never to be deactivated by the automatic system. with this in mind, the final blue led informs the user of the priority level of the plug concerned, by the frequency at which the led blinks. a low frequency indicates the low priority (0), a high frequency is used for high priority (1) and a constant light means the highest value of priority (2). the button is used to manually override some of the plug functionalities, which are separated by the pressing time on the button. a quick press toggles the electrical current on the plug on or off, a short duration press, of 1 second, activates the leds for a 5 second period. all except for the plug state led are off to reduce idle energy consumption. finally a longer press, of 2 seconds, advances trough the priority levels in a looping way. the controller presents an interface that allows for the configurations to be edited in the device itself, not having the need to initiate an ap network on boot up, as shown on figure 5. figure 5 boot up and settings page of the controller still on figure 5 it is possible to see several fields such as ‘power limit’, which represents the total power allowed to be consumed by the sum of all plugs connected to this controller and the value in field ‘sensitivity’, which represents the percentage of the maximum power that the real consumption value can oscillate around depending on if the ‘restrict mode’ is toggled. when ‘restrict mode’ is on, the system starts disabling plugs as soon as the total power they are consuming reaches the ‘max power’ minus the ‘sensitivity’. in case of it being off, then the system only intervenes when the total consumption rises above the ‘max power’ plus the ‘sensitivity’. finally, there is the ‘clock’, which lets the user configure the current hour and minutes, allowing it to save the consumption statistics on the correct time they were registered. since the system uses no external clock at this prototype stage, when the controller resets, the clock resets to 00h00. in order to connect the controller to the network, its information has to be written in the settings page of the controller on the ‘ssid’ (network name) and the ‘keyp’ (network password) fields as shown in figure 6. after being connected, the controller will look for all the enabled controllers in the network and assume itself the next available id. when no controllers are found, it assumes id 0. figure 6 wi-fi credentials on the setings page of the controller both fields are saved in flash memory so that when the microcontroller resets it tries to reconnect to the network automatically. if the network is not available it is necessary to reintroduce the credentials of the desired network. after the controller is connected to the network and configured as intended, it is capable of accepting new egplugs, connected through the use of the external configuration app and integrates them immediately into the system. the initial configuration of each eg-plug is shown in the figure 7. figure 7 prototype app while configurating the first eg-plug into controller id 0 in one of the controller’s pages the user can see the list of connected plugs as shown in figure 8. in this page it is possible to turn on or off any of the visible plugs by clicking m. bento et al. | i-etc, vol. 5, n. 1 (2019) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt the green toggle button, which indicates the current state of each plug (on/off). it is also possible to check the consumption statistics of each plug by pressing the ‘st’ button. in the test performed for the purpose of this article, the name of the connected plug was left on default, which is “module/” plus the internal id for the plug, being 0 the id of the first plug added for this controller. figure 8 controller’s module (eg-plug) list in this page, it is also possible to configure each plug and observe its immediate consumption by pressing the name of the module, opening the settings of the plug concerned as shown in figure 9 for the enabled module in which a 600[w] hairdryer with 2 levels of intensity, was connected on the first level, which consumed about 300[w]. figure 9 controller’s configuration menu for a specific eg-plug other tests were made to ensure that the devices work as expected. iv. future work the final prototype still presents some imperfections such as some reading errors caused by interferences on the sensor and the methodology with which the data is obtained and read. these reading errors result in several small flaws like the power offset value of 30[w] when no current is flowing to the appliance and the value read should be of 0[w]. another factor that influences the readings, when the appliance power is off, is the small amount of current that can flow through the ssr which has a value of roughly 25[ma] and causes an unwanted idle consumption of about 5[w] to 6[w]. a random delay in the order of milliseconds to two seconds has been noted while the system is working and a change of state is needed to switch off low priority devices; due to wifi access. there is also a problem, related to the system integrity; when the controller loses power, every connected plug becomes “lose” and has to be reset as well and reconfigured to be reintegrated into the system, since in the prototype version there are no mechanisms to automatically restore the plugs into the correct controller. this can be solved by storing the configuration on the flash memory. the mobile application is also necessary for user convenience, and it is also on the “to do” list. v. conclusions management of appliances at home, by switching off those that are considered less important that the one that is going to be switched on, is effectively done by this system, which allows the home inhabitants to have a more comfortable life with continuous access to their electrical energy supply. configuration by the users is set at a minimum so that an intuitive process is achieved. the system can be scaled up by adding more eg-plugs while using the same controller. this project was developed in a considerably short period of time. in spite of that, given the evolution level the project achieved, during public presentation, it was considered, by the evaluating jury, as a prototyping-ready system. the success of the project lays on its functionality, as it was capable of correctly managing a single eg-plug inside the given parameters without issues. for this reason it was chosen to be presented in two different exhibitions, one of them meant for the future engineering students and the other showing ‘the best of isel’ to isel students. there are some plans and ideas that would be interesting to develop and implement in order to improve the system further: first of all, to better organize the hardware components and reduce the eg-plug size. next, to overcome the memory constrains imposed by the chosen microcontroller, since they limit the user interface design capabilities and the number of eg-plugs that can be connected to a single controller. still related to the controller, the user interface should be updated in order to present a better design and to allow for more features to be added, like a lock screen or consumption graphics. further development of the external app, could connect it to a desired controller first and obtain all the needed information to automatically configure a desired eg-plug, thus facilitating the initial setup for the user. finally creating a scheduling system to allow the user to choose dates and hours in which specific eg-plugs turn on or off., this we believe is a high possibility to reduce the consumption wastes of electricity with electroges autonomous energy management system. m. bento et al. | i-etc, vol. 5, n. 1 (2019) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt references [1] erse-entidade reguladora dos serviços energéticos, “preços de referência no mercado liberalizado de energia elétrica e gás natural em portugal continental,” 2015. [2] j. amador, “produção e consumo de energia em portugal: factos estilizados,” bol. económico | banco port., 2010. [3] i. i. pǎtru, m. carabaş, m. bǎrbulescu, and l. gheorghe, “smart home iot system,” in networking in education and research: roedunet international conference 15th edition, roedunet 2016 proceedings, 2016. [4] f. k. santoso and n. c. h. vun, “securing iot for smart home system,” in proceedings of the international symposium on consumer electronics, isce, 2015. [5] m. mehta, “esp8266 : a breakthrough in wireless sensor networks and internet of things,” int. j. electron. commun. eng. technol., 2015. [6] oasis (organization for the advancement of structured information standards), “mqtt version 3.1.1,” 2014. m. bento et al. | i-etc, vol. 5, n. 1 (2019) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt radio over fiber (rof): a comparison of low-cost systems radio over fiber (rof): a comparison of low-cost systems márcio almeida adeetc, isel, lisboa, portugal 29420@alunos.isel.pt pedro pinho adeetc, isel, lisboa, portugal instituto de telecomunicações, aveiro, portugal ppinho@deetc.isel.pt fernando m. v. ramos universidade de lisboa, faculdade de ciências, lisboa, portugal fernando.ramos@di.fc.ul.pt keywords: phase modulation, radio-over-fiber (rof), remodulation, wireless access network, low-cost systems. abstract: in order to assess the potential of low-cost radio-over-fiber (rof) solutions, in this paper we make a comparison of three full-duplex rof systems. these systems are low-cost solutions that use remote modulation, with a single centralized light source used at the central station to generate a downlink wavelength that is reused at the remote location for upstream transmission. by avoiding the need for an additional light source at each remote location the cost of the solution is significantly reduced. the three systems evaluated in this paper differ by the type of optical modulation used for downlink and uplink. the first is an im-im system using intensity-modulation (im) for the downlink and uplink direction. the second scheme, pm-im, differs from the first by using phase-modulation (pm) for downlink. finally, the third system, pm-pm, uses phase modulation for downlink and uplink. 1. introduction by combining the enormous capacity and low transmission loss of optical fiber networks with the ubiquity and mobility of wireless networks, radioover-fiber (rof) transmission techniques form a powerful platform for the support of emerging applications and services [1]. by allowing the centralization of complex signal processing, they enable the implementation of simple, compact, and low-cost remote base stations. considering the downlink as an example, in a rof system the radio signals are processed and modulated at the central station (cs) and are then delivered to the remote unit (ru) using an optical fiber. the ru has the sole responsibility of demodulating and transmitting these signals wirelessly. all complex signal processing is done at the cs. in order to simplify the rus and reduce the associated costs, rof architectures with a single optical source have been proposed [2]. in these lowcost solutions the optical carrier is generated in the cs and modulated with the electrical radiofrequency (rf) signal for downlink transmission. in the ru the optical carrier is reused to modulate the uplink signal. both the downlink and the uplink can use different types of optical modulation (e.g. intensity modulation (im) and phase modulation (pm)) as presented in [2-4]. in this paper we make a comparison between three different low-cost systems with different combinations of optical modulations, namely im in both downlink and uplink, pm in downlink and im in uplink, and pm in downlink and uplink. for the comparison we consider two figures of merit: the error-vector-magnitude (evm) and the received constellations. we simulated the three rof i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-5 http://journals.isel.pt/index.php/iajetc systems described using the vpiphotonics™ simulator. the rest of this paper is organized as follows. in section 2 we describe and quantify the general parameters of the systems analyzed. the im-im, pm-im and pm-pm rof systems are presented and evaluated in sections 3, 4, and 5, respectively. finally, section 6 concludes this paper with a discussion on the three schemes compared. 2. general parameters the three systems use the same type of rf signals. as input 16-qam is used for downlink and qpsk modulation for uplink. the main features of these signals are presented in table 1. downlink rf signal modulation 16-qam bit rate 200mb/s radio frequency 5.9ghz bandwidth 60mhz uplink rf signal modulation qpsk bit rate 100mb/s radio frequency 6ghz bandwidth 60 mhz table 1 rf signals parameters. as the main purpose is to obtain a low-cost system, we use only one continuous wave (cw) distributed feedback laser (dfb) at the source. the optical fiber used in all systems is a dispersion shifted fiber (dsf). this type of fiber is typical for broadband wireless access [2]. finally, the photodetection is made by means of a pin photodiode. the main characteristics of each component are presented in table 2. laser type dfb average power 1mw emission frequency 192.3thz (1552.64nm) linewidth 10mhz optical fiber type smf-dsf reference frequency 192.3thz (1552.64nm) attenuation 0.2db/km dispersion 0,787ps/(nm km) length 25km photodiode type pin responsivity 1a/w dark current 0a thermal noise 10pa/hz 1/2 table 2 optical components parameters. the main difference between the three systems is the optical modulation used for downlink and uplink. when im is used the optical modulator is an electroabsorption modulator (eam). for pm the optical modulator used is an ideal phase modulator. in the reception, when the modulation is im, we use a single photodiode. in the case of pm, we use a 25ps delay interferometer (di) followed by a balanced receiver with two photodiodes. 3. im-im system the first system analyzed is presented in figure 1, where im is used in both directions. the optical carrier is modulated with an eam driven by a 16qam signal for downlink transmission in the fiber. in the ru the received signal is split and half of the optical signal power is converted to electrical by a photodiode. for the uplink, the second output of the splitter is remodulated by another eam driven by a qpsk signal. the uplink transmission and reception of the optical signal is identical to the downlink. figure 1 im-im system. 3.1 results: downlink transmission for this system, the evm as a function of the modulation index (mi) is shown in figure 2. 0% 1% 2% 3% 4% 5% 6% 7% 8% 9% 10% 1% 2% 4% 8% 16% 32% 64% e v m modulation index (im) figure 2 evm as a function of the mi in downlink for the im-im system. m.almeida et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc according to the wimax standard [5], the evm should be of 6% for 16-qam and of 12% for qpsk to guarantee good performance. as we are assuming this type of broadband signals, we use these values as reference in all figures. by analyzing figure 2 we can conclude that the mi value that resulted in the best evm (0.17%) is 15%. to guarantee an evm below the threshold of 6%, the mi has to be outside the range 30%-60%. in figure 3 and figure 4 we present the constellation and 16-qam signal spectrum in the receiver for the best mi value, respectively. as expected from the evm value obtained, we can observe that the symbols appear very well defined in the constellation and that the snr has an acceptable value. figure 3 constellation in downlink for an mi of 15%. figure 4 signal spectrum in downlink for an mi of 15%. for the worst value of mi (40%) the constellation appears noisy and the snr worsens, as we can see in figure 5 and figure 6, respectively. figure 5 constellation in downlink for an mi of 40%. figure 6 signal spectrum in downlink receiver for mi of 40%. in conclusion, by choosing an adequate mi we can achieve very good performance in downlink. 3.2 results: uplink transmission as in the downlink case, we also present the evm as a function of the mi. figure 7 shows the evm as a function of the mi in uplink with a fixed downlink mi of 15%. in this case, the rf signal is modulated in qpsk, increasing the limit of evm to 12%, according to the standard. figure 7 evm as a function of the mi in uplink with a fixed downlink mi of 15%. in the uplink the choice of the mi is less restrictive than in the downlink, due to the distance between symbols in the qpsk constellation be higher than in 16-qam. the mi which results in lower evm is 30% as can be seen in figure 7. the reception of the uplink signal is shown in figure 8. by analyzing the figure, we can see also the downlink signal, because the uplink receiver demodulates both im signals. figure 8 signal spectrum in uplink of im-im system. m.almeida et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc the reception of these two channels causes interference, which can be minimized by limiting the power of the downlink channel. this power can be limited by adjusting the downlink mi. it is therefore important to study how the evm varies in uplink as a function of the mi for downlink. in figure 9 we can observe this variation, by fixing the uplink mi to30%. figure 9 evm in uplink as a function of the mi of downlink. from this figure, we can conclude that in an imim system, the downlink mi should be limited to 30%, to allow bidirectional communication. 4. pm-im system in the pm-im system, presented in figure 10, pm is used for the downlink direction. this system was discussed in [2,6]. in the ru the received downlink signal is input to a 25ps delay interferometer (di) to demodulate the pm signal. the null frequency of the constructive port of the di is detuned by 5.9 ghz from the signal wavelength to eliminate one of first-order sidebands, in order to increase the received power by removing beating effects. the output of the di is detected by a balanced receiver to increase the signal to noise ratio (snr) [2]. figure 10 architecture of the pm-im system. 4.1 results: downlink transmission for the pm-im system, the evm as a function of the normalized pm index is shown in figure 11 for the downlink. the normalized pm index is the ratio between the pm index, in degrees, and the phase between adjacent symbols, which for 16-qam is 22.5º. we can see that the minimum evm value is 1.28% for an mi of 31.1%. the pm index must be inside the range 4.4%-70%. figure 11 evm as a function of the mi in downlink of the pm-im system. . the received downlink signal constellation of the pm-im system for the modulation index value that resulted in the best evm is shown in figure 12. the symbols appear very well defined in the constellation. however, note that this constellation is rotated due to the 25ps delay introduced in the di. this does not affect the evm because the receiver uses amplitude and phase correction. figure 12 received constellation in downlink of the pmim system. 4.2 results: uplink transmission the uplink of this system is the same as the one from the im-im system. figure 13 shows the evm in uplink as a function of the mi in downlink, for a fixed uplink mi of 30%. m.almeida et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 0,0% 0,5% 1,0% 2% 3% 6% 13% 25% 50% 100% e v m n o u l omi no dl figure 13 evm in uplink as a function of the mi in downlink with a fixed uplink mi of 30%. by analyzing the figure we can conclude that the downlink mi has no influence on the uplink transmission. contrary to the im-im low-cost solutions, due to the constant intensity of the pm signal the modulation index of the downlink signal is not sacrificed and the power budget of the uplink is improved. this fact is the main advantage of a pm-im solution when compared to im-im. the main limitation of this system is related to the existence of the phase modulation to intensity modulation (pm-to-im) conversion effect caused by chromatic dispersion of the optical fiber. to assess this problem, in figure 14 we present the uplink signal at the receiver with the pm downlink signal. figure 14 pm-to-im effect on uplink receiver. the pm-to-im effect causes no major constraint because the power of the signal generated by the pm-to-im effect is two orders of magnitude lower than the power of the uplink signal. 5. pm-pm system finally, we present a system using pm for both downlink and uplink. the architecture of the pmpm system is presented in figure 15. figure 15 architecture of a pm-pm system. as the downlink of the pm-pm system is equivalent to the downlink of the pm-im system, in this section we focus only in the uplink direction. 5.1 results: uplink transmission the evm as a function of the normalized pm index is shown in figure 16. figure 16 evm as a function of the mi for uplink in the pm-pm system. the evm is acceptable for all values of the modulation index from 2.2% to 72.2%. as we can see in the figure, the best evm value obtained is 1.43% for a mi of 22.2%. another important issue is the influence of the downlink mi for the uplink reception. although the downlink has constant intensity, the utilization of the same type of modulation causes interference in the uplink receiver, because the di of the uplink demodulates both the downlink and the uplink signals. the evm for uplink as a function of the mi for downlink, with a fixed uplink mi of 22.2%, is shown in figure 17. m.almeida et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 0% 2% 4% 6% 8% 10% 12% 14% 2% 3% 6% 13% 25% 50% 100% e v m u l modulation index dl figure 17 evm in uplink as a function of the mi for downlink in the pm-pm system. by analyzing figure 17 it is possible to conclude that the evm for uplink is indeed influenced by the downlink mi. the best evm values are obtained when the downlink uses low values of mi. by taking into account both constraints, we conclude that the downlink mi should be restricted to the range 4.4%-70%. figure 18 shows the uplink signal at the receiver using the best values of the mi, 31.1% for downlink and 22.2% for uplink. figure 18 – signal spectrum in uplink of pm-pm system. in the figure we can see the two channels present in the uplink receiver. this is due to the fact that the pm receiver in uplink also demodulates the pm downlink signal. this situation is similar to the imim case. 6. discussion to conclude the paper we present a brief discussion on the results obtained. considering the im-im system, it is possible to achieve good performance if the mi is outside the range 30%-60%. however, as the uplink is also influenced by the downlink mi, and therefore, the downlink mi is limited to 30% to allow bidirectional communication. the uplink mi is not limited for values above 2.2%. for the pm-im system the downlink mi is not limited by the uplink performance, due the constant intensity of the pm signal. the limitation of mi for both directions is thus independent. being less restriction in the choice of the mi is an important advantage of this system. the pm-pm system presents similar problems of the im-im system, as the pm receiver of the uplink is able to demodulate the pm downlink signal. the downlink mi thus limits the uplink performance. as a final note, it is important to emphasize that for applications where a high modulation index is not required, the solution with im is a preferable choice due the simplicity of the receiver. references [1] navid ghazisaidi and martin maier, “fiber-wireless (fiwi) access networks: challenges and opportunities”, ieee network, vol. 25, no. 1, pp. 3642, jan. /feb., 2011. [2] ho-chul ji, hoon kim, and yun chur chung “fullduplex radio-over-fiber system using phasemodulated downlink and intensity-modulated uplink.” ieee photon. technol. lett., vol. 21, no. 1, pp. 9-11, jan. 1, 2009. [3] j. j. v. olmos, t. kuri, and k. -i. kitayama, “60-ghzband 155-mb/s and 1.5-gb/s baseband time-slotted full-duplex radio-over-fiber access network,” ieee photon. technol. lett., vol. 20, no. 7, pp. 617–619, apr. 1, 2008. [4] z. jia, j. yu, and g. -k. chang, “a full-duplex radioover-fiber system based on optical carrier suppression and reuse,” ieee photon. technol. lett., vol. 18, no. 16, pp. 1726–1728, aug. 15, 2006. [5] ieee standard for local and metropolitan area networks, part 16: air interface for broadband wireless access systems. [6] márcio almeida, pedro pinho, and fernando m. v. ramos, “evaluation of a low-cost radio-over-fiber system”, ix symposium on enabling optical networks and sensors, jul. 1, 2011, aveiro, portugal. m.almeida et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc electrospinning pan/pdcl2 nanofiber – the influence of the preparation method electrospinning pan/pdcl2 nanofiber – the influence of the preparation method ana neilde rodrigues da silva1,2, maria lúcia pereira da silva1,2, sebastião gomes dos santos filho1 1department of engineering of electronics systems, polytechnic school university of sao paulo, sp, brazil 2electronics systems department, fatec são paulo, sao paulo, brazil neilde@lsi.usp.br, malu@lsi.usp.br, sgsantos@lsi.usp.br abstract – adsorbent and catalytic materials has received attention, especially if obtained as nanofibers. polyacrylonitrile (pan) has allowed the obtaining of carbon nanofibers from the electrospinning process followed by a thermal treatment. pd has the capability of storage/react several compounds, such as hydrogen. therefore, the aim of this work was obtaining in a single step electrospun nanofibers decorated with pd. four different preparation methods were evaluated and, mainly due to the ability of pd atom to form a complex with dimethylformamide (dmf). ftirs and xps analysis indicate that pd atom remains on fibers, even after sample heating. these fibers are uniform and scanning electron microscopy does not show cluster formation. a preliminary model explaining these data was proposed. keywords: nanofibers, pd, h2 storage, electrospinning i. introduction nowadays, the search for materials that can be applied in the hydrogen storage has received much attention. in order to storage it such materials must be able to adsorb and desorb hydrogen molecules efficiently [1,2,3]. as adsorption depends on surface area, nanofibers are an excellent option and have attracted much attention from scientific community, due their characteristics such as high surface to volume ratio and easiness of producing [4,5]. on nanofibers manufacturing process, electrospinning stands out by means of low-cost setups. moreover, electrospinning allows obtaining nanofibers from several different polymeric solutions and/or mixtures as long as the adequate solution characteristics, such as viscosity, are settled. polyacrylonitrile (pan) solutions normally presents such features and has been extensively studied in the last decade; furthermore, with pan is possible to easily obtain carbon nanofibers from an electrospinning process followed by a thermal treatment in inert ambient [4,5]. several different heavy metal compound can storage hydrogen, one of most efficient is palladium [3]. on the other hand, carbon fibers can be decorated with pd by immersion in pdcl2 chemical solutions. some authors have reported to reach high capacities of hydrogen storage with such approach [3], however the weak bond of c-pd can jeopardize the performance of this material [8]. therefore, the aim of this work is the development of simple ways for obtaining pan fibers with pd ions incorporated. the approach consisted in development of solution where pan and pd ion are already mixed in order to favor the production in a single step of nanofibers that present pd atom in a free form, i. e., adequate for hydrogen adsorption. ii. experimental method solution preparation: considering data present on the literature, i. e., pd complexation with dmf [6] and the inclusion of this solvent in the pan macrostructure during solubilization [7], several different preparation methods were tested, as shown on table 1. therefore, in solution 1, the important driving force is the competition between pan and pd ion for dmf molecules. on the other hand, solution 2 only pd atoms will compete, not only for dmf but also tor pan molecules. solution 3 is the opposite, in this case the complexation of pd atoms is privileged and pan molecules will compete afterwards. finally, the solution 4 corresponds only to the mixing of two well developed samples, which means that weak interactions were expected. the solution 5 was prepared in the conventional way and was used as reference. all the solutions were 6% w/w pan/dmf and 1% w/w pdcl2/pan, respectively. table 1 composition of the solutions # solution description 1 (pan+pdcl2): dmf pan and the pdcl2 dry powder are mixed before being stirred with the dmf during 12 h. 2 (pan+dmf): pdcl2 pan is dissolved in the dmf and stirred during 12 h. following, pdcl2 is mixed until total dissolution 3 (pdcl2+dmf): pan pdcl2 is dissolved in the dmf and stirred during 12 h. following, pan is added and the solution is mixed until the total dissolution 4 (pan+dmf): (pdcl2+dmf) pan is dissolved in dmf, pdcl2 is dissolved in the dmf, and both solutions are stirred during 12 h. following, the two resulting solutions are mixed and stirred until the total dissolution 5 pure pan pan is stirred with dmf until total dissolution i-etc: isel academic journal of electronics, telecommunications and computers vol. 6, n. 1 (2020) id-1 http://journals.isel.pt mailto:neilde@lsi.usp.br mailto:malu@lsi.usp.br mailto:sgsantos@lsi.usp.br figure 1: schematics of the electrospinning apparatus. fibers production: the fibers were electrospun using a conventional apparatus, as shown in figure 1. the electrospinning process was conducted as follows: a 5 ml syringe with a hypodermic needle (22g1) was filled with the solution and 15 kv was applied to the needle that was kept 15 cm from the grounded base. the electrospinning process was conducted during 20 min. fiber characterization: the fibers morphology and diameters were evaluated by optical and scanning electron microscopy (sem). the chemical bonds and their interaction between the polymer and the pdcl2 were evaluated by raman spectroscopy, fourier transform infrared spectroscopy (ftirs) and x ray photoelectron spectroscopy (xps). solution characterization: the viscosity was the solution property used to evaluate the macroscopic interaction between pan and pdcl2. to measure the absolute viscosity of the studied solutions the brookfield viscometer rv-dv ii was used. as the volume of the solution was less than the regular spindle can measure, a small sample adapter was used in the measurements. materials polyacrylonitrile (pan) and n,n dimethylformamide (dmf) were purchased from sigma-aldrich and, pdcl2, from synth. to make the fiber characterization easier, it was collected over a (100) silicon substrate. iii. results it is known that during the electrospinning process, the fiber diameters are influenced by the solution conductivity, viscosity and by the process parameters [3]. so, it is expected that an increase in the viscosity may result in larger fiber diameter [4,5]. however, as observed in figure 2, the diameter of the fibers decreases while the viscosity of the precursor solutions increases. the palladium chloride can form ion complex with dimethylformide [6], indicating that the preparation method of the solution can influence the chemical interaction among dmf, pan and pdcl2. as described in table i, the solutions 3 and 4 were prepared by dissolving pdcl2 in dmf before the addition of the polymer as powder or as dissolved in the solvent, respectively. as can be observed in figure 2, although the respective low solution viscosity these electrospun fibers have higher diameters. it is understood that the complex formation between pan and pdcl2 decreases the availability of free dmf molecules in the solution and influences the mechanism of polymer dissolution resulting in low viscosity. otherwise, there is less pd ions in the solution due to the complex formation and, possibly, during the electrospinning process, occurs less repulsion between the polymer molecules resulting in fiber with higher diameter. figure 2: variations in the viscosity of the studied pan/pdcl2 solutions and in the diameters of the fibers electrospun from these solutions. figure 3: variation in the solutions color evidencing the pd complex formation; a) as prepared; b) after one week; c) after one year. in solution 1, due to the proximity during solubilization, the interaction between pan molecule and pd atoms can be favored; therefore the “apparent” molecular weight of the polymer changes and viscosity increases. however, pan-pd interaction also corresponds to less complexation and the effective charge on the atom increases; thus repulsion is privileged and the fiber diameter decreases. solution 2 is an intermediary situation pa n+ dm f [5 ] (pa n+ pd cl) :dm f [1 ] (pa n+ dm f): pd cl [2] (pd cl+ dm f): pa n [ 3] pa n+ dm f:p dc l+d mf [4] 60 80 100 120 140 160 66 ,7 27 6,4 viscosity v is c o s it y ( c p ) solutions 200 300 400 500 600 cp diameter d ia m e te r (n m ) nm 01 02 0403 after one week pan/pdcl2 precursor solution a) b) after one year 01 02 0403 01 02 0403 c) at the time was prepared pd/dmf solution phase separation a.n.r. da silva et al. | i-etc, vol. 6, n. 1 (2020) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt between solution 1 and 5, pd atoms probably are not completely complexed when interact with pan molecules whereas this polymeric structure is already involved by dmf molecules. pan/dmf solutions are transparent and light yellow. however, as can be seen in figure 3, pan/pdcl2 solutions are dark brown indicating that pd complexation occurred. furthermore, solution ageing for over one year shows different behavior for each solution. whereas for solutions 3 and 4 a homogeneous dark color liquid material remains, in solution 1 occurs phase separation with the formation of a dark solid material and remained solution with light yellow color that resembles pan/dmf solutions and solution 2 presents some solid material dispersed in the liquid. the formation of this solid is an indicative of strong interaction between pd atom and pan molecule, which can lead to three-dimensional rigid structures during ageing. ftir analysis was used to unravel chemical environment and the respective species interactions. films were produced by dropping small solution amount on a silicon wafer; thus, no chemical stress due to unexpected driving forces is presumed to be present on these samples. on the other hand, the electrospinning process and the consequent electrostatic fields can impose such different interactions among chemical species. as expected, [9,10,11], for all fibers and films, ftir spectra show the main bands of pan molecule: amine groups (~ 3300 3500cm-1), c n (~2240 cm-1), c-h (~1450 cm-1) and figure 4a shows typical results. on fibers, the addition of pdcl2 changes peak maximum and its shape on the 1000 – 1800 cm-1 region and details in figure 4b) point out these features. moreover, the spectra show different behavior, depending on solution manufacturing methods, to the bands 1658 cm-1 and 1608 cm-1. cipriani et al [11] indicates bands at 1660 cm-1 region as c=o species, due to the solvent presence (dmf molecules) and pan film in the figure shows a single well resolved band on this region. the addition of pdcl2 salt on this environment ([pan+dmf]/pdcl2) compete with the dmf c=o bond and a 1608-cm-1 band is partially revealed. with two different solutions being mixed ([pan+pdcl2]), both conditions (pan and the salt involved by dmf molecules) are settled down before the mixture and the ftir shows two resolved bands. finally, the mixture of polymer and salt prior to the addition of the solvent favors a competition mechanism and the bands are found, but completed mixed (not resolved). the 1608 cm-1 is uncommon on pan films or fibers. nonetheless, huang et al [9] assigned this band as c=n species, presents in oxidized pan fibers; this oxidized material can also crosslink with pan molecule. badii et al [10] found a similar band (c=n), also due to oxidation/cyclization, that was assigned at 1592 cm-1 and cipriani et al [11] considers the range 1610-1575 cm-1 indicative of pan degradation and formation of c=c and c=n. since these films were not exposed to thermal treatment or oxidant reactants, a possible explanation to the appearance of this vibration band at 1608 cm-1 is the pd atom connected to the nitrogen (n) in the dmf molecule, but this band is not found in [pd(dmf)2cl2] infrared spectrum [12], or a direct interaction with the nitrile species on pan molecule, as explained later on the qualitatively model. figure 4: ftirs analysis: (a) typical spectra for film, fiber and after thermal annealing, and details of (b) fibers and thermal annealing of (c) films and (d) fibers regarding [pd(dmf)2cl2] spectrum, a strong band is expected in 1629 cm-1 and, in fact, some fibers actually show vibration on this range, especially if pd was solved by dmf before the addition of pan molecules. 500 1000 1500 2000 2500 3000 3500 4000 0,70 0,75 0,80 0,85 0,90 0,95 t ra n s m it a n c e ( a .u .) wavenumber (cm-1) pan film [pan+dmf]/[pdcl2+dmf] film [pan+dmf]/pdcl2 [pan+pdcl2]/dmf [pdcl2+dmf]/pan film from precursor solutionsa) 1000 1200 1400 1600 1800 0,70 0,75 0,80 0,85 0,90 film from precursor solutionsb) t ra n s m it a n c e ( a .u .) wavenumber (cm-1) pan film [pan+dmf]/[pdcl2+dmf] film [pan+dmf]/pdcl2 [pan+pdcl2]/dmf [pdcl2+dmf]/pan 1000 1500 2000 2500 3000 3500 4000 92 94 96 98 100 102 fibers annealed t ra n s m it a n c e ( % ) wavenumber (cm-1) (pan+pdcl):dmf fiber annealed (pan+dmf):pdcl fiber annealed (pdcl+dmf):pan annealed pan+dmf:pdcl+dmf annealed c) 1000 1200 1400 1600 1800 90 92 94 96 98 100 102 fibers annealed t ra n s m it a n c e ( % ) wavenumber (cm-1) (pan+pdcl):dmf fiber annealed (pan+dmf):pdcl fiber annealed (pdcl+dmf):pan annealed pan+dmf:pdcl+dmf annealed d) a.n.r. da silva et al. | i-etc, vol. 6, n. 1 (2020) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt thermal annealing (500ºc) of films (figure 4c) and fibers (figure 4d) leads to similar behavior, with prominent band on 1550 cm-1 due to cyclization and other ch bands. however, the decrement of 1660 cm-1, assigned as solvent [11], is more evident on fibers, probably due to the high surface/area ratio. for both, films and fibers, the preservation of 1660 cm-1 seems to be linked with pd complexation by dmf molecules. it is worth noting that raman analysis did not find graphite. since ftir analysis shows different chemical environment on fibers decorated with pd, xps analysis was carried out to evaluate pd chemical shifts. figure 5a shows typical results. the horizontal baseline on these spectra indicates that there is no contamination and c, pd and o peaks are evident. regarding pan molecules, c peak, on figure 5b, points out that there is no significant chemical shift. according to [13], xps sign of pd(0) is expected at ~335 ev, which does not occur on these spectra, and 3d5/2 sign of pdcl2 complex range from 334.8 ev to 336.4 ev but reaches 338 ev for nanoparticles or films. two different solutions clearly showed peaks on such range and figure 5c presents such peaks. furthermore, on these fibers c/pd atomic weight ratio can achieve up to 0.0097 and 0.0018 for solutions 2 and 4, respectively, i.e., an expressive amount. thus, this data allows one to infer that the fibers electrospun from solutions 2 and 4 have freer pd than the other solutions which probably means that competition for dmf molecules (solution 1) as well as the complexation of the pd ion (solution 3) binds more strongly the pd ion to the polymer structure. figure 5: xps analysis. (a) full spectra and details of (b) c and (c) pd signals. on the other hand, besides the variation in the electrospun fiber diameters and in the chemical environment on the films, there is no variation in their morphology that can be observed in the sem images as shown in figure 6. that seems to indicate that there is no clusterization of pd atoms. a preliminary model representing the chemical interaction between pd and the solvent is proposed based on the previous discussions. table 2 shows the most probably interaction of pan/dmf/pd for each solution, the main characteristics, fiber diameter, viscosity and chemical bonds. figure 6: sem images from the fibers electrospun of the solutions without pd compared to the fibers electrospun from solutions 2 and 4 with pd. iv. conclusions this work presented the preliminary results of the influence on the preparation method of pan/pdcl2/dmf electrospinning precursor solutions. due to pd complexation with dmf, the preparation method influences the interaction between pd and pan. depending on the preparation method the quantity of incorporated pd on the fiber surface can vary, thus it is expected that the action of pd as a catalytic or in the hydrogen storage is influenced. it can also be observed that although the solution viscosity decreases the diameters of electrospun fibers is larger than expected. the ftirs analysis confirms that pd complexes with dmf by means of bonds to n; furthermore, thermal treatment does not remove significant bands regarding pd complex. xps pointed out pd(ii) incorporation and high c/pd atomic ratio. finally, the low-cost setup can be easily operate to produce in a single step nanofibers decorated with pd. thus, the results suggest that the fiber electrospun from solutions 2 and 4 are promising to produce fiber decorated with pd after the thermal treatment. 1000 800 600 400 200 0 0 200 400 600 800 1000 1200 1400 1600 c 1s pd 1s pan (pan+dmf) + pdcl2 (pan+dmf)+(pdcl2+dmf) c /s binding energy (ev) 292 288 284 280 276 400 600 800 1000 1200 1400 pan (pan+dmf) + pdcl2 (pan+dmf)+(pdcl2+dmf) c /s binding energy (ev) c 1s 360 356 352 348 344 340 336 332 328 460 480 500 520 540 560 pan (pan+dmf) + pdcl2 (pan+dmf)+(pdcl2+dmf) c /s binding energy (ev) pd 1s a) b) c) a.n.r. da silva et al. | i-etc, vol. 6, n. 1 (2020) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt table 2 proposed electronic structure after the chemical interactions for each solution, the main characteristics, fiber diameter, viscosity and chemical environment. s o lu ti o n m a in c h a ra c te ri st ic s v is u a l o b se rv a ti o n f ib e r d ia m e te r v is c o si ty c h e m ic a l e n v ir o n m e n t most probably interaction pan / dmf / pd 1 c o m p le x a ti o n c o m p e ti ti o n p h a se s e p a ra ti o n lo w e st h ig h e st t w o n o t re so lv e d b o n d s 2 p a n /d m f s tr o n g i n te ra c ti o n p a rt ia l p h a se s e p a ra ti o n m e d iu m m e d iu m p a rt ia ll y r e v e le d b o n d f re e r p d 3 p d /d m f s tr o n g i n te ra c ti o n h o m o g e n e o u s a p p e a ra n c e h ig h lo w e st 4 p a n a n d p d /d m f s tr o n g in te ra c ti o n h o m o g e n e o u s a p p e a ra n c e h ig h e st lo w t w o r e v e le d b o n d s, f re e r p d similar to solution 2 and solution 3 c n pd oc n c cn o c n c h h solução 1 c n c solution 2 c n c c solution 3 c n c pd oc n h o h c n h o a.n.r. da silva et al. | i-etc, vol. 6, n. 1 (2020) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt v. references [1] lee, hye-min; heo, young-jung; an, kay-hyeok; jung, sang-chul; chung, dong chul; park, soo-jin; kim, byung-joo; a study on optimal pore range for high pressure hydrogen storage behaviors by porous hard carbon materials prepared from a polymeric precursor; int. journal of hydrogen energy 43 (2018) 5894-5902. doi.org/10.1016/j.ijhydene.2017.09.085 [2] konda, suresh k.; chen, aicheng; palladium based nanomaterials for enhanced hydrogen spillover and storage; materials today 19 (2016) 100-108. doi.org/10.1016/j.mattod.2015.08.002 [3] kim, hongyeun; lee, daehee; moon, jooho; coelectrospun pd-coated porous carbon nanofibers for hydrogen storage applications; international journal of hydrogen energy 36 (2011) 3566-3573, doi:10.1016/j.ijhydene.2010.12.041 [4] nandana bhardwaj, subhas c. kundu, “electrospinning: a fascinating fiber fabrication technique”, biotechnology advances 28 (2010) 325–347 [5] gomes, demetrius s., silva, ana n. r. da, morimoto, nilton i., mendes, luiz t. f., furlan, rogerio, & ramos, idalia. (2007). characterization of an electrospinning process using different pan/dmf concentrations. polímeros, 17(3), 206-211. doi.org/10.1590/s0104-14282007000300009 [6] pd palladium: palladium compounds, william p. griffith, stephen d. robinson, kurt swars, springer science and business media, 2013, 355 p, isbn: 9783662091883 [7] youngho eom, byoung chul kim, solubility parameterbased analysis of polyacrylonitrile solutions in n,ndimethylformamide and dimethyl sulfoxide, polymer, volume 55, issue 10, 2014, pages 2570-2577, issn 0032-3861, https://doi.org/10.1016/j.polymer.2014.03.047 [8] martyna baca, krzyszto cendrowski, wojciech kukulka, grzegorz bazarko, dariusz moszyński, beata michalkiewicz, ryszard j. kalenczuk, beata zielinska; a comparison of hydrogen storage in pt, pd and pt/pd alloys loaded disordered mesoporous hollow carbon spheres, nanomaterials 2018, 8(9), 639; doi.org/10.3390/nano8090639 [9] fei huang, yonggen lu, li chen, liguo liu and junqi jiang, a new polyacrylonitrile fiber for direct carbonization without oxidation, j mater sci (2018) 53:8232–8240, https://doi.org/10.1007/s10853-018-2158-y [10] khashayar badii, jeffrey s. church, gelayol golkarnarenji, minoo naebe, hamid khayyam; chemical structure-based prediction of pan and oxidized pan fiber density through a non-linear mathematical model. polymer degradation and stability 131 (2016) 53 e 61, http://dx.doi.org/10.1016/j.polymdegradstab.2016.06.019 [11] e. cipriani, m. zanetti, p. bracco, v. brunella, m.p. luda, l. costa, crosslinking and carbonization processes in pan films and nanofibers polymer degradation and stability 123 (2016) 178e188 [12] cationic and neutral chloride complexes of palladium(ii) with the nonaqueous solvent donors acetonitrile, dimethyl sulfoxide, and a series of amides. mixed sulfur and oxygen coordination sites in a dimethyl sulfoxide complex, bradford b. wayland and robert f. schramm, inorganic chemistry, vol. 8, no. 4, april 1969 [13] maggy f. lengke, michael e. fleet, gordon southam langmuir 2007, 23, 17, 8982-8987, doi.org/10.1021/la7012446 a.n.r. da silva et al. | i-etc, vol. 6, n. 1 (2020) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt lempel-ziv sliding window update with suffix arrays lempel-ziv sliding window update with suffix arrays artur ferreira1,3,4 arlindo oliveira2,4 mário figueiredo3,4 1instituto superior de engenharia de lisboa (isel) 2instituto de engenharia de sistemas e computadores – investigação e desenvolvimento (inesc-id) 3instituto de telecomunicações (it) 4instituto superior técnico (ist), lisboa, portugal arturj@isel.pt aml@inesc-id.pt mtf@lx.it.pt keywords: lempel-ziv compression, suffix arrays, sliding window update, substring search. abstract: the sliding window dictionary-based algorithms of the lempel-ziv (lz) 77 family are widely used for universal lossless data compression. the encoding component of these algorithms performs repeated substring search. data structures, such as hash tables, binary search trees, and suffix trees have been used to speedup these searches, at the expense of memory usage. previous work has shown how suffix arrays (sa) can be used for dictionary representation and lz77 decomposition. in this paper, we improve over that work by proposing a new efficient algorithm to update the sliding window each time a token is produced at the output. the proposed algorithm toggles between two sa on consecutive tokens. the resulting sa-based encoder requires less memory than the conventional tree-based encoders. in comparing our sa-based technique against tree-based encoders, on a large set of benchmark files, we find that, in some compression settings, our encoder is also faster than tree-based encoders. 1 introduction the lempel-ziv 77 (lz77) [14, 19], and its variant lempel-ziv-storer-szymanski (lzss) [14, 16], are lossless compression algorithms that are the basis of a wide variety of universal data compression applications, such as gzip, winzip, pkzip, winrar, and 7-zip. those algorithms are asymmetric in terms of time and memory requirements, with encoding being much more demanding than decoding. the main reason for this difference is that the encoder part requires substring search over a dictionary, whereas decoding involves no search. most lz-based encoders use efficient data structures, such as binary trees (bt) [6, 11], suffix trees (st) [5, 7, 9, 13, 17], and hash tables, thus allowing fast search at the expense of higher memory requirement. the use of a bayer-tree, along with special binary searches on a sorted sliding window, has been proposed to speedup the encoding procedure [6]. suffix arrays (sa) [7, 10, 15], due to their simplicity, space efficiency, and linear time construction algorithms [8, 12, 18], have been a focus of research; e.g., sa have been used in encoding data with antidictionaries [4] and to find repeating sub-sequences for data deduplication [1], among other applications. recently, algorithms for computing the lz77 factorization of a string, based on sa and auxiliary arrays, have been proposed to replace trees [2, 3]. these sa-based encoders have the two following memoryrelated advantages over tree-based encoders: they require less memory; the amount of allocated memory is constant and a priori known, being independent of the dictionary contents. in contrast, encoders based on hash tables or trees encoders, usually require allocating a maximum amount of memory. the main disadvantage of sa-based encoders is their encoding time, which is typically above that of tree-based encoders, attaining roughly the same compression ratio. regarding previous work on sa for lz decomposition, it has been found that the main drawback of the method in [2] is the absence of a strategy to update the sa as encoding proceeds: the entire sa is repeatedly rebuilt. the proposals in [3] for lz decomposition with sa are memory efficient, but the encoding time is above that of tree-based encoders. i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-4 http://journals.isel.pt/index.php/iajetc 1.1 our contribution in this paper, we improve on previous approaches [2, 3] by proposing an algorithm for sliding window update using sa and a fast technique for finding the tokens. the application of these techniques to lz77/lzss encoding does not involve any changes on the decoder side. our sa-based encoder uses a small amount of memory and can be faster than the tree-based ones, like 7-zip, being close to gzip in encoding time on several standard benchmark files, for some compression settings. the rest of the paper is organized as follows. section 2 reviews basic concepts of lz77/lzss encoding and decoding as well as the use of sa for this purpose. section 3 describes the proposed algorithms. the experimental results are presented and discussed in section 4, while section 5 contains some concluding remarks. 2 lempel-ziv basics the lz77 and lzss [14, 16, 19] lossless compression techniques use a sliding window over the sequence of symbols to be encoded with two subwindows: • the dictionary which holds the symbols already encoded; • the look-ahead-buffer (lab), containing the next symbols to be encoded. as the string in the lab is encoded, the window slides to include it in the dictionary (this string is said to slide in); consequently, the symbols at the far end of the dictionary are dropped (they slide out). at each step of the lz77/lzss encoding algorithm, the longest prefix of the lab which can be found anywhere in the dictionary is determined and its position stored. for these two algorithms, encoding of a string consists in describing it by a token. the lz77 token is a triplet of fields, (pos, len, sym), with the following meanings: • pos location of the longest prefix of the lab found in the current dictionary; • len length of the matched string; • sym the first symbol in the lab that does not belong to the matched string (i.e., that breaks the match). in the absence of a match, the lz77 token is (0,0,sym). each lz77 token uses log2(|dictionary|)+ log2(|lab|) + 8 bits, where |.| denotes length (number of bytes); usually, |dictionary|�|lab|. in lzss, the token has the format (bit,code), with the structure of code depending on value bit as follows: { bit = 0 ⇒ code = (sym), bit = 1 ⇒ code = (pos, len). (1) in the absence of a match, lzss produces (0, sym), otherwise (1, pos, len). the idea is that, if a match exists, there is no need to explicitly encode the next symbol. besides this modification, storer and szymanski [16] also proposed keeping the lab in a circular queue and the dictionary in a binary search tree, to optimize the search. lzss is widely used in practice since it typically achieves higher compression ratios than lz77 [14]. the fundamental and most expensive component of lz77/lzss encoding is the search for the longest match between lab prefixes and the dictionary. in lzss, the token uses either 9 bits, when it has the form (0,sym), or 1 + log2(|dictionary|) + log2(|lab|) bits, when it has the form (1,(pos,len)). figure 1 shows an example of lz77 encoding for a dictionary of length 16 and lab with 8 symbols. 2.1 decoding procedures assuming that the decoder and encoder are initialized with equal dictionaries, the decoding of each lz77 token (pos,len,sym) proceeds as follows: 1) len symbols are copied from the dictionary to the output, starting at position pos of the dictionary; 2) the symbol sym is appended to the output; 3) the string just produced at the output is slid into the dictionary. for lzss decoding, we have: 1) if the bit field is 1, len symbols, starting at position pos of the dictionary, are copied to the output; otherwise sym is copied to the output; 2) the string just produced at the output is slid into the dictionary. both lz77 and lzss decoding are low complexity procedures which do not involve any search, thus decoding is much faster than encoding. 2.2 using a suffix array a suffix array (sa) represents the lexicographically sorted array of the suffixes of a string [7, 10]. for a string d of length m (with m suffixes), the suffix array p is the set of integers from 1 to m, sorted by the lexicographic order of the suffixes of d. for instance, if we consider dictionary d=business-machine (with m=16), we get a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 1: illustration of lz77 encoding with dictionary “business-machine”. we show the corresponding outputted lz77 tokens for the encoding of the prefix of the lab. figure 2: lz77 and lzss with dictionary “business-machine” and its representation with sa p. the encoding of the prefix “s-mak” of the lab can be done with substrings ranging from d[p[le f t]] to d[p[right]]. we choose the 4-symbol match at p[13] producing the depicted lz77/lzss tokens. p ={9,11,1,12,16,6,13,14,4,10,15,5,8,3,7,2}, as shown in figure 2. each integer in p is the suffix number corresponding to its position in d. finding a substring of d, as required by lz77/lzss, can be done by searching array p; for instance, the set of substrings of d that start with ‘s’, can be found at indexes 3, 7, and 8 of d, ranging from index 13 to 15 on p. in this work, we use the suffix array induced sorting (sa-is) algorithm to build the sa [12]. 3 sliding window update algorithm in this section we present the proposed algorithm for sliding window update as well as an accelerated technique to obtain the tokens over a dictionary. this work addresses only the encoder side data structures and algorithms, with no effect in the decoder. decoding does not need any special data structure and follows standard lz77/lzss decoding, as described in subsection 2.1. 3.1 accelerated encoder the lz77/lzss tokens can be found faster if we use an auxiliary array of 256 integers (named li – leftindex). this array holds, for each ascii symbol, the first index of the suffix array where we can find the first suffix that starts with that symbol (the left index for each symbol, as shown in figure 3). for symbols that are not the start of any suffix, the corresponding entry is labeled with -1, meaning that we have an empty match for those symbols. figure 3 shows the li array for the dictionary of figure 2. the left value, as depicted in figure 2, is computed by a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 3: the li (leftindex) auxiliary array: for each symbol that starts a suffix it holds the index of the sa p in which that suffix starts. for the symbols that are not the start of any suffix, the corresponding entry is marked with -1, meaning that we have an empty match for substrings that start with that symbol. le f t ← p[li[lab[1]]]. the right indicator is found by iterating p starting on index left and performing a single symbol comparison. if we want lz77/lzss “fast” compression, we choose pos = left; for “best” compression, we choose left ≤ pos ≤ right, such that we have the longest match between substrings starting at d[pos] and lab[1]. 3.2 proposed algorithm the key idea of our sliding window technique is to use two sa of length |dictionary|, named pa and pb, and a pointer p (to pa or pb) to represent the dictionary. at each substring match, that is, each time we produce a token, we toggle pointer p between the two sa and we also update the li array. if the previous token was encoded with pa, the steps next described are carried out using pb, and vice-versa. this idea somewhat resembles the double buffering technique, since we are switching from one sa to the other, every time a token is produced. if we used a single sa, we would have a slow encoder, because we would have to perform several displacements of the integers on the unique large sa. these integer displacements would leads to us to a situation in which the encoder would be slow. for both lz77/lzss encoding, each time we output a token encoding l symbols, the dictionary is updated as follows: r. remove suffixes {1,...,l} (they slide out); i. insert in a lexicographic order the suffixes ranging from |dictionary|−l + 1 to |dictionary| (they slide in); u. update suffixes {l + 1,...,|dictionary|}; these are subtracted by l. figure 4 shows these r, i, and u actions, for the dictionary in figs. 2 and 3, after encoding s-mak with l=5 symbols. the removal action (slid out) is implicit by toggling from sa pa to pb; the updated suffixes keep the order between them; the inserted (slid in) suffixes are placed in the destination sa, in lexicographic order. algorithm 1 details the set of actions taken by our proposed algorithm. algorithm 1 runs each time we produce a lz77/lzss token; in the case of lz77, we set l = len + 1; for lzss l = len. we can also update the sa with the entire lab contents using l = |lab|, after we produce the set of tokens encoding the entire lab. this algorithm also works in the “no match” case of lzss, in which the token is (0, sym), with l=1. notice that we use two auxiliary arrays with length up to |lab|; we thus have a memory-efficient sliding window algorithm. the amount of memory for the encoder data structures is msa = 2|p|+ |li|+ |ps|+ |i| (2) bytes. figure 5 illustrates algorithm 1 using the dictionary shown in figs. 3 and 4, after encoding s-mak with l = 5. we see the sa pa as origin and pb as destination; we also show the contents of ps and i with 5 positions each. a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 4: illustration of our sliding window update algorithm with the r i u actions, after encoding “s-mak” with l=5: r) suffixes 1 to 5 slide out of pa; i) suffixes 12 to 16 are inserted in lexicographic order into pb; u) suffixes 6 to 16 are updated being subtracted by 5. figure 5: sliding window update with pointer p set to pa initially; the first update is done using pb as destination. array i holds the indexes where to insert the new suffixes into pb. the update of the indexes is done over pb and there is no modification on pa or integer displacement on pb. 4 experimental results our experimental tests were carried out on a laptop with a 2 ghz intel core2duo t7300 cpu and 2 gb of ram, using a single core. the code was written in c, using microsoft visual studio 2008. the linear time sa construction algorithm sa-is [12] (available at yuta.256.googlepages.com/sais) was used. for comparison purposes, we also present the results of nelson’s binary tree (bt) encoder [11], gzip1, and the lz markov chain algorithm (lzma2). these three encoders were chosen as benchmark, 1www.gzip.org/ 2www.7-zip.org since they represent the typical usage of tree and hash tables data structures for lz77 compression. the btencoder represents the dictionary with a binary tree data structure. the well-known gzip encoder uses trees and hash tables. lzma is the default compression technique employed by the 7z format in the 7zip program. both the gzip and lzma encoders perform entropy encoding of the tokens produced by the lzma algorithm. this allows for these algorithms to attain a higher compression ratio than our algorithms and the bt-encoder (it does not perform entropy encoding of the tokens). the test files are from the standard corpora calgary (18 files, 3 mb) and silesia (12 files, 211 mb), a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc yuta.256.googlepages.com/sais www.gzip.org/ www.7-zip.org available at www.data-compression.info. we use the “best” compression option, by choosing the longest match as discussed in section 3.1 and depicted in figure 2. 4.1 experimental setup in our tests, we assess the following measures: encoding time (in seconds, measured by the c function clock); compression ratio (in bits per byte, bpb); amount of memory for encoder data structures (in bytes). nelson’s bt-encoder [11] uses 3 integers per tree node with |dictionary|+ 1 nodes, occupying mbt = 13×|dictionary|+ 12 (3) bytes, using 4-byte integers. larsson’s suffix tree encoder (available at algorithm 1 sa sliding window algorithm input: pa,pb, m-length sa; p, pointer to pa or pb; pdst , pointer to pb or pa; lab, look-ahead buffer; li, 256-position length leftindex array; l ≤ |lab|, number of symbols in the previously produced token(s). output: pa or pb updated; p pointing to the recently updated sa. 1: if p points to pa then 2: set pdst to pb. {/*r action. implicit removal.*/} 3: else 4: set pdst to pa. 5: end if 6: compute the sa ps for the encoded substring (with l positions). 7: using li and ps, fill the l-length array i with the insertion indexes (slide in suffixes). 8: for i = 1 to l do 9: pdst [ i[i] ] = ps[i]+ |dictionary|−l. {/*i action.*/} 10: end for 11: do nupdate = |dictionary|−l; 12: do j=1. {/*perform |dictionary|−l updates.*/} 13: for i = 1 to |dictionary| do 14: if (p[i]−l) > 0 then 15: while ( j ∈ i) do 16: j = j + 1. {/*make sure that j is an update position.*/} 17: end while 18: pdst [ j] = p[i]−l. {/*u action.*/} 19: j = j + 1. 20: nupdate = nupdate 1. 21: if (nupdate==0) then 22: break. {/*destination sa is complete.*/} 23: end if 24: end if 25: end for 26: set p to pdst . {/*p points to recently updated sa.*/} www.larsson.dogma.net/research.html) uses 3 integers and a symbol for each node, occupying 16 bytes, placed in a hash table [9], using the maximum amount of memory mst = 25×|dictionary|+ 4×hashsz + 16 (4) bytes, where hashsz is the hash table size. the gzip encoder occupies mgzip=313408 bytes, as measured by sizeof c operator. the lzma encoder data structures occupy mlzma = 4194304+      9.5|dict.|, if mf = bt2 11.5|dict.|, if mf = bt3 11.5|dict.|, if mf = bt4 7.5|dict.|, if mf = hc4 , (5) bytes, depending on the match finder (mf) used as well as on |dictionary| with bt# denoting binary tree with # bytes hashing and hc4 denoting hash chain with 4 bytes hashing. for instance, with (|dictionary|,|lab|) = (65536,4096), we have in increasing order msa =627712, mbt =851980, mst =1900560, and mlzma=4816896 bytes. if we consider an application in which we only have a low fixed amount of memory, such as the internal memory of an embedded device, it may not be possible to instantiate a tree or a hash table based encoder. the gzip and lzma3 encoders are built upon the deflate algorithm, and perform entropy encoding of the tokens achieving better compression ratio than our lzss encoding algorithms. these encoders are useful as a benchmark comparison, regarding encoding time and amount of memory. for both compression techniques, we have compiled their c/c++ sources using the same compiler settings as for our encoders. the compression ratio of our encoders as well as that of the bt-encoder can be easily improved by entropy-encoding the tokens. our purpose is to focus only on the construction and update of the dictionary and searching over it, using less memory than the conventional solutions with trees and hash tables. 4.2 comparison with other encoders we encode each file of the two corpora using lzss and compute the total encoding time as well as the average compression ratio, for different configurations of (|dictionary|,|lab|), with “best” compression option and l = |lab| for algorithm 1. table 1 shows the results of these tests on the calgary corpus. our saencoder is faster than bt, except on tests 5 to 7; on test 6 (the gzip-like scenario), bt-encoder is about 3lzma sdk, version 4.65, 3 feb. 2009, available at www.7-zip.org/sdk.html a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc www.data-compression.info www.larsson.dogma.net/research.html www.7-zip.org/sdk.html table 1: amount of memory, total encoding time (in seconds), and average compression ratio (in bpb), for several lengths of (|dictionary|,|lab|) on the calgary corpus, using “best” compression. gzip “fast” leads to time=0.5 and bpb=3.20 while gzip “best” yields time=1.2 and bpb=2.79. the best encoding time is underlined. calgary corpus sa (proposed) bt lzma # |dict.| |lab| msa time bpb mbt time bpb mlzma time bpb 1 2048 1024 25600 2.2 5 77 26636 3 92 5 65 4217856 4 7 2 99 2 4096 1024 41984 2.5 5 40 53260 4 3 4 98 4241408 4 8 2 82 3 4096 2048 50176 2.4 5 75 53260 11 1 5 48 4241408 4 8 2 82 4 8192 2048 82944 3.8 5 49 106508 11 7 4 88 4288512 5 1 2 69 5 16384 256 134144 9 1 4 36 213004 4 5 4 12 4382720 5 2 2 61 6 32768 256 265216 18 4 4 31 425996 5 5 4 08 4571136 4 9 2 54 7 32768 1024 271360 11 1 4 86 425996 7 5 4 40 4571136 4 9 2 54 8 32768 2048 279552 9.5 5 16 425996 15 8 4 57 4571136 4 9 2 54 table 2: amount of memory, total encoding time (in seconds) and average compression ratio (in bpb), for several lengths of (|dictionary|,|lab|) on the silesia corpus, using “best” compression. gzip “fast” obtains time=19.5 and bpb=3.32 while gzip “best” does time=74.4 and bpb=2.98. the best encoding time is underlined. silesia corpus sa (proposed) bt lzma # |dict.| |lab| msa time bpb mbt time bpb mlzma time bpb 1 2048 1024 25600 118.7 5 66 26636 249 5 5 65 4217856 333 53 3 05 2 4096 1024 41984 116.9 5 41 53260 303 4 5 25 4241408 349 05 2 90 3 4096 2048 50176 112.9 5 68 53260 694 9 5 63 4241408 349 05 2 90 4 8192 2048 82944 143.4 5 44 106508 668 9 5 27 4288512 356 77 2 76 5 16384 256 134144 319 1 4 55 213004 254.6 4 44 4382720 366 47 2 62 6 32768 256 265216 542 7 4 41 425996 318.1 4 31 4571136 356 34 2 52 7 32768 1024 271360 322.2 4 80 425996 382 6 4 64 4571136 356 34 2 52 8 32768 2048 279552 302.3 5 02 425996 979 8 4 81 4571136 356 34 2 52 3.5 times faster than sa. table 2 shows the results for the silesia corpus. in these tests, the sa-encoder is the fastest except on tests 5 and 6. on test 3, the saencoder is about 5 times faster than the bt-encoder, achieving about the same compression ratio. notice that that for the bt and lzma encoders, the amount of memory only depends on the length of the dictionary. for our sa-encoder the amount of memory for the encoder data structures also depends on the length of the lab, due to the use of the ps and i arrays, as given by (2). figure 6 shows the performance measure time× memory on the encoding of the calgary and silesia corpora, on the tests shown on tables 1 and 2, for sa and bt-encoders, including gzip “best” test results for comparison. regarding the calgary corpus test results, the sa-encoder has better performance than bt-encoder on tests 1 to 4 and 8; on silesia corpus, this happens on all tests except on test 6. for all these encoders searching and updating the dictionary are the most time-consuming tasks. a high compression ratio like those of lzma and gzip can be attained only when we use entropy encoding with appropriate models for the tokens. the sa encoder is faster than the bt encoder, when the lab is not too small. our algorithms (without entropy encoding) are thus positioned in a trade-off between time and memory, that can make them suitable to replace binary trees on lzma or in substring search. the usage of the li array over the sa allows to quickly find the set of substrings that start with a given symbol acting as an accelerator of the encoding process. 5 conclusions in this paper, we have proposed a sliding window update algorithm for lempel-ziv compression based on suffix arrays, improving on earlier work in terms of encoding time, with similar low memory requirements. the proposed algorithm uses an auxiliary array as an accelerator to the encoding procedure, as well as a fast update of the dictionary based on two suffix arrays. it allows a priori computing the exact amount of memory necessary for the encoder data structures without any waste of memory; usually this may not be the case when using (binary/suffix) trees. we have compared our algorithm on standard corpora against tree-based encoders, including gzip and lzma. the experimental tests showed that our ena.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 x 10 6 gzip best # test time x memory on calgary corpus sa bt 1 2 3 4 5 6 7 8 0 0 5 1 1 5 2 2 5 3 3 5 4 x 10 8 gzip best # test time x memory on silesia corpus sa bt figure 6: time × memory performance measure for sa and bt on the calgary and silesia corpus, on the 8 encoding tests of tables 1 and 2. we include gzip “best” performance for comparison. coders always occupy less memory than tree-based encoders; moreover, in some (typical) compression settings the sa-encoders are also faster than treebased encoders. the position of the proposed algorithm in the time-memory tradeoff makes it suitable as a replacement of trees and hash tables, for some compression settings. these compression settings include all the situations in which the length of the lookahead-buffer window is not too small, as compared to the length of the dictionary. references [1] c. constantinescu, j. pieper, and t. li. block size optimization in deduplication systems. in dcc ’09: proc. of the ieee conference on data compression, page 442, washington, dc, usa, 2009. [2] m. crochemore, l. ilie, and w. smyth. a simple algorithm for computing the lempel-ziv factorization. in dcc ’08: proc. of the ieee conference on data compression, pages 482–488, washington, dc, usa, march 2008. ieee computer society. [3] a. ferreira, a. oliveira, and m. figueiredo. on the use of suffix arrays for memory-efficient lempel-ziv data compression. in dcc ’09: proc. of the ieee conference on data compression, page 444, washington, dc, usa, march 2009. ieee computer society. [4] m. fiala and j. holub. dca using suffix arrays. in dcc ’08: proc. of the ieee conference on data compression, page 516, washington, dc, usa, march 2008. ieee computer society. [5] gaston h. gonnet, ricardo a. baeza-yates, and tim snider. new indices for text: pat trees and pat arrays. information retrieval: data structures and algorithms, pages 66–82, 1992. [6] ulrich gräf. sorted sliding window compression. in dcc ’99: proc. of the ieee conference on data compression, page 527, washington, dc, usa, 1999. [7] d. gusfield. algorithms on strings, trees and sequences. cambridge university press, 1997. [8] j. karkainen, p. sanders, and s. burkhardt. linear work suffix array construction. journal of the acm, 53(6):918–936, 2006. [9] n. larsson. structures of string matching and data compression. phd thesis, department of computer science, lund university, sweden, september 1999. [10] u. manber and g. myers. suffix arrays: a new method for on-line string searches. siam journal on computing, 22(5):935–948, october 1993. [11] m. nelson and j. gailly. the data compression book. m & t books, new york, 2nd edition, 1995. [12] g. nong, s. zhang, and w. chan. linear suffix array construction by almost pure induced-sorting. in dcc ’09: proc. of the ieee conference on data compression, pages 193–202, march 2009. [13] michael rodeh, vaughan pratt, and shimon even. linear algorithm for data compression via string matching. j. acm, 28(1):16–24, 1981. [14] d. salomon. data compression the complete reference. springer-verlag london ltd, london, fourth edition, januray 2007. [15] m. salson, t. lecroq, m. léonard, and l. mouchard. dynamic extended suffix arrays. journal of discrete algorithms, in press, corrected proof, 2009. [16] j. storer and t. szymanski. data compression via textual substitution. journal of acm, 29(4):928–951, october 1982. [17] e. ukkonen. on-line construction of suffix trees. algorithmica, 14(3):249–260, 1995. [18] s. zhang and g. nong. fast and space efficient linear suffix array construction. in dcc ’08: proc. of the ieee conference on data compression, page 553, washington, dc, usa, march 2008. ieee computer society. [19] j. ziv and a. lempel. a universal algorithm for sequential data compression. ieee transactions on information theory, it-23(3):337–343, may 1977. a.ferreira et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc introduction our contribution lempel-ziv basics decoding procedures using a suffix array sliding window update algorithm accelerated encoder proposed algorithm experimental results experimental setup comparison with other encoders conclusions apc speech coding techniques applied to ecg signals adaptive predictive coding speech coding techniques applied to electrocardiogram signals d. silvaa, g. martinsa, a. lourençoab, c. menesesac aelectronic, telecommunication and computer department, isel, portugal bcardioid, portugal cm2a – multimedia and machine learning group 9cdanielsilva@gmail.com guimbmartins@hotmail.com alourenco@deetc.isel.ipl.pt cmeneses@deetc.isel.ipl.pt abstract — this paper describes a lossy ecg signal coder with an adaptive predictive coding scheme initially proposed for speech coders. the predictors include linear predictive coding that takes advantage of the correlation between consecutive samples and long-term predictor that takes advantage of the signal quasi-periodicity. the prediction residue, with less dynamic range and therefore able to be encoded with less bits than the original, is transmitted sample by sample. the prediction coefficients and the amplitude of the residue are transmitted once for each heartbeat, with a negligible number of bits compared to the total bit rate. the long-term predictor is shown to obtain reliable performance when the heart rate does not change rapidly. linear predictive coding, on the contrary, is more reliable and presents better prediction gain. the best developed coder uses double prediction and with 45% compression ratio allows a prediction gain of 24.8 db. keywords: ecg, speech, adaptive predictive coding, linear prediction coding, long-term prediction, signal to noise ratio. i. introduction signal coding is intended to decrease the signal representation binary rate. the main applications are to transmit or store signals using a low bit rate, which leads to the use of cheaper and lower power modems and less storage memory. in addition to the traditional electrocardiogram (ecg) medical applications, ecg signal applications [1-2] are increasingly emerging on devices such as smartwatches, sport watches or chest straps, not only to measure heart rate but also to check for fatigue, to predict heart failure, or authenticate the user. in an increasing number of applications, the ecg signal must be transmitted and stored to smartphones or to the cloud. ecg signals are also stored in hospital information systems (his), in the patient history. in all these cases, there is a need for transmission and storage, for which decreasing bit rate is an important contribution. there are two main methods of signal coding: lossless and lossy methods [3]. lossless methods obtain an exact reconstruction of the signal, but the compression ratio is low. lossy methods can achieve bigger compression ratio but do not represent the exact original signal. the main goal of lossy methods is to achieve high compression ratio without compromising the quality. for ecg this paper reflects the work for the final-year project of a 3-year cycle of studies in electronic, telecommunications and computer engineering at isel (instituto superior de engenharia de lisboa). signals, this corresponds to maintain the diagnostic capability of the original signal. speech signal coding has a history [4-10] of decades and is a very mature field. depending on the applications and the tradeoff between compression ratio and quality requirements, it is possible to find standard coders with bit rates between 800 bit/s [8] and 64 kbit/s [9,10]. adaptive predictive coding (apc) [11] is a low complex and high-quality speech coder that is a good compromise between quality and bit rate. the apc coder predicts the speech signal taking advantage of the almost periodic structure in voiced regions and the high correlation between adjacent samples. only the prediction residue is transmitted sample by sample, reducing the bit rate. given that the ecg and speech signals have in common an almost periodic structure, the long-term predictor used in speech that takes advantage of this characteristic can be applied to ecg signals [12]. at the same time, a small variation in some parts of the ecg signal also reveals a correlation between consecutive samples, capable to predict one sample from the immediately previous (linear predictive coding [13-14]), making it necessary to find out the best prediction order and the prediction capacity. the apc coder can be, therefore, an alternative solution to more traditional ecg coding methods [15-16]. this paper presents the development of an ecg signal lossy coder using the apc speech coder scheme. section ii characterizes the ecg signal. section iii presents the apc coder. section iv presents the proposed method, including the database, the measures to assess the coders performance and the development method. section v presents the results and discussion, including the optimization of each parameter and the all quantized coders for each type of predictor. section vi finishes the paper with the conclusions and directions for future work. ii. ecg signal ecg signals represent the electrical activity of the heart and are recorded by electrodes connected to the body. the signal has a quasi-periodic structure, being each period one heartbeat. the same quasi-periodic structure can be found in voiced regions (produced with vocal folds vibration) of speech i-etc: isel academic journal of electronics, telecommunications and computers vol. 6 , n. 1 (2020) id-5 http://journals.isel.pt mailto:9cdanielsilva@gmail.com mailto:guimbmartins@hotmail.com mailto:alourenco@deetc.isel.ipl.pt mailto:cmeneses@deetc.isel.ipl.pt signals. in each heartbeat it is possible to find 5 well-defined fiducial points, represented by the letters p, q, r, s, t, as in fig. 1. analysis of the waveform between these points and the relative latency time and wave signal allows to evaluate the transients of the electrical stimulus from the auricles to the ventricles, analyse the cardiac rhythm (regular or arrhythmias), evaluate possible hypertrophy of the cardiac cavities and to evaluate signs of deficient irrigation of the heart, for example, in coronary heart disease or ischemic. fig. 1. ecg signal despite the assumption of quasi-periodic structure, ecg signals can have variability between consecutive periods, depending on the subject activity, which will change heart rate, as can be seen in fig. 2, either with changes in amplitude, period and shape. also, in case of heart diseases such as arrhythmias, the quasi-periodic structure is also called into question. fig. 2. non-periodic ecg signal. iii. apc coder the adaptive predictive coding method [11], presented in fig. 3, quantizes in pulse code modulation (pcm) the prediction residue r[n], defined as the difference between the input original signal s[n] and a prediction sp[n] estimated from the last quantized samples sq[n]. in the receiver, the quantized prediction residue rq[n] is added to the prediction to calculate the actual quantized sample. the better the predictor works, the lower the dynamic range of the prediction residue and the better the final quality of the quantized signal. taking advantage of the quasi-stationarity of the signals, the prediction coefficients must be estimated frame by frame and transmitted to the receiver. typically, speech frames are 5 to 30 ms long. fig. 3. apc transmitter. the receiver is embedded in the transmitter. a. lpc prediction linear predictive coding (lpc) [13] takes advantage of the correlation between consecutive samples to predict one sample from a linear combination of past samples, as in equation (1), that translates the equation of a finite impulse response (fir) filter. 𝑠𝑝[𝑛] = −∑ 𝑎𝑘𝑠𝑞 𝑝 𝑘=1 [𝑛 − 𝑘]. (1) the prediction coefficients, 𝑎𝑘, are estimated in order to minimize the prediction residue and have information about the spectral envelope. b. long-term prediction for quasi-periodical signals, as the ecg and speech signals in voiced regions, one entire heartbeat period can be predicted by replication of the previous period. this predictor is known as long-term (lt) predictor as one sample is predicted with a delay of one heartbeat period 𝑇𝑃, equation (2), and not consecutive samples as in lpc prediction. 𝑠𝑝[𝑛] = 𝑎𝑝𝑠𝑞[𝑛 − 𝑇𝑃], (2) where 𝑎𝑝 is the lt prediction coefficient. to estimate the lt period, 𝑇𝑃, maximum autocorrelation or similar methods [17] are normally used in speech analysis. qrs detection [18] can also be used to estimate periodicity. setting always the same initial point in the period is also desirable. align the r peaks can be done with an adaptive threshold comparison and absolute maximum detection. to accommodate period change, the lt period, 𝑇𝑃, must be interpolated/decimated sample by sample in order to time warp the previous period to have the same length as the period to predict. as can be seen in fig. 4, where 12 consecutive heartbeat periods are interpolated to have the same duration, this procedure aligns the pqrst points to improve the prediction. the lt prediction coefficient, 𝑎𝑝 is estimated in order to minimize the prediction error and corresponds to the normalized correlation with delay 𝑇𝑃, between the periods to predict and the interpolated/decimated previous period, receiver q/c c -1 predictor s[n] r[n] rq[n] sq[n] sp[n] – receiverr t synthesis window analysis window s q p d. silva et al. | i-etc, vol. 6, n. 1 (2020) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt fig. 4. interpolated heartbeats. 𝑎𝑝 = 𝑅[𝑛−𝑇𝑃] 𝐸 , (3) on what 𝑅[𝑛 − 𝑇𝑃] is the autocorrelation with delay 𝑇𝑃 and e is the energy of the interpolated/decimated previous period. c. double prediction the lpc residue has the same periodicity as the original signal. therefore, the lt predictor can be applied to this residue, resulting in a double predictor, minimizing even more the dynamic range of the double prediction residue and increasing the quality. fig. 5 shows the complete block diagram of the apc encoder with double prediction. fig. 5. apc with double prediction (lpc+lt). iv. proposed method this section presents the proposed method, including the database, the measures to assess the coders performance and the development method. a. database the massachusetts institute of technology (mit) and the boston's beth israel hospital (now the beth israel deaconess medical center) developed an ecg database, the mit-bih arrhythmia database [19], that contains 48 half-hour excerpts of two-channel ambulatory ecg recordings, obtained from 47 subjects studied by the bih arrhythmia laboratory. this database is available since 1980 and is one of the reference databases in the field. the signals are sampled at 360 hz and stored in pcm with 11 bits per sample. from this database, a set of 19 signals are chosen, of which 10 seconds are extracted to develop and test the ecg coder. all the signals are normalized in amplitude. only integer periods from the second heartbeat period are considered to the quality measure, as the first period cannot be predicted with lt predictors. these correspond to an average of 8.2 seconds and 10.2 heartbeat periods per signal, in a total of 155.5 seconds and 193 heartbeat periods. the average heart rate is 74 heartbeats per minute. b. quality assessment to assess the quality of the coder, the signal to noise ratio, (snr) given by the ratio of a reference or original signal power p to the noise power n, in decibels as in equation (4), is used, 𝑆𝑁𝑅𝑑𝐵 = 10𝑙𝑜𝑔10 ( 𝑃 𝑁 ). (4) in coders assessment, the noise takes origin in samples and parameters quantization, and the snr is denominated as quantization snr. in this study, the quantization noise is the difference between the 11-bit pcm signal, taken as the reference, and the output of the coder. the quantization snr average between the 19 signals of the database is estimated for each coder and is assumed to be the quantization snr of that coder. for the apc coder, the increase in the signal to noise ratio in relation to the pcm direct coding or reference coder, denominated the prediction gain, is, 𝐺𝑝𝑑𝐵 = 10𝑙𝑜𝑔10 ( 𝑉2 𝑉1 2) , (5) where 𝑉 is the maximum quantization value in pcm (reference coder) and 𝑉1 is the maximum quantization value in apc. the better the predictor, the lower the maximum quantization value and the greater the prediction gain. c. development method taking advantage of the quasi-stationarity of the signals, lpc can be transmitted frame by frame. the length of the synthesis window (frame length) is chosen to be a heartbeat period between r peaks (fig. 1), resulting in a variable bit rate coder. the analysis window to estimate the lpc coefficients is extended in one third. the quantization of the prediction residue consumes most of the quantization bits, since this signal is transmitted sample by sample and not by heartbeat period, practically defining the final bit rate. the number of bits to quantize the prediction residue is fixed and a pcm coder (equivalent to an apc coder without prediction) is assessed and taken as a coder reference. with the same bits per sample to quantize the prediction residue of the reference coder, but without any further quantization, each predictor (lt, lpc and lt+lpc double predictor) is evaluated and adjusted based on the prediction gain, defined as the snr difference in relation to the reference coder. after adjusting the predictors, each quantizer is trained to minimize the quantization error and the number of bits to quantize the predictor parameters (prediction residue amplitude, lt coefficient and lpc coefficients) are tune individually, based in the snr loss. to train each quantizer [20], 10 ecg signals are used, with no quantization of any parameter beyond the prediction error. the quantizers are trained from the corresponding non quantized values distributions. the remaining 9 signals are used to test if the trained quantizer is generalizing. receiver q/c c -1 lpc s[n] r n] rq n] sq[n] – receiver lt rq n] rq n-tp] d. silva et al. | i-etc, vol. 6, n. 1 (2020) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt v. results and discussion this section follows the development method from section iv, starting by to define the reference coder, with which all results will be compared, and all the parameters optimized. then the best lpc order is found, followed by the individual optimization of the quantizers. this section ends with the presentation and discussion of the complete coder. a. reference coder a pcm coder with 6 bits per sample was taken as a reference, corresponding to 2160 bit/s, a 45% reduction in relation to the 11-bit original pcm coder. to define a reference in terms of quality, the 19 signals from the database were re-quantized in pcm, corresponding to the removal of the two predictors. the average quantization snr obtained in the 19 ecg signals of the database is 30.9 db. for each additional quantization bit, a gain of 6.02 db is obtained, but the bit rate also increases in 360 bit/s. b. heartbeat period estimation and quantization the alignment of the r peaks is achieved with an adaptive threshold comparison of 0.3 and absolute maximum detection. the heartbeat period, 𝑇𝑃, is estimated from the time between consecutive r peaks. a minimum heart rate of 30 beats per minute and a maximum of 232 beats per minute are assumed. at a sample rate of 360 hz, this corresponds to 720 to 93 samples per heartbeat. the range of values is 720-93 = 627, requiring 10 code bits for each heartbeat, assuming that the heartbeat period is a multiple of the sampling period and does not suffer from additional quantization error. the maximum bit rate added due to this parameter is 39 bit/s for a heart rate of 232 heartbeats per minute. on average, for the 19 signals, the heart rate is 74 heartbeats per minute, corresponding to 1.24 bit/s for each coding bit per heartbeat period. using 10 bits to code each heartbeat, 12.4 bit/s are added. the lt prediction coefficient, 𝑎𝑝, and lpc prediction coefficients, 𝑎𝑘, are estimated per heartbeat. the maximum quantization value, 𝑉1, which depends on the prediction gain, 𝐺𝑝, is also estimated per heartbeat and transmitted to the receiver. this value cannot be constant, as a value that is too low implies a slope overload and a value that is too high implies a decrease in the prediction gain. c. lpc order typically, order 10 is used in speech coders, a good tradeoff between spectral envelope definition and bit rate, as these coefficients must be transmitted to the receiver. one question to be answered when using ecg signals is which order of prediction to use, assuming this tradeoff. table i presents the prediction gain compared to the reference coder (30.9 db), for different orders of the lpc predictor, without quantization of the coefficients. the covariance method [13] to estimate the lpc coefficients is chosen since it can achieve better results than the more traditional autocorrelation method. the order 3 of the lpc is chosen since the prediction gain increases considerably up to that order. from that order, the increase in the order of the lpc only slightly increases the snr but increases the complexity and the bit rate. table i prediction gain with different lpc orders lpc order snr [db] gp [db] 1 42.6 11.7 2 50.1 19.2 3 53.2 22.3 4 53.7 22.8 5 53.7 22.8 10 54.0 23.1 d. prediction residue quantization as the prediction coefficients are transmitted per heartbeat period, prediction residue quantization bits correspond to most of the transmitted bits. table ii presents the snr for the different predictors (lt, 3rd order lpc and lt+lpc), where no parameters are quantized beyond the prediction residue. for lt single prediction, as presented in fig. 6, the prediction coefficient distribution is located around 1, indicating that consecutive periods have high similarity. using a constant coefficient 𝑎𝑝 = 1, the snr even increases 0.4 db, so this value is adopted as it does not need to be transmitted to the receiver. table ii snr for 6-bit quantizers pcm lt lt ap=1 lpc lt+lpc lt+lpc ap=0.6 snr [db] 30.9 43.2 43.6 53.2 55.8 55.8 fig. 6. prediction values distribution for lt single prediction. for the double prediction, as presented in fig. 7, the prediction coefficient distribution is not located around a value. using a constant value of 0.6, as presented in table ii, the snr is the same, so this value is adopted as it does not need to be transmitted to the receiver. the best snr with only the quantization of the prediction residue is achieved with lt+lpc double prediction, obtains 55.8 db, a prediction gain of 25 db than the reference coder. fig. 7. lt prediction coefficient distribution for double prediction. d. silva et al. | i-etc, vol. 6, n. 1 (2020) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt e. prediction residue amplitude quantization the maximum quantization value, 𝑉1, is an important parameter to be set. too high implies a decrease in the snr, and too low implies slope overload. to solve this problem, the maximum quantization value is estimated and transmitted per heartbeat period. fig. 8, 9 and 10 presents the maximum quantization value distributions for lt single prediction, lt+lpc double prediction and lpc prediction, respectively. as it can be seen, values from lt single prediction are higher than for lpc prediction and the lowest are from lt+lpc double prediction, in line with the increase in the snr. table iii presents the quantization loss when using different number of bits to code the prediction residue amplitude, compared to results from table ii. table iii snr quantization loss for v1 quantization # of bits 5 6 7 lt snr loss [db] 0.4 0.1 0.1 lpc snr loss [db] 1.6 0.3 0.1 lt+lpc snr loss [db] 1.1 0.5 0.1 fig. 8. v1 distribution for lt single prediction. fig. 9. v1 distribution for lpc prediction. fig. 10. v1 distribution for lt+lpc double prediction. to code the prediction residue amplitude, considering the tradeoff between quality and bit rate, 7 bits are chosen, as it corresponds to only 0.1 db of snr loss. this corresponds to 8.7 bit/s at an average heart rate of 74 heartbeats per minute. f. lpc coefficients quantization the direct transmission of the lpc coefficients is not recommended as the quantization error can change significantly the spectral envelope or turn the filter instable. to solve this problem, the use of a line spectrum pair (lsp) transformation [21], widely used in speech coders [4][6-7][22], guarantees the stability and minimizes the sensitivity of the filter. after the lsp transformation, the lsp parameters must be quantized. lsp values are in ascending order and between 0 and 0.5 (0.5 corresponds to /2 radians or half of the sample frequency). the stability of the filter is guaranteed by imposing that the lsp coefficients maintain the ascending order after quantization. since the lsp parameters are not uniformly distributed, as presented in fig. 11, 12 and 13 for order 3, respectively for the first, second and third coefficients, the quantizers for each coefficient must be trained to minimize the quantization error. fig. 11. first lsp coefficient distribution fig. 12. second lsp coefficient distribution fig. 13. third lsp coefficient distribution table iv shows the snr loss using 3 bits per coefficient. the degradation in the entire database is 0.1 db for lpc prediction and lt+lpc double prediction. these values are used in the final coder corresponding to 11.1 bit/s. table iv snr quantization loss for lsp quantization training set test set entire database lpc snr loss [db] 0.06 0.15 0.11 lt+lpc snr loss [db] 0.13 0.02 0.08 g. full quantized coders table v presents the bit rate distribution for each quantized parameter. it is considered an average heart rate of 74 heartbeats per second. d. silva et al. | i-etc, vol. 6, n. 1 (2020) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt table v final bits assignment parameter [bit] bit rate [bit/s] prediction residue 6 2160 heartbeat period 10 12.4 prediction residue amplitude 7 8.7 lpc/lsp (3+3+3) 9 11.1 total 2192 table vi presents the quantized snr, the prediction gain, the bit rate and the compression ratio for the full quantized coders, in the 19 ecg signals from the database. table vi final results snr [db] prediction gain [db] mean bit rate [bit/s] compression ratio [%] pcm 30.9 ---2160 45.5 lt 43.5 12.6 2181 44.9 lpc 53.0 22.1 2192 44.6 lt+lpc 55.7 24.8 2192 44.6 comparing the results from table vi and table ii, the snr quantization loss due to quantization is 0.1, 0.2 and 0.1 db, respectively to the lt, lpc and lt+lpc quantizer. the lt+lpc double predictor presents a prediction gain of only 2.7 db compared to the lpc predictor. this is because the first predictor, the lpc, decorrelates is residue and, when there is physical activity of the subject, the heart rate is constantly altered and the assumption of the periodicity of the signal is no longer valid. the latter reason also applies for the lt single predictor, where the prediction gain is also lower, 12.6 db, compared to the lpc predictor, with 22.1 db prediction gain. as already pointed out, the main origin of the bit rate is the coding of the prediction residue, transmitted sample by sample. the other parameters, transmitted by heartbeat period, correspond to only 32 bit/s out of 2192 bit/s total bit rate. the difference between the compression ratio with the different predictors is less than 1%, with a compression ratio of about 45%. vi. conclusions this paper describes an ecg signal lossy coder with an adaptive predictive coding scheme initially proposed for speech coders. it was concluded that the lt predictor is the worst predictor with a gain of only 12.6 db, due to variations in heart rate that occur during physical activity. after the lpc predictor, the lt predictor even has a lower gain of only 2.7 db. the variation in the cardiac rhythm and the 5 distinct parts of the heartbeat explains the low third order of the lpc predictor for the ecg signal, comparing with the order 10 for speech signals. the quantization loss is less than 0.2% for all the predictors, a value negligible in the final snr. as expected, the best predictor is the lt+lpc double predictor with a prediction gain of 24.8 db, a total of 55.7 db and a compression ratio of 44.6%. since for each bit per sample in pcm a gain of 6.02 db is obtained, 4 bits are needed to have the same quality just re-quantizing in pcm, but this procedure corresponds to 66% increase in the bitrate and a compression ratio of only 9%. the ecg signal can be divided into two zones. one corresponds to the signal points belonging to the qrs complex. the other corresponds to the points between peak s of a complex and next peak q. as a future work, it is suggested to implement the division of the lpc in these two zones, as it can significantly improve the quality of the coding. one of the reasons that the lt predictor does not produce a high-quality gain is because the cardiac cycle period does not coincide with a multiple of the sampling period. it is suggested to solve this problem through fractional pitch techniques already used in speech signal coding. in addition, further testing will be performed with ecg acquired using less intrusive settings, as off-the-person approaches, where ecg is acquired while the user is interacting only with the hands with an ecg sensing device. references [1] h. silva, c. carreiras, a. lourenço, a. l. n. fred, r. césar das neves, r. ferreira, off-the person electrocardiography: performance assessment and clinical correlation, health and technology, vol. 4, numb. 4, 2015 [2] j. ribeiro pinto, j. s. cardoso and a. lourenço, evolution, current challenges, and future possibilities in ecg biometrics, in ieee access, vol. 6, pp. 34746-34776, 2018.doi: 10.1109/access.2018.2849870 [3] g. vijayvargiya, s. silakari, r. pandey, a survey: various techniques of image compression, (ijcsis) international journal of computer science and information security, vol. 11, no. 10, october 2013. [4] j. r. crosmer, t. p. barnwell, a low bit rate segment vocoder based on line spectrum pairs, proc. of the int. conf. acoust., speech and signal processing, pp. 240-243, 1985. [5] i.m. trancoso, j. s. marques, c. meneses ribeiro, two solutions for speech coding at 4.8-9.6 kbps, speech communications journal, vol 9 5/6, pp.389-400, december 1990. [6] r. salami, c. laflamme, b. bessete, j-p adoul, itu-t g.729 annex a: reduced complexity 8 kb/s cs-acelp codec for digital simultaneous voice and data, ieee communication magazine, 1997 [7] j. p. campbell, jr. the dod 4.8 kbps standard, advances in speech coding, ed. b. atal, v. cuperman and a. gersho, kluwer academic publishers, 1990. [8] b. mouy, p. de la noue, g. goudezeune, nato stanag 4479: a standard for an 800 bps vocoder and channel coding in hf-eccm system, proc. of the int. conf. acoust., speech and signal processing, pp.480-483, 1995. [9] https://www.itu.int/rec/t-rec-g.711 [accessed on october 2020] d. silva et al. | i-etc, vol. 6, n. 1 (2020) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt https://www.itu.int/rec/t-rec-g.711 [10] https://www.itu.int/rec/t-rec-g.722 [accessed on october 2020] [11] b. s. atal, m. r. schroeder, adaptive predictive coding of speech signals, bell system technical journal, vol. 49, pp.1973-1986, october 1970. [12] y. zigel, a. cohen, a. katz, ecg signal compression using analysis by synthesis coding, in ieee transactions on biomedical engineering, volume: 47 , issue: 10 , oct. 2000. [13] j. makhol, linear prediction: a tutorial review, proc. of the ieee, vol. 63, nº 4, 1975. [14] justin l. c. loong, khazaimatol s. subari, rosli besar, muhammad k. abdullah, a new approach to ecg biometric systems: a comparitive study between lpc and wpd systems, world academy of science, engineering and technology, international journal of medical, health, biomedical, bioengineering and pharmaceutical engineering, vol. 4, 2010 [15] l. rebollo-neira, effective high compression of ecg signals at low level distortion, sci rep 9, 4564, 2019. [16] m. elgendi, a. mohamed, r. ward, efficient ecg compression and qrs detection for e-health applications, sci rep 7, 459, 2017. [17] l. r. rabiner, on the use of autocorrelation analysis for pitch detection, ieee trans. on acoustics, speech and signal processing, vol. assp-25, n°1, february, 1977. [18] j. pan e w. j. tompkins, a real-time qrs detection algorithm, ieee transactions on biomedical engineering, vol.bme-32, nº 3, pp. 230 236, 1985. [19] pysionet mit-bih arrhythmia database [online]. https://physionet.org/physiobank/database/mitdb/. [accessed on october 2020]. [20] j. max, quantizing for minimum distortion, ire trans. inform. theory, vol it-6, pp. 7-12, 1960. [21] f. soong, b. juang, line spectrum pair (lsp) and speech data compression, proc. of the int. conf. acoust., speech and signal processing, 1.10.1-1.10.4, 1984. [22] g. s. kang, l. j. fransen, application of line-spectrum pairs to low-bit-rate speech encoders, proc. of the int. conf. acoust., speech and signal processing, pp.244-247, 1985. d. silva et al. | i-etc, vol. 6, n. 1 (2020) id-5 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt https://www.itu.int/rec/t-rec-g.722 https://ieeexplore.ieee.org/author/37394190600 https://ieeexplore.ieee.org/author/37285890200 https://ieeexplore.ieee.org/author/37395646700 https://ieeexplore.ieee.org/xpl/recentissue.jsp?punumber=10 https://ieeexplore.ieee.org/xpl/recentissue.jsp?punumber=10 https://ieeexplore.ieee.org/xpl/tocresult.jsp?isnumber=18878 simulating long term evolution self-optimizing based networks simulating long term evolution self-optimizing based networks marco carvalho1,2, pedro vieira1,2, 1área departamental de engenharia de electrónica e telecomunicações e de computadores (adeetc), instituto superior de engenharia de lisboa (isel), lisboa, portugal, e-mail: 35832@alunos.isel.pt, 2instituto de telecomunicações (it), lisboa, portugal, e-mail: pvieira@deetc.isel.pt keywords: wireless communications, lte, son, automatic neighbor relation, handover optimization abstract: with the first 3rd generation partnership project (3gpp) long term evolution (lte) networks being deployed more complexity is added to current existing cellular mobile networks and more capital (capex) and operational (opex) effort will be needed. in addition, the rising demand of users for new services and higher data rates demands more efficiency from operators. for this matter, 3gpp release 8 as introduced the self-organizing network (son) concept, a set of self-configuration, self-optimizing and self-healing functions that allow the automation of labor-intensive tasks, reducing operational and capital costs. while requirements on cutting operational expenditure remain, operators still remain skeptical with the efficiency of these functions. in this paper, physical cell identity (pci) conflict detection and resolution, automatic neighbor relation (anr) and automatic handover parameter optimization (hpo) functions are proposed as part of a simulator for lte son based networks. based on user defined inputs, these functions allow operators to closely predict and gather optimal policy input values for son algorithms, while maintaining desirable network performance. based on a real network scenario, results show simulator’s clear benefit when compared with other proposals. 1 introduction with the arrival of the 4th generation standards, especially the 3rd generation partnership project (3gpp) long term evolution (lte), more complexity is added to the current existing networks. the rising number of parameters from multiple coexisting standards, in most cases from different suppliers, combined with the increasing demand of 3rd party services, demands more management effort from mobile network operators. introduced since 3gpp release 8 specifications, the self organizing networks (son) concept aims to reduce most of common planning, optimization and operational tasks through automated mechanism such as self-configuration, self-optimization and self-healing mechanisms, reducing operator’s operational and capital costs (opex/capex). with first commercial lte networks being deployed, operators question the performance and reliability of these automated functions in their current networks. for this matter some proposals have been made to evaluate network performance under these circumstances, [1], [2]. despite the variety of features, these simulators don’t allow, in most cases, the recreation of a typical operator network configuration, putting into question the reliability and usefulness of the obtained results in a real life scenario. in this paper, a new simulation tool is proposed for lte son based networks. this simulator aims to evaluate physical cell identity (pci) conflict detection and resolution, automatic neighbor relation (anr) and automatic handover parameter optimization (hpo) functions based on customizable scenarios using user defined network inputs such as geographical positioning of evolvednodebs (enodebs), antennas orientation or radio propagation environment characterization. the rest of the paper is organized as follows. section 2 presents a brief overview of son current state-ofthe-art. the implemented algorithms are presented in section 3 followed by a brief description of the simulator in section 4. a simulation scenario and performance results are set in section 5 and 6, respectively. finally, section 7 presents the overall conclusions. 2 self-organizing network concept started by ngmn (next generation mobile networks) and later included by 3gpp in the i-etc: isel academic journal of electronics, telecommunications and computers cetc2011 issue, vol. 2, n. 1 (2013) id-8 http://journals.isel.pt/index.php/iajetc evolved umts terrestrial radio access network (e-utran) specification process since release 8, son is the key driver to improve operators o&m (operations & maintenance). by automating most of common planning, optimization and operational tasks, son aims to reduce capital and operational costs by reducing time-consuming manual processes in network management. reference [3] establish the guidelines needed to create autonomous functions that can be organized essentially in the following groups, in the format of user-cases: • self-configuration; • self-optimization; • self-healing. self-configuration compromises the automation of tasks related to the deployment of new evolved nodeb (enb). the self-configuration process works in a pre-operational state, starting from the moment the enb is powered on until the rf transmitter is switched on. this process evolves the transport link detection, connection with the core network elements, download and upgrade of software version, setup of initial configuration parameters, including neighbour relations, self-test and finally rf transmitter activation. self-optimization is the process in which user equipment (ue) and enb measurements are used to auto-tune the network. this is an operational state process which starts when the rf interface is switched on. this autonomous optimization allows a more fast and accurate resolution of network problems. the optimizations tasks within this function are: • neighbor list optimization; • coverage and capacity optimization; • mobility load balancing optimization; • radio access channel (rach) optimization; • inter-cell interference coordination. finally, self-healing functions aim to automatically detect and localize failures in network elements, and take the appropriate decisions, e.g. load balance traffic in case of high traffic element failure, reduce cell power in case of high temperature failure or fallback to previous software version in case of errors during network software update. 3 son functions 3.1 pci conflict detection and resolution the pci is the physical identification of a cell contained in sch (synchronization channel). there are totally 504 unique pcis defined in the evolved umts terrestrial radio access network (eutran) spread over 168 designated physical cell identity groups, where each one contains three unique identifiers. each pci plays an important role in that each allows synchronization signals (including primary synchronization signal (pss) and secondary synchronization signal (sss)) and the reference signals (including cell-specific reference signal (crs) and ue-specific reference signal (urs)) to be generated and distinguish by ues. in addition, is also from the pci that scrambling sequences of most of the physical channel such as pbch, pcfich, phich, pdcch, pdsch, and pucch are generated, [4]. as earlier stated, one of the main goals of son is to automate most of common planning, optimization and operational tasks. as in utran, with the allocation of sc (scrambling codes), in e-utran, pci assignment is a task that requires great care. bad pci allocation may lead to interference that can reduce network performance due to call drop rise. replacing manual pci planning methods by automatic functions may pose some risks. depending on propagation environment or terrain morphology, algorithm may fail to correctly predict interference. for that matter, advanced pci conflict detection and resolution mechanisms are needed to maximize network availability and reduce interference probability. to provide pci conflict detection and resolution mechanism two algorithms were implemented, respectively. the first algorithm is based on the one proposed in [5] witch, through joint collaboration with the anr function (later explained) and using ue measurements reports, detects pci conflicts. in addition to the algorithm presented in [5], periodic measurements reports are also considered to increase pci conflict detection probability. the work presented in [5] also presents a mechanism based on transmission gaps where ues can measure surrounding neighbours while serving cell is not transmitting and thus increase pci conflict detection probability. this latter mechanism was not considered because it is not clear as to how the ue can distinguish serving cell signal from others when a near cell contains the same pci as the serving cell. anytime a cell receives a measurement report the algorithm presented in figure 1 is set. in figure 1, if local cell nrt (neighbour relation table) is not empty, the algorithm checks if reported ecgi exists. m carvalho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc if exists, the algorithm checks if the corresponding nr (neighbour relation) pci is different from reported cell pci. if different the algorithm proceeds to the pci resolution process. figure 1: pci conflict detection algorithm. once a pci conflict is detected, a designated resolution algorithm takes place. in this algorithm, a new pci is assigned to the conflicting cell. as also described in [5], to avoid new conflicts, a compilation of a set of locally conflicting pcis is made by retrieving the pcis of the neighbours and neighbours of neighbours to the selected conflicting cell. in figure 2, if a cell has detected a pci conflict, the resolution algorithm starts by contacting the conflicting cell. once the conflicting cells are notified about the pci conflict, a new pci generation process begins. this process starts by gathering all neighbour cells pci. for each neighbouring cell, pcis are gathered from their neighbouring cells. once all neighbouring pci are gathered, a new pci is generated excluding the ones previously detected. this new pci may be generated taking into account a pre-defined list of available pci, provided during configuration download, or can be locally generated as follows: ��� = 3��� ( ) +��� ( ) (1) where 3��� ( ) is the physical cell identity group raging from 0 to 167 and ��� ( ) the physical layer identity raging from 0 to 2. figure 2: pci resolution algorithm. 3.2 anr neighbour relation creation is one of the most intensive and important tasks during network planning. typically neighbour planning is based on tools that, using drive test and networks information can predict signal level and thus help in neighbour’s definition. by automatically define neighbours based on ue reports, better mobility can be achieved since missing neighbour situations will be avoided and thus post-integration optimization is no longer needed. implemented at cell level, the anr function aims to automatically create neighbor relations based on ue triggered or periodic measurements reports. based on [3] an anr algorithm was implemented to provide automatic neighbour creation. in figure 3, once a measurement report event condition is fulfilled, most likely a3 or a5 as stated in [3], an intra-frequency handover is triggered. once the enodeb receives ue measurement report, a coverage evaluation takes places and a handover m carvalho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc decision is made. this handover decision is based on rsrp (reference signal received power) and rsrq (reference signal received quality) measurements from ue neighbouring cells. if enodeb can identify a candidate neighbouring cell from whose coverage conditions are better, the handover proceeds. in case handover is started, the anr function request the ue to get the candidate cell ecgi (e-utran cell global identifier). having received the candidate cell ecgi, if this already exists in nrt, a pci conflict detection procedure, as stated in section 3.1, takes place. once the pci conflict detection procedure is made an x2 connection to the target enodeb is made in order to exchange nrts. during nrt exchange, beyond the resource reservation, pci conflict detection is also made. finally, local nrt is updated and handover procedure is terminated by sending a connection reconfiguration message to the ue. start anr procedure end anr procedure yes no measurement report received ? ecgi exists ? ho decision exchange nrt proceed ho? yes no no yes pci conflict detection reconfigure ue establish x2 interface update nrt request target cell ecgi figure 3: anr algorithm. 3.3 handover optimization as mentioned, one of the son main goals relies on the automatic optimization of radio parameters. one of these functions is the automatic handover (ho) parameter optimization. based on performance indicators, this function allows each cell to readjust the hysteresis and time-to-trigger (ttt) parameters, avoiding failed handovers, dropped calls or ping-pong effect. in order to keep continuity of communication, enhanced capacity and good user perceived qos, evaluation methodologies are needed. therefore, handover performance indicators (hpi) must be calculated. in this simulator the implemented optimization algorithm is based on the one described at [6] which takes into account the ping-pong handover performance indicator (hpi���) to optimize hysteresis and ttt values. the hpi��� measures the event rate where a call is handed over from cell a to cell b and is handed back to cell a in period of time that is less than a designated critical time (t����). it is calculated as the number of pingpong handovers (n����) divided by itself plus the number of non-ping-pong handovers (n�����) and the number of failed handovers (n������), ��!"" = #$%&& #$%&&'#$%(&&'#$%)*+, . (2) using cell measured hpi���, this algorithm continuously search for the best values at any time. additionally, the ue type of traffic and speed is taken into account for a more accurate handover decision. 4 system-level simulator overview 4.1 simulator overview the system-level simulator provided by [1] is a non-commercial open source simulator available for academic research. this simulator allows the study of various aspects related to cell planning, scheduling and interference. developed in matlab®, this object-oriented programming (oop) simulator is well organized, presenting a good understandable and maintainable structure that suits for development and testing of new algorithm and functions. for this matter, it was decided to adopt this simulator as the basic foundation for the developed lte son simulator. 4.2 additional features by implementing additional functions together with a graphical user interface (gui), the system-level simulator is able to simulate a user customized lte son network. among others parameters, the proposed simulator allows the specification of the following parameters: m carvalho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc � enodeb geographic location, and height. � antenna orientation, radiation pattern, mechanical down tilt (mdt) and electrical down tilt (edt). � pathloss model including modified cost 231 [7] and 3gpp tr36.942 [8]. figure 4 shows a simulator network scenario example. in this scenario, a georeferenced topographic map is loaded. each enodeb is manually placed and configured. ues are pseudo randomly placed, and their direction of movement is also defined in a pseudo-randomly matter. figure 4: simulation scenario example (olhão, portugal). 5 simulation 5.1 simulation parameter to evaluate the simulator performance, a real network scenario is configured. table 1 presents the general parameters taken into account in each simulation. parameter value propagation enviroment urban area frequency 2,6 ghz cellular layout 9 cell sites, 3 sectors per site propagation model urban macro, [7] (ts 36.942) shadow fading log-normal, µ=0; σ=10 (db) multipath fading 3gpp veha minimum coupling losses (bs↔ue) 70 db (ts 36.942) number of users 270 (10 users per cell) total bs tx power 43 dbm – 5 mhz carrier table 1: simulation parameter the specific parameters of each algorithm are presented in the following sections. the type of traffic remains the same along each system-level simulation. there is no transmission delay between ue and enb communication. due to computational requirements, time values (expressed in tti (transmission time interval)) are normalized. all algorithms take into account the signalling sequence order proposed in [3]. 6 simulation results 6.1 pci detection and resolution to evaluate the pci conflict detection and resolution algorithm performance, a new enodeb commissioning was simulated. to new enodeb, already in use pcis will be assigned. the percentage of blocked users due to interference caused by conflicting cells is used as metric to evaluate network stability during integration. figure 5 presents the initial scenario. in this scenario, “enb 5” represents the new bts. as can be seen, the pci assigned to “enb 5” will interfere with the ones assigned to already existing “enb 4”. the main goal here is to the algorithm discover the conflicting pci and assign new ones to avoid interference. figure 5: initial scenario for pci conflict detection and resolution figure 6 presents the final scenario after simulation using the developed algorithm. it can be seen that enodeb 4 and 5 have now different assigned pci and that there isn’t any near cell with the same pci. m carvalho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc figure 6: final scenario for pci conflict detection and resolution figure 7 presents the number and percentage of blocked users during a 500 tti simulation. as can be seen, blocked users are progressively eliminated and so, it can be concluded that network stability is achieved at the end. figure 7: blocked users do to pci interference between cells close to each other the results presented in [5] are not clear as to the effectiveness achieved by the algorithm since it is considered that the network stability is achieved when all expected neighbor relationships are created which may not be entirely true. so it was not possible to establish a term of comparison between our results and those presented in [5]. 6.2 anr as stated in [1], to reduce run-time computational complexity, the measured link quality is abstracted using sinr (signal to interference and noise ratio) as metric. each ue has a sensibility of about -133 dbm (link-budget). when accessing a new cell, during handover process, if measured rsrp is lower than ue sensibility the handover will fail. in the same way, if ue received sinr drops below a certain value, when accessing a new cell, the handover will also fail. figure 8 show the 50 tti average handover count and fail percentage for a required sinr of -10 db for a 1 mbps dual-antenna receiver terminal according to the link-budget presented in [9]. figure 8: intra-frequency handover count and fail percentage for a minimal required sinr of -10 db and a dual antenna receiver. as can be seen there is an average handover failure of about 10 % mainly due to poor signal coverage areas. these areas exist because considered simulation area is larger than predicted coverage area due to polygon draw limitations. it can also be seen a large number of performed handovers. these results allow us to conclude that the implemented algorithm is properly working. 6.3 handover optimization as previously mentioned handover optimization algorithm takes into account the speed and type of service of the ue. to properly evaluate the performance of algorithm, different ue types of service and speeds are simulated. table 2 presents the initial simulation parameters taken into account for handover optimization as described in [7] m carvalho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc parameter value initial {hysteresis,timeto-trigger) {3 db ,3 tti} critical time (tcrit) 15 tti delta hysteresis increment/decrement 1 % delta ttt increment/decrement 50 % initial good performance counter 5 tti initial bad performance counter 5 tti simulation time 500 tti tti length 1 ms table 2: handover optimization parameters, [7] figure 9: handover oscillation rate for http service without handover optimization. figure 10: handover oscillation rate for voip service without handover optimization. figures 9 and 10 show the results for a nonoptimized scenario where hysteresis and ttt are fixed and for http and voip traffic, respectively. the results are expressed in hpi��� rate during 500 ttis for the case where the ue maximum speed is 3, 60 and 120 km/h. as can be seen, in both ue traffic type and speed scenarios, a high percentage of handovers are classified as ping-pong handovers and, as expected, at high speed, the hpi��� percentage is smaller, reaching 64 and 62% for http and voip service types, respectively. in the sequence, figures 11 and 12 present the equivalent results but now considering the handover optimization function. figure 11: handover oscillation rate for http service type using handover optimization. figure 12: handover oscillation rate for voip service type using handover optimization. m carvalho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc as can be seen, in both types of service, and for overall ue speeds, there is a significant hpi��� reduction especially at low speeds. when using http, there is a reduction of about 71% on hpi��� when ues are moving at 3 km/h. as the speed increases, the hpi��� reduction decreases. when using voip, there is also a significant reduction of about 70% on hpi��� when ues are moving at 3 km/h. similar to the http service, as the speed increases, the hpi��� reduction also decreases. finally, it can also be seen a larger reduction in hpi��� when using http service type. when comparing to the results presented in [7], where an empirical scenario is used, the obtained results show that a greater hpi��� reduction is indeed achieved when using a real scenario. table 3 presents the final optimized averaged hysteresis and ttt for http and voip services and for each simulated ue speed. 3 km/h 60 km/h 120 km/h http {2,4;10} {2,01;11} {2,8;12} voip {2;8} {2,0;10,8} {2,5;10} table 3: handover optimization parameters, [6] as presented in [7], in overall, hysteresis is smaller when service type is voip when compared with http. however, at low speed scenarios, hysteresis is higher when compared to the high speed scenarios. 7 conclusion in this paper, we propose a lte son capable simulator through which self-configuration and selfoptimization functions can be visualized and evaluated over user defined scenarios. in this simulator, pci conflict detection and resolution, anr and handover optimization functions are implemented and evaluated using a real network scenario. the obtained results show an overall benefit of the developed simulator in predicting and gathering the best configuration parameters for son based networks. when compared to [2], the developed simulator reveals himself more practical and useful thanks to greater detail that can be achieved in each scenario and to added handover optimization function that can estimate optimal hysteresis and ttt values. due to the simulation conditions considered in [5] a proper comparison in the pci detection and resolution functions was not possible. still, the developed algorithm is able to eliminate conflicts of pci using only ues measurement reports. with regard to the handover optimization function the achieved results reveal themselves better when using a real scenario compared to the ones obtained with the empirical scenario used in [6]. 8 acknowledgements the authors would like to acknowledge instituto de telecomunicações (it) and instituto superior de engenharia de lisboa (isel), portugal, for the support. references [1] ikuno, j., wrulich, m., rupp, m., "system level simulation of lte networks", institute of communications and radio-frequency engineering, vienna university of technology, http://www.nt.tuwien.ac.at/ltesimulator. [2] quan, h., åström, t., jern, m., moe, j., gunnarsson, f., kallin, h., “visualization of self-organizing networks operated by the anr algorithm”, proc. ieee research, innovation and vision for the future, rivf, pp 1-8, 2009. [3] 3rd generation partnership project; technical specification group radio access network; evolved universal terrestrial radio access (e-utra); overall description, ts 36.300, 2011 [4] wu, y., jiang, h., wu, y., zhang, d., “physical cell indentity self-organization for home enodeb deployment in lte”, ieee, 2010 [5] amirijoo, m., frenger, p., gunnarsson, f., kallin, h., moe, j., zetterberg, k., “neighbor cell relation list and physical cell identity self-organization in lte” icc’08, ieee, 2008. [6] carvalho, m., vieira, p., “an enhanced handover oscillation control algorithm in lte self-optimizing networks”, wpmc’11, ieee, 2011 [7] 3rd generation partnership project; technical specification group radio access network; spatial channel model for multiple input multiple output (mimo) simulations (release 10), 3gpp tr 25.996, 2011 [8] 3rd generation partnership project; technical specification group radio access network; evolved universal terrestrial radio access (e-utra); radio frequency (rf) system scenarios (release 10), 3gpp ts 36.942, 2010 [9] holma, h., toskala, a., “wcdma for umts – hsdpa evolution and lte, fourth edition”; john wiley & sons, ltd., 2007 m carvalho et al. | i-etc cetc2011 issue, vol. 2, n. 1 (2013) id-8 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt/index.php/iajetc simulation of plasmonic effects of metal (au,ag and al) nps and rgo embedded in aqueous solutions simulation of plasmonic effects of metal (au,ag and al) nps and rgo embedded in aqueous solutions vladan stojkovica, paula louroa,b a electronics telecommunication and computer dept. isel/ipl, r. conselheiro emídio navarro, 1, 1949014 lisboa, portugal, b cts-uninova, quinta da torre, monte da caparica, 2829-516, caparica, portugal. 42437@alunos.isel.ipl.pt plouro@deetc.isel.ipl.pt abstract – graphene [1] is a material that has been extensively explored in recent years as a material with optical properties that enable its application as active material in sensing devices. in this work we will study plasmonic effects and optical properties of graphene and metal nanoparticles (aunps), comparing its results, whenever possible, with results obtained in previous studies. analysis will be supported by simulation results obtained with matlab (“mie analysis”). keywords: graphene, graphene oxide; reduced graphene oxide, silicon, au, ag, al, mie analysis, lspr, plasmonics. i. introduction carbon [2] nanomaterials such as graphene, graphene oxide (go) and reduced graphene oxide (rgo) have been extensively studied in recent times as promising materials for several important applications, mainly due to their excellent physical and chemical properties. on the other hand, metal plasmonic nanostructures, such as gold (au) nanoparticles (np), have been widely used due to their excellent optical properties. these two materials, rgo and au, are attractive due to the possibility of easy molecules bonding of to their surfaces, and to the localized surface plasmon resonance phenomenon (lspr). lspr is related to electromagnetic modes caused by electron oscillations of noble metal nanoparticles embedded in a dielectric medium. among the optical properties, the excitation by electromagnetic radiation of metallic nanostructures smaller than the wavelength of incidence light causes the radiation electric field to engage with collective electronic oscillations. when the incidence frequency equals the natural resonant frequency of the electronic density oscillation in the material, the oscillations are maximized, inducing the formation of intense local electric fields in the vicinity of the nanoparticle. these fields themselves act on the electrons by reinforcing the oscillations, leading to the lspr phenomenon. through nanoparticles, therefore, it is demonstrated that it is possible to manipulate and benefit from optics at nanometer scales, making common optical fields capable of producing strong evanescent waves confined to nanoparticle dimensions. one of the important scientific areas in which these phenomena and nanomaterials can be applied is sensors and biosensors. detection of chemical compounds and biological structures is of great importance in various sectors of human activity, especially in the health field, such as medical diagnosis. thus, the purpose of this work is to deepen the study of graphene, in the form of reduced graphene oxide nanocomposite (rgo) and gold nanoparticles (np), using transmission/reflection analysis technique of light. as the final objective goes beyond this work, an evaluation of the potential of this nanocomposite for future use as a functional element in biosensors will be developed. the plasmonic properties of rgo nanocomposite with metal np (au, ag and al) will be simulated as a function of the size and simulations will be used to evaluate whether their combined characteristics can be efficiently employed in the field of biosensors. ii. what is graphene? graphene was discovered in 2004 by researchers at the university of manchester, a work that awarded them the nobel prize in physics in 2010 [3]. the term graphene was adopted in 1962, from the junction of graphite with the suffix -ene, due to the existing double bond. it consists of a flat monolayer of carbon atoms, organized into hexagonal cells with sp2 hybridized atoms, resulting in a free electron per carbon atom in the porbital and making graphene a usable material for various applications [4]. this unique structure gives graphene several superior properties such as high electrical and thermal conductivity, good transparency, good mechanical strength, inherent flexibility and very high specific surface area. the electrical conductivity (up to 2×104 s/cm) and high electronic mobility (2×105 cm2/v s, which is more than 100 times higher than silicon) in the graphene monolayer result from a small effective mass. electrons in a solid are restricted to certain energy bands. in an insulator or semiconductor, an electron attached to an atom can be released only if it gets enough energy to jump the gap between the valence and the conduction band. but in graphene the difference between these two bands is infinitesimal, which explains why graphene electrons can move very easily and quickly. thus, electrons in a single graphene layer behave as massless particles moving at a speed of approximately 106 m/s. it is the thinnest material ever known and the strongest ever measured in the universe, has an extremely high young modulus (1 tpa) and the highest intrinsic resistance (approximately 130 gpa) ever measured. i-etc: isel academic journal of electronics, telecommunications and computers vol. 8 , n. 1 (2022) id-2 http://journals.isel.pt mailto:42437@alunos.isel.ipl.pt the thermal conductivity of graphene at room temperature can reach 5000 w/mk (comparatively speaking it can be mentioned that copper is 400 w/mk), which enhances its applicability in thermal control. it has a very high surface area (2600 m2/g), much larger than the surface areas of graphite (10 m2/g) and carbon nanotubes (1300 m2/g). in table i main electrical properties of graphene are summarized. table i properties of graphene [4]. transmittance ~97.7 % absorbance 2.3 % density 0.77 mg m-2 charge carrier density 1012 cm-2 resistivity 10-6  cm surface area 2630 m2g-1 stiffness young modulus 1110 gpa strength 125 gpa thermal conductivity 5000 wm-1k-1 eletrical mobility 200000 cm2v-1s-1 in relation to optical properties, graphene has an almost total transparency. it can absorb a fraction of 2.3% of light (table 1). its optical properties are strongly related to its electronic properties as well as its low energy electronic structure. these properties (table 1), provide graphene as a material that can be used in applications ranging from polymeric materials to sensors, transistors, portable electronic devices and electrochemical energy storage systems. among the most promising materials in the field of biosensors stand out two variants of graphene: graphene oxide (go) and reduced graphene oxide (rgo) [5][6]. both materials, go and rgo, are able to be mixed with nano materials, such as noble materials like gold (au) and silver (ag) and enhance their resonant plasmonic capacities. iii. lspr localized surface plasmon resonance (lspr) is an optical phenomenon that occurs in np's smaller than the electromagnetic wavelength of the incident light [7][8][9][10], as shown in figure 1. an electromagnetic wave (light) can excite the electrons located in the valence band of the noble metals, so that there is a transition of electrons from valence band energy level to conduction band level. at the initial moment there is the displacement of electrons in the opposite direction to the electric field of the incident wave (figure 1). this displacement of charges promotes the induction of an electric dipole in the particle. the induced dipole promotes the appearance of a restorative electric field in np, which has the function of restoring the equilibrium given by the distortion of the charges. this restorative force and dipole induction, when connected, generate the plasmon resonance. the cloud around the np has its own oscillation and will have maximum oscillation when the incoming wave is at the resonant cloud oscillation frequency. thus, plasmon in an np can be considered as a harmonic oscillator driven by the resonant light wave, where the electron cloud oscillates as a simple dipole in the direction parallel to the electric field of electromagnetic radiation. only the wave often resonant with cloud oscillation is capable of producing the lspr phenomenon. localized surface plasmon resonance is influenced by the size of np, its geometry (involves changing the energy conditions on its surface), the change of local dielectric conditions surrounding the np, as well as the type of ligand (nanoparticle-interacting molecules that can change the dielectric constant) and solvent type (which also changes the dielectric constant). it is due to the plasmon resonance phenomena in noble metals that we are able to use the uv-vis spectroscopy technique and the induced changes in the plasmon oscillation frequency caused by the change in the surrounding dielectric constant. the presence of ligands and solvents also cause dielectric constant changes and alter plasmon resonance conditions. fig. 1. incidence of light over metallic np and oscillator model [11]. iv. mie theory in 1908 gustavo mie presented a satisfactory resolution for maxwell's equations, through extensive mathematical work done by hand, in which he considered an electromagnetic wave interacting with a conducting sphere [11]. in this work, mie conditioned the system under the following boundary conditions: a polarized electromagnetic wave on a given plane falling on a homogeneous sphere surrounded by a real dielectric medium, with equivalent real refractive index (re). considering the conducting sphere, the phenomenon of absorption of the electromagnetic wave should be evaluated, since the dielectric function of a conductor is complex (it has an imaginary component, im) and, in turn, has a complex refractive index, which is a function of the frequency of the incident electromagnetic wave. considering only nanoparticles that have 2r << λ or approximately 2r < λmax/10, and using an approximation known as quasi-static approximation (since the electric field of the incident electromagnetic wave is considered to be static over the nanoparticle over a given period), mie v. stojkovic et al. | i-etc , vol. 8, n. 1 (2022) id-02 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt obtained the expression for the light extinction crosssection, described in equation 1 [12][13]: (1) where v is the volume of the particle, ω is the angular frequency of extinguished light, c is the speed of light, and εm and εnp (ω) = ε1(ω) + iε2(ω) are the dielectric functions of the medium and nanoparticle (the material) , respectively. the resonance condition is reached when: ε1(ω)= -2εm (2) from equations 1 and 2, we can conclude that the plasmonic properties of any material are defined by their dielectric function ℇnp (ω). we can also predict that the growth of εm leads to an increase in the negative value of ε1, which is necessary to satisfy resonance conditions. this shifts the plasmonic peak to the infrared side, which corresponds to increase the wavelength. from the physical point of view, this increase in εm corresponds to an increase of the restorative force in the polarized electron cloud, which in turn lowers the lspr frequency. we can conclude that lspr is very sensitive to changes in the surrounding environment, εm, of np (or change in the refractive index of the medium). we can also infer that σext depends on the particle volume. σext is maximum when the denominator is minimized, i.e., when plasmon is excited at the frequency that satisfies the condition ε1(ω) = -2εm, resulting in a sharp increase in absorption and/or scattering at that wavelength. v. simulations in this section we present simulation results developed in matlab for the evaluation of the extinction efficiency, predicted by the mie theory, as a function of the light wavelength. we used three different materials for the embedded nanoparticles (au, ag and al) in different media (water, rgo matrix). it was analyzed the influence of the np particle size and of the particle’s embedding medium. 1. influence of gold np size in water in figure 2 it is displayed the simulated extinction coefficient variation with light wavelength for gold np of sizes ranging from 30 nm up to 70 nm. it was assumed that np were embedded in water. the wavelength range varied from 300 nm up to 900 nm. a) b) c) v. stojkovic et al. | i-etc , vol. 8, n. 1 (2022) id-02 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt d) e) fig. 2. simulation for au np in water with size: a) 30 nm; b) 40 nm; c) 50 nm; d) 60 nm and e) 70 nm. results show that the increase on the particle size results in a shift of the plasmonic peak to longer wavelength range. with au np of 30nm the peak appears at 536 nm and with 70 nm at 622 nm. from 30 nm up to 50 nm the magnitude of the lspr peak increases with the size, and in the range 50 nm up to 70 nm it decreases again. thus themaximum forau np is observed with 50 nm at 570 nm. in figure 3 it is displayed the qext variation with light wavelength for gold np in water with sizes ranging from 30 nm up to 70 nm. fig. 3. qext variation with light wavelength for au np in water with sizes ranging from 30 nm up to 70 nm. data show that the plasmonic signal occurs in the range from 540 nm to 650 nm, depending on the np size. the increase of the np size shifts the ressonance signal to longer wavelengths and broadens the signal. the maximum resonance peak occurs at 570 nm for au np of 50 nm radius. 2. influence of silver np size in water in figure 4 it is displayed the simulated extinction coefficient variation with light wavelength for silver np of sizes of 50 nm. this size corresponds to the size of au np with maximum plasmonic peak. it was assumed that np were embedded in water. the wavelength range varied from 300 nm up to 900 nm. fig. 4. simulation for ag np in water with size 50 nm. results show that when we compare np of the same size, the silver np present the plasmonic peak at 496 nm, which corresponds to a shift to shorter wavelengths. gold np of 50 nm exhibt the plasmonic peak at 570 nm. in figure 5 it is displayed the qext variation with light wavelength for silver np in water with sizes ranging from 30 nm up to 70 nm. fig. 5. qext variation with light wavelength for ag np in water with sizes ranging from 30 nm up to 70 nm. results show that for silver np the plasmonic signal is maximum for small size particles (30 nm). when the np size increases it is observed that the plasmonic peak v. stojkovic et al. | i-etc , vol. 8, n. 1 (2022) id-02 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt exhibits lower magnitude, shifts to the longer values and broadens in a wider range of values. 3. influence of aluminum np size in water in figure 6 it is displayed the simulated extinction coefficient variation with light wavelength for aluminum np, embedded in water, of sizes ranging from 30 nm up to 70 nm. fig. 6. qext variation with light wavelength for al np in water with sizes ranging from 30 nm up to 70 nm. in al np the plasmonic signal is well defined for 40 nm size. for the other np sizes the signal is broad and of low magnitude, corresponding to a low excitation effect. as the range of interest is limited to 300 nm up to 900 nm, we will not consider particles of smaller sizes, that will provide plasmonic response at wavelengths below 300 nm. 4. influence of the rgo as embedding medium in figure 7 it is displayed the simulation of the extinction coefficient for different metallic np composed of gold, silver and aluminum. the np sizes was varied in the range 30 nm – 70 nm. a) b) c) fig. 7. simulation of qext for different metallic np (sizes in the range: 30 – 70 nm) embedded in a rgo medium composed of: a) au, b) ag and c) al. accordingly, to the simulation results it is possible to infer that the highest magnitude of the resonance peaks occurs either with gold or silver np of 30 nm. the increase on the np size decreases the peak magnitude. this behavior is more evident in silver than in gold. the aluminum np produces a reduced response at the resonance wavelength and a shift of the peak to lower wavelengths range. 5. influence of np in water with a rgo mantle in figure 8 it is displayed a pictorial representation of the simulation conditions of the next results, where metallic np were involved in a rgo mantle and immersed in water. thus the embedding medium has to take into acount the interaction of both rgo and water. fig. 8. metallic np involved in a rgo mantle. in figure 9 it is displayed the qext variation with light wavelength for gold np of variable sizes immersed in water and with a rgo mantle (10 nm thick). v. stojkovic et al. | i-etc , vol. 8, n. 1 (2022) id-02 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt a) b) c) d) e) fig. 9. qext variation with light wavelength for gold np in water with a rgo mantle 10 nm thick. the size of the np is: a) 30 nm, b) 40 nm, c) 50 nm, d) 60 nm and e) 70 nm. the introduction of rgo around the np changes the plasmonic response. the main differences are related to the magnitude of the peak, that is reduced as well as the peak wavelength that exhibits a slight shift to longer wavelengths. in figure 10 it is displayed the qext variation with light wavelength for gold np with radius of 30 nm wrapped with a rgo mantle of variable size (10-30 nm) and immersed in water. fig. 10. qext variation with light wavelength for gold np of fixed size (30 nm) wrapped with a rgo mantle of variable size immersed in water. v. stojkovic et al. | i-etc , vol. 8, n. 1 (2022) id-02 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt the use of a thicker rgo mantle around the au np induces a strong reduction of the plasmonic peak and a shift of its spectral position to longer wavelengths. the variation of 20 nm in the mantle thickness resulted in a decrease of the plasmonic peak of 45% and a shift of 40 nm (from 600 nm to 640 nm). in figure 11 it is displayed the qext variation with light wavelength for silver np with variable radius (30-50 nm) wrapped with a rgo mantle 10 thick and immersed in water. a) b) c) fig. 11. qext variation with light wavelength for silver np of different sizes in water with a rgo mantle 10 nm thick. the size of the np is: a) 30 nm, b) 40 nm and c) 50 nm. the use of silver np with the rgo mantle produces similar results to the ones obtained with gold np. there is a reduction of the plasmonic peak magnitude when compared with silver np of the same size immersed in water. besides, the shift of the peak to longer wavelengths is more evident. in figure 12 it is displayed the qext variation with light wavelength for silver np, with 30 nm of radius, in water with a rgo mantle with variable thickness (0 – 20 nm). fig. 12. qext variation with light wavelength for silver np with radius of 30 nm in water with a rgo mantle of variable thickness (0-20 nm). results show the occurrence of the plamonic peak in the blue region of the spectrum (450 nm) when the rgo mantle is not present, and a shift to longer wavelengths when the rgo wraps the np. this shift increases with the size of the rgo mantle. at 20 nm the plamonic peak is located near 550 nm. in figure 13 it is diplayed the the qext variation with light wavelength for aluminum np, with variable radius (30 nm, 40 nm and 50 nm), in water with a rgo mantle 10 nm thick. a) v. stojkovic et al. | i-etc , vol. 8, n. 1 (2022) id-02 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt b) c) fig. 13. qext variation with light wavelength for aluminum np of different sizes (a) 30 nm, b) 40 nm and c) 50 nm) in water with a rgo mantle 10 nm thick. results demonstrate again that the increase of the np size shifts the plasmonic peak from the uv range (around 360 nm) to the visible range (480 nm). the magnitude of the peak exhibits a slight decrease. in figure 14 it is displayed the is displayed the qext variation with light wavelength for aluminum np, with 30 nm of radius, in water with a rgo mantle with variable thickness (0 – 20 nm). fig. 14. qext variation with light wavelength for aluminum np with radius of 30 nm in water with a rgo mantle of variable thickness (0-20 nm). the variation of the mantle dimension influences mainly the magnitude of plasmonic peak. the position in the spectral range shows a negligible variation. vi. conclusions results obtained for np composed of different metals, with different sizes and used in different media, show that maximum plasmonic values are strongly dependent on the following factors: size of np, their geometry and surrounding environment, water and graphene, as medium or mantle. it was also demonstrated that, for au, ag and al np of different sizes and different mantle thicknesses of rgo, that the wavelength λlspr at which the maximum plasmonic peak occurs, depends on several factors, namely, metallic composition of the np, size of np's and surrounding environment (rgo). in addition, it was concluded that the change of the environment and the change of k values caused changes in plasma maximum values λlspr. for gold np in water, it was demonstrated that the maximum peak value λlspr shifts to longer wavelengths when the size of np increases and exhibits the maximum value for 50 nm radius at 570 nm. for silver and aluminum np in water, it was observed that for larger particles the peaks shifted to shorter wavelengths. in nps of ag and al the correspondent maximum is assigned to the smallest np size (30 nm). when nanoparticles are embedded into rgo as surrounding medium, it was observed a shift of the maximum peak to longer wavelengths. the same effect is observed when the metallic np is wrapped with a rgo mantle of variable size and embedded in water. in this case it is also noted that there is also strong dependence of thickness of rgo layer regarding maximum peaks values . in summary, the results confirmed that the maximum plasmonic peak values λlspr depend on changes in refractive index and extinction between layers of np, size of np and variations in refractive index and extinction (dielectric function) between np and or surrounding environment (water/rgo). vii. future work next steps, related to the study of functional biosensors, will have to include the addition of antibodies (ab), interaction with metallic np as well as lspr effects on metallic np + rgo + ab nanostructures embedded in water. depending on the obtained results, decisions will be taken on the possible ways to pursue the research. there are two possible directions. the study may proceed on liquid samples involving lspr effects (current study) or may continue on the research of plasmonic effects (spr) with solid samples deposited on the graphene layer metal surface. viii. references [1] de jesus, karla acemano, estevão freire, and maria josé oc guimarães. "grafeno: aplicações e tendências tecnológicas." (2012). v. stojkovic et al. | i-etc , vol. 8, n. 1 (2022) id-02 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt [2] pei, q. x., y. w. zhang, and v. b. shenoy. "a molecular dynamics study of the mechanical properties of hydrogen functionalized graphene." carbon 48.3 (2010): 898-904. [3] novoselov. k. s., geim, a. k., morozov, s. v., jiang, d., zhang, y., dubonos, s. v., grigorieva, i. v., firsov, a. a. electric field effect in atomically thin carbon films. science, 306, 666-669, 2004. [4] segundo, josé etimógenes duarte vieira, and eudésio oliveira vilar. "grafeno: uma revisão sobre propriedades, mecanismos de produção e potenciais aplicações em sistemas energéticos." revista eletrônica de materiais e processos 11.2 (2016). [5] khalil, ibrahim, et al. "graphene–gold nanoparticles hybrid—synthesis, functionalization, and application in a electrochemical and surface-enhanced raman scattering biosensor." materials 9.6 (2016): 406. [6] zarbin, aldo jg, and marcela m. oliveira. "nanoestruturas de carbono (nanotubos, grafeno): quo vadis." química nova 36.10 (2013): 1533-1539. [7] turcheniuk, kostiantyn, rabah boukherroub, and sabine szunerits. "gold–graphene nanocomposites for sensing and biomedical applications." journal of materials chemistry b 3.21 (2015): 4301-4324. [8] ferreira, jacqueline, et al. "ressonância de plasmon de superfície localizado e aplicação em biossensores e células solares." química nova. vol. 39, n. 9 (nov. 2016), p. 1098-1111 (2016). [9] cittadini, michela, et al. "graphene oxide coupled with gold nanoparticles for localized surface plasmon resonance based gas sensor." carbon 69 (2014): 452459. [10] cittadini, michela, et al. "graphene oxide coupled with gold nanoparticles for localized surface plasmon resonance based gas sensor." carbon 69 (2014): 452459. [11] amendola, vincenzo, et al. "surface plasmon resonance in gold nanoparticles: a review." journal of physics: condensed matter 29.20 (2017): 203002. [12] hergert, w, and thomas wriedt. the mie theory: basics and applications. berlin: springer, 2012. internet resource. [13] quinten, michael. optical properties of nanoparticle systems: mie and beyond. john wiley & sons, 2010. v. stojkovic et al. | i-etc , vol. 8, n. 1 (2022) id-02 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt author's info i-etc: isel academic journal of electronics, telecommunications and computers vol. 8 , n. 1 (2022) id-3 http://journals.isel.pt a new lightweight and efficient fpga based architecture for aes algorithm targeting iot security abiy tadesse abebea aschool of electrical and computer engineering, addis ababa institute of technology, addis ababa university, addis ababa, ethiopia abiy.tadesse@aait.edu.et abstract1— different platforms such as resource limited devices and high-performance processors are used in iot networks each with its own set of resource, performance, and security needs. it is critical to optimize existing standard cryptographic algorithms to meet the needs of today's networks, yet this is a difficult undertaking. in this paper, a compact and efficient architecture for the advanced encryption standard (aes) is developed and implemented using several fpga devices, with the goal of addressing both constrained and high-performance platforms in iot networks. to implement a compact and efficient fpga-based aes architecture, a hybrid optimization technique is applied. the implementation takes advantage of fpga embedded resources such as brams and dsp slices. to synthesize and implement it on the xilinx virtex-7 device, the vivado hls tool 2019.2 is used. similarly, as devices older than the xilinx 7-series platforms are not directly supported by vivado hls tool, xilinx 14.5 ise tool is used to synthesize and implement it. compared to existing research results found in the literature, reductions of 78.33 %, 63.12% and 78.14% of the number of slices utilized on virtex-5, virtex-6 and virtex-7, respectively, are obtained. also, improvements of 0.33%, 0.37% and 7.42% of throughput on virtex-5, virtex-6 and virtex-7, respectively are achieved. keywords: aes algorithm, cryptography, fpga based implementation, iot security, parallel pipelining architecture i. introduction there are a number of well-known standard cryptographic algorithms with proven security capabilities, such as aes [1]. although the benefits of standard algorithms have long been known, the contemporary network requires lightweight and efficient architectures since it includes constrained as well as high-performance platforms that differ in resource, performance and security needs. optimizing existing standard cryptographic algorithms and improving their performance based on the needs of today's networks are on going research but are challenging. after the data encryption standard (des) was broken due to its small key size, aes was developed as a standard symmetric key technique [1]. the aes algorithm is a strong symmetric key block cipher that has been used to secure a variety of applications. although it is not an authenticated encryption technique in and of itself, it serves as the foundation for several authenticated encryption algorithms [2], [3]. it is a well-organized standard symmetric key algorithm that can be implemented in both hardware and software. hardware implementation of aes provides stronger physical security and higher speed compared to the software implementation. for this reason, implementing aes in hardware is vital [4]. in this paper, a compact and efficient architecture for aes is developed and implemented using several fpga platforms, with the goal of addressing the security of both constrained and high-performance platforms in iot networks. to create compact and efficient fpga-based architecture for aes, a hybrid optimization technique is applied. for the implementation, fpga embedded resources such as brams and dsp slices are combined with a reasonable number of typical fpga logic elements. the contribution of this research work is summarized as follows: considering the resource, performance and security requirements of the contemporary iot network that incorporates high performance platforms and constrained devices, an fpga-based aes architecture is proposed and implemented on different fpga devices and optimized to achieve reduced area and improved throughput. the hardware-based small footprint architecture (in terms of small number of fpga slices and embedded hard-cores) with a good throughput performance achieved using virtex-5 device is intended for constrained devices’ security in iot application environment. conversely, the virtex 6 and virtex 7 implementations of the proposed architecture with higher throughputs are intended for high performance platforms in current iot networks. the rest of the paper is organized as follows: section ii reviews the aes algorithm. fpga based implementation approaches for aes are discussed in section iii. the proposed architecture is presented in section iv. section v presents the discussion and analysis of the achieved results in comparison to existing research results found in the literature. finally, section vi concludes the paper. ii. overview of aes algorithm aes is a 128-bit block cipher algorithm with three possible key sizes: 128-bit, 192-bit, and 256-bit, respectively, with 10, 12, and 14 round operations. as http://journals.isel.pt/index.php/iajetc mailto:abiy.tadesse@aait.edu.et a. abebe et al. | i-etc, vol. 8, n. 1 (2022) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt shown in figure 1, the aes encryption process executes four essential operations to provide data confidentiality: subbytes (byte substitution), shiftrows (rows shifting), mixcolumns (column mixing), and addroundkey (adding round keys). as illustrated in figure 1, there is no mixcolumns transformation for the last round. the 128 bit block data (plaintext) is placed in a 4x4 matrix known as state before the execution of these four basic operations can begin. state elements are elements with an 8-bit size that are found in each cell of a state. before the round operations begin, the state is xored with the initial round key. fig. 1. the structure of aes encryption rounds the first round key (128 bits) is combined with 128 bits plaintext to complete the initialization process. then, until the desired round number is reached, all four operations are done in order, on the resulting state outputs (after each round step). the ciphertext is the end product of the encryption procedure. a separate key is utilized for each round operation, which is generated using a key generator function [1]. the subbytes function is a nonlinear byte replacement that uses a substitution table (s-box) to act independently on each byte of the state. s-boxes are produced in two stages: finding the multiplicative inverse in the gf(28) for the numbers 00h-ffh (zero has no reciprocal, thus it is replaced by itself); then, performing the affine transformation to them. this entails multiplying a finite field by a matrix a, then adding a finite field (exclusive or) to a vector x of hexadecimal value 63 as: [a] * b + x, which is extended as indicated in eq. (1). thus, applying all 256 possible bytes to the matrix in eq. (1) yields a lookup table that implements the subbytes transformation. shiftrows is a transformation that cyclically shifts rows to the left. in this scenario, the first row remains unchanged, but the second row shifts one byte, the third row shifts two bytes, and the fourth row shifts three bytes to the left [1]. the mixcolumns operation is processed by multiplying each state column by a matrix of constant numbers to produce an updated column [1] as shown in eq. (2) and eq. (3) for encryption and decryption processes, respectively [1]. the mixcolumns transform is composed of multiplication and addition operations, with 16 multiplications and 12 additions. when the input is multiplied by one, it can be directly taken. addroundkey is a function that adds the round key word to each column of the state matrix. one column at a time is processed by addroundkey. to generate fresh state, the state column is xored with the key generated by the key generator. inversesubbytes, inverseshiftrows, addroundkey (which is the inverse of itself), and inversemixcolumns are the inverse operations used in aes decryption process. the construction of aes is described in detail in [1] and [5]. iii. fpga based implementation approaches for aes different studies have mostly concentrated on the aes s-box for implementing the aes algorithm on fpga [4], [6]. this is because the s-box is the only non-linear component of the algorithm that has a significant impact on its performance. the shiftrows operation is a permutation of bytes that does not require any hardware. the mixcolumns operation is a linear column mixing transformation. addroundkey is a 128bit word that is xored with another 128-bit word. however, the aes s-box is a nonlinear byte replacement with a block length of 128 bits and a key length of 128 bits, considering aes-128. the substitution box is the most expensive portion of aes in terms of hardware resources, necessitating effective hardware optimization for efficient implementation [4]. as a result, new approaches of reconstructing it (the s-box) for high speed or compact area optimization targets have been presented in [4], [6], [7], without compromising the algorithm's basic purpose. ram-based (using pre-stored s-box values) [8], composite field-based (using combinational logic circuits rather than pre-stored values) [9], and lutbased (using fpga logic elements) [8], [10], are some of the implementation options for aes on fpga. to build substitution boxes, the ram-based technique makes use of the block rams found inside fpgas. subbytes and inversesubbytes are thus stored in brams. this approach can save logic elements because current fpgas include brams [8], [10]. however, it (1) (2) (3) http://journals.isel.pt/index.php/iajetc a. abebe et al. | i-etc, vol. 8, n. 1 (2022) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt necessitates a large area. the composite field technique is the computation of subbytes to directly implement the multiplicative inverse and affine transforms [9]. small-area architecture can be achieved using this method. however, it does require some transformation logic [4], [11]. the lut-based implementation method employs generic fpga logic elements [12]. the design might not be accommodated due to a lack of resources if it is sophisticated and requires more such logic parts and the fpga does not give as much of these resources as needed. the t-box (t-table) [13] is another implementation alternative that can be used instead of s-box implementations. a t-box is a lookup table technique that includes subbytes, shiftrows, and mixcolumns in addition to subbytes. a t-box can be expressed algebraically as demonstrated in eq. (5): a. optimization techniques for small area aes on fpga area optimization refers to the use of a small amount of fpga resources for compact implementation [6] results, especially for use in constrained devices. there are many methods of area optimization for fpga-based aes implementations [4], in addition to the composite field method [9]. the looping (iterative) design is the most basic way to implement aes on fpga since it takes numerous computing cycles (as it is basically an iterative algorithm). it allows the use of the same fpga hardware resource resulting in a decreased area utilization while requiring more clock cycles [4]. an iterative design on reconfigurable hardware for aes implementation is described in [12]. for aes-128, this method used a single round and processed it ten times iteratively. as demonstrated in eq. (6), the throughput may be calculated by multiplying the maximum frequency reached by the data block size (128 bits) and dividing the result by ten (number of rounds) [12]. throughput (mbps) = (fmax(mhz) x 128)/10 therefore, looping architecture necessitates the repetition of the same operation for a large number of computation cycles. as shown in figure 2 [4], this architecture employs a feedback loop in which the data is iteratively modified by round functions. fig. 2. iterative (looping) architecture resource sharing is another way for small-area optimization. this optimization technique allows different functions and operations of the same algorithm to handle equivalent jobs on the same hardware. this method aids in the reduction of hardware requirements for various components of the algorithm that would otherwise necessitate independent hardware for each [4], [6]. in general, the applicability of the algorithm for area optimization and the coding styles associated to the structure of the fpga device might influence optimization techniques for reduced resource consumption. b. optimization techniques for high-speed aes using pipelined architectures, it is possible to boost the throughput of the aes architecture at the cost of additional area. as shown in figure 3, registers are inserted at each cycle of aes to form the pipeline for concurrent processing [4], [14]. the depth of the pipeline can be determined limiting the amount of data blocks that can be processed at the same time. if full pipeline is required, the total number of rounds of the aes is chosen as the pipeline depth to obtain higher throughput [4], [9], [15]. pipeline architecture improves the performance of the encryption process as numerous blocks of data are executed at the same time. fig. 3. pipelined aes sub-pipeline architecture is created by inserting registers within the round functions of aes, as shown in figure 4 [4], [9], [10]. registers are also inserted within each round unit in this situation. if each round unit has x stages with equal delay, an n-round sub-pipelined design can reach x times the speed of an n-round pipelined architecture, with some additional registers and control logic causing some area increase [4]. fig. 4. sub-pipeline aes all rounds are implemented independently in hardware in a loop unrolling architecture [4], [16], as illustrated in figure 5. in this case, the registers placed at each round in the aes pipelined architecture are eliminated, and multiple aes rounds are processed in the same clock cycle. each round has the same delay, which is determined by the combinational logic (5) (6) http://journals.isel.pt/index.php/iajetc a. abebe et al. | i-etc, vol. 8, n. 1 (2022) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt employed. however, the aggregate of many delays produces the total delay, which makes these systems slow. unrolled architectures can increase hardware complexity by allowing a large number of rounds to be implemented independently in hardware [4]. fig. 5. loop unrolling architecture c. area versus speed trade-offs in the design and implementation of fpga-based aes, there is always a trade-off between throughput and hardware resource utilization. designing a better architecture for a small-area and high-throughput application is demanding and challenging. in general, some applications require extremely high throughput, such as e-commerce servers, a wide range of throughput, such as cell phone design, and very tiny area and low power implementations, such as rfid cards [4]. despite the fact that many studies have provided diverse implementation strategies for fpgabased aes, efficiency could be significantly improved by using efficient architectures and optimization techniques linked to devices and algorithms [4]. when aes needs to be smaller in size while still performing at a high speed, balancing area and timing efficiency is essential [12], [13], [17]-[19]. however, the area and throughput trade-offs are dependent on the application's specific requirements [17]-[19]. iv. the proposed architecture using a hybrid approach, an efficient fpga-based aes architecture is proposed as shown in figure 6. figure 6 depicts the proposed architecture, which contains two aes-128 cores running in parallel with full outer pipelining stages of rounds, including the initial round, intermediate rounds, and final round. it also depicts inner partial sub-pipelining and round-by-round processes. except for the last round, which eliminates the mixcolumn operation, a round consists of subbytes, shiftrows, mixcolumns, and addroundkey operations. subbytes and shiftrows actions are made to function in tandem with mixcolumns and addroundkey operations, as illustrated in figure 6. these round activities are also partially pipelined. to provide high throughput, the two aes-128 cores are designed to run in parallel with their respective full outer pipelining and parallel sub-pipelining modes. a pre-stored s-box bram and a hardware key scheduling module are also included in the proposed design for data access and key expansion duties, respectively. the plaintext (pt) is provided in parallel mode using prestored brams, as shown in figure 6. the two concurrent aes-128 cores each accept 128-bit input data and process it in parallel. in both cores, the initial round is first conducted by xoring the plaintext (pt) with the initial round key in parallel. the intermediate rounds are then processed in the same way, with each round's processed state output xored with the addroundkey of its associated round in both aes-cores in parallel. after the intermediate rounds are finished, the final rounds are run in parallel on both aes-cores, yielding the ciphertext (ct). for both aes cores, the same round keys are utilized. to speed up the procedure, pipelined registers are employed. the key schedule generates the appropriate keys and stores them in brams, which are then utilized by the two aes-cores during encryption. all of these procedures continue until all of the input messages have been processed. to balance hardware resource use and improve throughput efficiency, this design uses both the fpga general fabrics and specialized hard-cores such as dsp48e1 and brams. the activities of these components and the parallel operations of the modules are synchronized using a control block. the proposed method varies from other analogous designs in that it combines several methodologies to provide a compact and efficient aes-128 architecture based on a hybrid approach, which is implemented utilizing dsp and bram hard-cores, as well as balanced resources from generic fpga fabrics. rather than employing all traditional logic parts or all fpga hard-cores entirely, it balances the usage of both hardware resources. this enhances implementation flexibility while utilizing the modern fpga resources for appropriate sections of the algorithm. the proposed architecture's combined impact is intended to generate a balanced trade-off between throughput and area, with more emphasis on throughput. fig. 6. a two core parallel pipelined aes architecture http://journals.isel.pt/index.php/iajetc a. abebe et al. | i-etc, vol. 8, n. 1 (2022) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt techniques from [4], [16], and [20] were employed and merged in the proposed method. in [4], generic fpga fabrics were used to build fully pipelined aes. they didn't employ parallel aes cores or dsp hardcores. dsp slices and brams were employed in [16] to construct a high-throughput round-based and unrolled pipelined architecture. they didn't employ parallel aes-128 cores in their method. in [19], a similar technique to [16] was adopted for implementing aes and extending the method to include gcm and optimize aes-gcm utilizing dsp and bram. they followed the architecture used in [16], which included roundbased and unrolled architectures. unlike the method proposed in [20], where four parallel aes-cores were employed for aes-gcm implementation, only two parallel aes-128 cores are used based on hybrid optimization technique in the present proposed architecture. the area needed by two parallel aes-128 cores is half that of four parallel aes-cores. the usage of dsp and bram hard-cores proposed by [16] for implementation of unrolled pipelined architecture is followed; however, it is only utilized to create parallel aes-128 cores in the current study, not for a single round based and unrolled pipelined aes architecture. the proposed work uses the same fully pipelining approach as [4] to fully pipeline the two parallel aes-128 cores; however, this time it is also based on dsp slices and brams, rather than using only generic fpga logic elements as in [4]. in addition, the partial sub-pipelining strategy uses simultaneous subbytes and shiftrows operations with mixcolumns and addroundkeys operations. the key schedule shown in figure 7 is implemented in hardware, but the required round constants shown in table 1 are stored in brams. the process of creating all round keys from the original input key is known as key-expansion in aes-128. the key-expansion method develops nr + 1, 128-bit round keys from a single 128bit key if the number of rounds is given by nr. before beginning the encryption or decryption operations, an initial round key is added to the input. the circular keys are made in a word-by-word fashion. for the 128-bit key size, ten round keys of 16 bytes are created. subword applies the s-box to the 32-bit input word, rotword rotates the word one byte to the left, and the round constant rci is an 8-bit constant associated with each round, as illustrated in figure 7. table i aes round constants round constant (rcon) 1 (01 00 00 00)16 2 (02 00 00 00)16 3 (04 00 00 00)16 4 (08 00 00 00)16 5 (10 00 00 00)16 6 (20 00 00 00)16 7 (40 00 00 00)16 8 (80 00 00 00)16 9 (1b 00 00 00)16 10 (36 00 00 00)16 fig. 7. key expansion in aes v. implementation approaches the proposed architecture was first implemented on xilinx virtex-7 (part: xc7vx690t, speed-grade -3) platform. it was tested using the xilinx vivado 2019.2 high-level synthesis (hls) tool. a functional test for the encryption section of aes-128 is shown in figure 8. the rtl output was then synthesized on xilinx virtex5 (xc5vlx155t, speed-grade -3) and virtex-6 (xc6vlx75t, speed-grade -3) fpgas. fig. 8. functional test for aes-128 encryption xilinx vivado hls is a cutting-edge eda tool that allows to specify a design in software, synthesize it, and generate rtl for the specified design. it gives the freedom to optimize the design implementation using a variety of optimization directives to get the results needed. the synthesized vhdl code was implemented on xilinx virtex-7 device. xilinx ise 14.5 was used to implement and analyse the performance of the proposed architecture on earlier generations of xilinx fpga http://journals.isel.pt/index.php/iajetc a. abebe et al. | i-etc, vol. 8, n. 1 (2022) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt devices that are not directly supported by the vivado hls tool, including virtex-5 and virtex-6. a simple finite state machine (fsm) was used for synchronization of the activities of the different components of the proposed architecture including the pipeline registers found in dsp slices. exclusive-or (xor) operations are employed to accomplish binary addition operations in hardware, hence dsp slices were used to conduct the xor operations. brams were used to hold the s-box values and constants. as a result, the brams are read whenever the data and keys for processing the aes states are required. to execute 128 bit operations, cascaded dsp slices were employed. vi. results comparisons and analysis xilinx virtex-5, virtex-6, and virtex-7 fpga platforms were used to synthesize the proposed architecture. table ii, table iii and table iv show comparisons of the synthesis results of proposed architecture on virtex-5, virtex-6 and virtex-7 devices, respectively, including resource utilization and throughput performance, against existing research findings found in the literature. it's difficult to obtain a balance between throughput performance and resource consumption as there is always a tradeoff between them. nonetheless, the goal of this research is to achieve better throughput while using relatively less hardware resources. in comparison to existing research outcomes found in the open literature, a lesser number of lut slices are utilized for synthesizing the proposed aes architecture on xilinx virtex-5, virtex-6, and virtex-7 fpga devices, while achieving improved throughput performance. thus, reductions of 78.33 %, 63.12% and 78.14% of the number of slices utilized on virtex-5, virtex-6 and virtex-7, respectively, are obtained. similarly, improvements of 0.33%, 0.37% and 7.42% of throughput on virtex-5, virtex-6 and virtex-7, respectively, are achieved compared to the outcomes found in the open literature for the related works. these results are achieved at the cost of smaller number of fpga hard-cores. from the results shown in table ii, table iii and table iv, it is noted that the achieved frequencies of the present work are higher. this is because the reported results in this paper are post-synthesis timing analysis and not post place-and-rout (par) simulation analysis that considers full design routing. moreover, only the aes encryption part is considered for the post-synthesis timing analysis. vii. conclusions it's critical to improve the performance of existing standard cryptographic algorithms to meet today's security standards. despite the fact that the benefits of such standard algorithms have long been recognized, the contemporary network's resource, performance, and security needs demand lightweight and efficient designs since it includes constrained and high-performance platforms. an efficient architecture for fpga-based aes algorithm is proposed which takes into account the iot security. a hybrid technique is employed, and the proposed aes architecture is implemented on xilinx virtex-5, virtex-6, and virtex-7 fpga devices. in comparison to existing research outcomes found in the open literature, lower number of lut slices are utilized for implementation of the proposed aes architecture on xilinx virtex-5, virtex-6, and virtex-7 fpga devices, while achieving improved throughput at the cost of smaller number of fpga hard-cores. as a result, 78% reduction of the number of slices on virtex-5 device and 7.42% improvement of throughput on virtex-7 device are achieved. in the future, post-implementation timing analysis will be performed considering full routing of the design. http://journals.isel.pt/index.php/iajetc a. abebe et al. | i-etc, vol. 8, n. 1 (2022) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt table ii results for implementation of the proposed aes architecture on virtex-5 device methods device slices brams dsps freq. (mhz) throughput (gbps) [21] virtex-5 3420 199.18 25.50 virtex-5 3788 232.30 29.73 [22] virtex-5 7385 638.16 81.68 [23] virtex-5 2940 704.70 90.40 this work virtex-5 637 8 32 708.60 90.70 table iii results for implementation of the proposed aes architecture on virtex-6 device methods device slices brams dsps freq. (mhz) throughput (gbps) [21] virtex-6 4095 463.42 5.93 [17] virtex-6 423 191.98 2.50 [21] virtex-6 5566 237.45 3.03 virtex-6 4095 463.42 59.31 [22] virtex-6 6361 849.18 108.69 [23] virtex-6 2537 740.70 94.81 [17] virtex-6 5759 732.28 93.73 virtex-6 9531 457.58 58.57 virtex-6 5759 849.18 108.69 [24] virtex-6 1626 166.66 0.24 [6] virtex-6 1551 190.66 0.56 this work virtex-6 572 8 32 828.60 109.0 table iv results for implementation of the proposed aes architecture on virtex-7 device methods device slices brams dsps freq. (mhz) throughput (gbps) [21] virtex-7 4089 495.32 6.34 [7] virtex-7 126040 288 31.29 [25] kintex-7 4493 202.26 21.92 [23] virtex-7 2617 813.0 104.06 this work virtex-7 572 8 32 878.0 112.40 references [1] j. daemen and v. rijmen, the design of rijndael: aes the advanced encryption standard, in information security and cryptography. springer, 2002. [2] d. mcgrew, j. viega, the galois/counter mode of operation (gcm), submission to nist, may 2005. [3] h. wu and b. preneel, aegis: a fast authenticated encryption algorithm (v1. 1), submission to caesar, 2016. [4] a. tadesse and p.s. kumar, effective implementations techniques for fpga based aes algorithm, 2016 kics korea and ethiopia ict international conference, 2016. [5] pub, nist fips. 197. specification for the advanced encryption standard (aes), 2001-11-26). ht-tp://csrc. nist. gov/publications/fips/fips197/fips-197. pdf 2001. [6] p. rajasekar and h. mangalam, design and implementation of power and area optimized aes architecture on fpga for iot application, circuit world, 2020. [7] s. chen, w. hu, z. li, high performance data encryption with aes implementation on fpga, in 2019 ieee 5th intl conference on big data security on cloud (bigdatasecurity), ieee intl conference on high performance and smart computing,(hpsc) and ieee http://journals.isel.pt/index.php/iajetc a. abebe et al. | i-etc, vol. 8, n. 1 (2022) id-3 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt intl conference on intelligent data and security (ids), 149-153. ieee, 2019. [8] c.j. chang, c.w. huang, h.y. tai, m.y. lin, and t.k. hu, 8-bit aes fpga implementation using block ram, in iecon 2007-33rd annual conference of the ieee industrial electronics society, 2654-2659. ieee, 2007. [9] x. zhang and k. k. parhi. high-speed vlsi architectures for the aes algorithm, ieee transactions on very large scale integration (vlsi) systems, 12(9):957-967, 2004. [10] f. wu, l. wang, j. wan, a low cost and inner-round pipelined design of ecb-aes-256 crypto engine for solid state disk, in 2010 fifth ieee international conference on networking, architecture, and storage, ieee, 485-491, 2010. [11] c. arul murugan, p. karthigaikumar and sridevi sathya priya, fpga implementation of hardware architecture with aes encryptor using sub-pipelined s-box techniques for compact applications, automatika, 61(4), 682-693, 2020. [12] f. x. standaert, g. rouvroy, j. j. quisquater, and j. d. legat, efficient implementation of rijndael encryption in reconfigurable hardware: improvements and design tradeoffs, in international workshop on cryptographic hardware and embedded systems-ches, 334-350, springer, berlin, heidelberg, 2003. [13] v. fischer, and m. drutarovský, two methods of rijndael implementation in reconfigurable hardware, in proceedings of the conference on cryptographic hardware and embedded systems (ches’01). 2160, 7792, springer, berlin, heidelberg, 2001. [14] c. p. fan and j. k. hwang, implementations of high throughput sequential and fully pipelined aes processors on fpga, in 2007 international symposium on intelligent signal processing and communication systems, 353-356, ieee, november, 2007. [15] a. hodjat and i. verbauwhede, a 21.54 gbits/s fully pipelined aes processor on fpga, in 12th annual ieee symposium on field-programmable custom computing machines, 308-309, ieee, 2004. [16] s. drimer, t. güneysu, and c. paar, c., dsps, brams, and a pinch of logic: extended recipes for aes on fpgas, acm transactions on reconfigurable technology and systems (trets), acm, 2010. [17] a.q. al-khafaji, m.f. al-gailani and h.n. abdullah, fpga design and implementation of an aes algorithm based on iterative looping architecture, in 2019 ieee 9th international conference on consumer electronics (icce-berlin), 1-5. ieee, september 2019. [18] a. soltani, s. sharifian, an ultra-high throughput and fully pipelined implementation of aes algorithm on fpga, microprocess. microsyst. 39(7), 480–493, 2015. [19] e.b. kavun, n. mentens, j. vliegen and t. yalçın, efficient utilization of dsps and brams revisited: new aes-gcm recipes on fpgas, in 2019 international conference on reconfigurable computing and fpgas (reconfig), 1-2. ieee, december 2019. [20] l. henzen and w. fichtner, fpga parallel-pipelined aes-gcm core for 100g ethernet applications, in 2010 proceedings of esscirc, ieee, 202-205, 2010. [21] h. zodpe, and a. sapkal, an efficient aes implementation using fpga with enhanced security features,journal of king saud university-engineering sciences, 32(2) 115-122, 2018. [22] s. oukili, s. bri, high throughput fpga implementation of advanced encryption standard algorithm, telkomnika. 15(1), 494–503, 2017. [23] m. b. chellam, r. natarajan, aes hardware accelerator on fpga with improved throughput and resource efficiency, arabian journal for science and engineering, 43(12), 6873-6890, 2018. [24] s. oukili, s. bri, hardware implementation of aes algorithm with logic sbox, journal of circuits, systems and computers. 26(9), 1750141, 2017. [25] s.s. rekha and p. saravanan, low-cost aes-128 implementation for edge devices in iot applications, journal of circuits, systems and computers, 28(4) 1950062, 2019. [26] shengiian, l., ximing, y., senzhan, j. and yu, p., a fast hybrid data encryption for fpga based edge computing, in 2019 14th ieee international conference on electronic measurement & instruments (icemi) 18201827. ieee, 2019. http://journals.isel.pt/index.php/iajetc vlc system for the determination of a vehicle’s position and speed vlc system for the determination of a vehicle’s position and speed f. rodrigues a , m. vieira a,b , p. louro a,b a electronics telecommunications and computer dept. isel/ipl, r. conselheiro emídio navarro, 1949-014 lisboa, portugal; b cts-uninova, quinta da torre, monte da caparica, 2829-516, caparica, portugal. a33240@alunos.isel.pt mv@isel.ipl.pt plouro@deetc.isel.ipl.pt abstract — in recent years, lighting solutions have gradually been replaced by more efficient features, taking advantage of light emitting diodes (leds) that have progressively conquered the market with increasingly high optical powers, low energy consumption and variable color temperatures. along with this evolution, visible light communication (vlc) technology has also been developed to use this existing lighting infrastructure and the inherent characteristic of leds being easily switched to high frequency to build data transmission systems. the applications of this communication technology using electromagnetic signals in the visible range are currently in a development stage with promising applications in several domains. this paper intends to study an optical communication system based on vlc to establish communication between road infrastructures and. vehicles. for this purpose, four communication channels established through the modulation of white trichromatic led emitters are used. detection of the optical signals is performed with a photodiode based on two stacked pin structures made of a-si:h and a-sic:h. this device works as an optical filter in the visible spectrum and its spectral sensitivity can be adjusted through stationary optical bias. onoff-keying (ook) modulation is used. the structure of the data blocks to be transmitted was designed to avoid undesirable effects related to ambient light (flickering and/or perceptible variations in color temperature of the white light). the experimental tests of the proposed model were performed using a small-scale prototype. the results show that with the proposed system it is possible to transmit information between road infrastructure and vehicles. keywords: visible light communication, infrastructure to vehicle communication, light emitting diode, photodiode, onoff-keying, multiplexing, dynamic current control, dataframe, semiconductor, wavelength, absorption coefficient. i. introduction in the past few years, lighting solutions are passing through a disruptive change in many ways. energy issues are a priority in a global agenda and making everything more efficient is a demand, that also includes lighting solutions. in the past almost twenty years, the traditional incandescent bulbs were replaced, first for fluorescent and fluorescent compact bulbs and, more recently, for led based bulbs. the major advantage of leds – light emitting diodes – is related to high power saving, comparatively to incandescent bulbs, or even to fluorescent bulbs. this brings new options for different kinds of illumination, for instance, at home, with light dimmable options and with different color temperatures accordingly if it the room is intended to work or to stay comfortably in living rooms, or even for decoration purposes. its use is also being adopted in public spaces, industry facilities, for signaling, etc. as led technology for lighting purposes becomes ubiquitous, the led’s switching capability brought a new possibility for data access networks in a whole new frequency range, without all the inconveniences that a brand-new infrastructure installation would cause. besides that, the use of visible spectrum offers some characteristics that are completely attractive in telecommunication systems, such as its short propagation distance and its inability to cross walls and objects. this establishes new challenges and brings new possibilities. a new emerging field for optical communication, i. e., visible light communication (vlc) that found in the possibility of led modulation an efficient way of taking advantage of the visible part of the electromagnetic spectrum to transmit information. this technology can be used in different fields extending from indoor to outdoor applications. the communication through visible light holds special importance when compared to existing forms of wireless communications. the visible light spectrum is completely untapped for communication and can complement the radio frequency (rf)-based mobile communication systems. modern vehicles are equipped with many electronic sensors, which monitor the vehicle’s speed, position, heading, and lateral and longitudinal acceleration. although the technology already exists, vehicles rarely communicate this information wirelessly to other vehicles or roadside infrastructure. researchers are anticipating the deployment of wireless vehicle communication to improve safety and reduce congestion. this particular application is known as connected vehicles. recently, the transportation lighting infrastructure such as street lamps, traffic lights, automotive lamps, etc., is changing to light emitting diodes (leds). in the case of an its based on visible light communication (vlc), it will be possible to make use of the conventional automotive and traffic leds. compared to rf-based communications, vlc offers robustness against jamming attacks, a smaller interference domain, and a large license-free spectrum. i-etc: isel academic journal of electronics, telecommunications and computers vol. 5 , n. 1 (2019) id-2 http://journals.isel.pt mailto:a33240@alunos.isel.pt mailto:mv@isel.ipl.pt mailto:plouro@deetc.isel.ipl.pt vehicular communication systems are an emerging type of network in which vehicles and roadside units are the communicating nodes, providing each other with information, such as safety warnings and traffic information. the vehicular communication is composed of infrastructureto-vehicle (i2v), vehicle-to-vehicle (v2v) and vehicle-toinfrastructure (v2i) communications. the i2v applications focus on utilizing the traffic related infrastructure, such as traffic light or streetlight to communicate useful information. so, vlc can be realized as a secondary application in led arrays that are used for lighting. in the recent past, we have developed a wavelength division multiplexing (wdm) device that enhances the transmission capacity of the optical communications in the visible range. the device was based on tandem a-sic:h/asi:h pin/pin light-controlled filter with two optical gates to select different channel wavelengths. when different visible signals are encoded in the same optical transmission path, the device multiplexes the different optical channels, performs different filtering processes (amplification, switching, and wavelength conversion) and finally decodes the encoded signals recovering the transmitted information. this device can be used as receiver and helps developing automated vehicle technologies that allow vehicles to communicate with the surrounding ‘environment’. ii. system architecture the proposed vlc system includes an outdoor scenario of infrastructure-to-vehicle communication. the led luminaires are used to perform two tasks, street illumination and transmission. the block diagram of the vlc system is depicted in (figure 1). fig. 1. block diagram of the vlc system for illumination, positioning and data transmission from infrastructure side, luminaires working as data emitters define a vlc network with clusters of four emitters. receivers are implemented in vehicles, which are mobile users of the network. a. emitter the vehicular communication system emitter for data transmission between the road infrastructure and the vehicle (i2v) is based on the use of rgb leds, together with an additional violet led, placed in public luminaries, which will take the role of vlc emitters, as well as public illumination. for illumination purposes, just rgb leds have an active role, as the three wavelengths, together, result in white light. for vlc emitter purpose, the four wavelengths are used. so, four communication channels will be available. details related to the use of the emitter channels will be further explained in the chapter regarding the network topology [1]. the characterization of the optical sources was done through the measurement of the output spectra of each biased chip junction of the rgb white led with the driving current. in figure 2 it is plotted the normalized output spectra of the rgb white leds used in this experiment. 400 450 500 550 600 650 700 0.0 0.2 0.4 0.6 0.8 1.0 r g b drive current 0.5 ma 1 ma 2 ma 3 ma in te ns ity (a .u .) wavelength (nm) fig. 2. output normalized spectrum of the rgb white led using different values of driving currents in the range 0.5ma – 3 ma. the measurement was done using a compact ccd spectrometer from thorlabs, model ccs 200/m, that allows spectral characterization of optical sources in the 200 1000 nm spectral range with a 2 nm accuracy [2]. this experiment was done using different driving currents for the emitters of the rgb white led. results demonstrated that the central wavelength and linewidth were similar. main difference was obviously related to the peak intensity, as the increase on the magnitude of the driving current results in an increase of the output optical power delivered by the led. the output spectrum covers the wavelengths assigned to the blue, green and red regions, with wavelengths centered, respectively at 470 nm, 535 nm and 630 nm. the full width half maximum (fwhm) is 22 nm for the blue chip, nearly 48 nm for the green and 13 nm for the red chip. usually the fwhm of led devices increases with the wavelength. however as this is a white led, the magnitude and width of each rgb peaks are optimized for the white. the green component is lowest because the human eye has a maximum sensitivity at 530 nm [3] [4]. in fig. 3 it is plotted the normalized output spectra of the emission spectrum of the violet leds used to soak the device with steady state background light either from the back or front sides. fig. 3. normalized emission spectrum of the violet led of the background light 350 375 400 425 450 0,0 0,2 0,4 0,6 0,8 1,0  peak = 390 nm fwhh = 15 nm 30 ma r el at iv e in te ns ity wavelength (nm) f. rodrigues et al. | i-etc, vol. 5, n. 1 (2019) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt the violet led used for the background exhibits a s single peak centered at 390 nm and a narrow full width half maximum (fwhm) of 15 nm. table i optical characteristics of the rgb and violet leds at 25 ºc red green blue violet dominant wavelength (nm) 619 624 520 540 460 480 370 – 425 luminous intensity (mcd) 355 900 560 1400 180 505 fwhm (nm) 24 38 28 15 b. receiver the vlc receiver that will be installed in vehicles, and integrates the vlc vehicular communication i2v system, should be able to detect the different wavelengths (channels) that are being used. in this case, these wavelengths are located at 620nm(r), 530nm(g), 470(b) and 400nm(v). the device will be able to convert a visible light signal to an electrical signal, in a way that it can be demodulated and extracted the information related to each input signal. for this purpose, the receiver must include a photodiode. through the wavelength that is supposed to detect, that means, through the wavelength that the photodiode should absorb, it must be designed considering the absorption coefficient of the semiconductor material of the absorber layer. the used photodiode must consider the frequency of the signal to be detected, as the time constant, resultant from the photodiode junction capacity, limits the maximum frequency of the emitted signal modulations. the penetration depth of a wavelength is the inverse of its absorption coefficient. the depth penetration of the wavelength of interest must reach the depletion region of the photodiode. pin photodiodes, composed by one p and one n semiconductor layers, separated by an intrinsic (i), active layer, offer improved characteristics regarding the device capacitance, allowing higher frequencies. besides that, they exhibit a wider depletion region, which increase the detection efficiency of the spectral range. the chosen semiconductor bandgap is an important attribute for receiver sizing, as this material should be chosen in order to increase the number of electron-hole pairs produced by the wavelength that is supposed to detect. figure 4 shows the simplified cross-section view of the photodetector. it is based on two pin heterostructures on a glass substrate with two transparent electrical contacts of indium tin oxide. the device presents an asymmetrical configuration. the front pin photodiode (pin1) is a thin structure with 200 nm and it is based on a-sic:h. the back device (pin2) is manufactured with a-si:h and is 1000 nm thick. fig. 4. pinpin photodiode used in receiver prototype [5] due to the bandgap differences of the sensitive materials of both front and back devices, the front one with a bandgap of 2.1 ev is sensitive to wavelength light up to 550 nm, which includes the blue and green parts of the visible spectrum and excludes the red part. the back device, with a bandgap of 1.8 ev is sensitive to wavelength light higher than 520 nm which corresponds to the range of green light of longer wavelengths and to the red spectrum. the thickness of both structures optimizes the detection of light of short wavelength to the front photodiode and the longer wavelengths to the back device [5] [6]. background steady state light was supplied by violet leds (390 nm, 15 nm of fwhm) that illuminate the device by the back or the front side. the white light produced by the rgb leds is directed to the front side and in each led the red and the blue chips were modulated with a specific bit sequence. the device was reverse biased at – 8v and the photocurrent was measured between the front and back electrical contacts. the output spectral characteristics of the photodetector are shown in figure 5 using background light from both front and back sides, and without any background light (which corresponds to the condition of not having any optical bias). fig. 5. spectral photocurrent under dark conditions and using front and back violet light results show that the use of steady state illumination as a background light changes the device spectral sensitivity. for long wavelengths (red at 630 nm) it is observed an amplification of the photocurrent under front optical bias while under back optical bias the signal is reduced. for shorter wavelengths the opposite trend is observed with a small amplification under back bias and a minor reduction under front bias. this means that the modulated signal of the red chip will be enhanced under front light and shortened 400 450 500 550 600 650 700 10-10 10-9 10-8 10-7 front side bias back side bias no optical bias 470 nm 630 nm p ho to cu rr en t ( a ) wavelength (nm) f. rodrigues et al. | i-etc, vol. 5, n. 1 (2019) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt under back illumination, while the blue signal will be amplified under back light and slightly reduced under front light. the analysis of the device transient photocurrent to the optical excitation from the different emitters was done with a square waveform driving current. the optical signals illuminated the photodetector from the front side [7]. in figure 6 it is displayed the measured signal due to the overlap of the four independent input channels without applied optical bias (dark) and under front and back irradiation. on the top the driving signal applied to each r, g, b and v led is presented, the bit sequence was chosen in order that when one channel is on the others are always off. fig. 6. transient photocurrent without (dark) and under front (front) and back (back) 390 nm irradiation data analysis shows that the photocurrent depends, under irradiation, on the irradiated side and on the incoming wavelength, the irradiation side acting as the optical selector for the input channels. under front irradiation, the long wavelength channels are enhanced, and the short wavelength channels quenched while the opposite occurs under back irradiation. note that, under back lighting, as the wavelength increases the signal strongly decreases while the opposite occurs under front irradiation. the quantification of the signal amplification under front and back bias is determined by the optical gain (αf and αb for the front and back gains, respectively), defined at each wavelength (λ) as the ratio between the signal magnitudes measured with and without optical bias. the gain assigned to each channel is displayed in figure 7. 400 450 500 550 600 650 700 0.1 1 r gbv front violet background 2800 wcm-2 g ai n ( v ) wavelength (nm) front bias fig. 7. spectral gain under violet front optical bias (αv) only. the arrows point towards the optical gain at the analyzed r, g, b and v input channels [8] the photocurrent produced by photodiode is then converted into a voltage signal, using a transimpedance lna., afterwards it will be demodulated, and the transmitted information recovered and decoded. since the used modulation is ook, ‘1’ or ‘0’ bits are represented by the existence of a signal, or not, respectively [9] [10]. c. modulation scheme and dataframe one of the system requirements is related to the quality of the perceived light by the users. this affects the modulation schemes as it is necessary to prevent flickering effect, or changes in the perceived light color [8]. the modulation of the emitted light was done through the modulation of the driving electrical current of the semiconductor chips of each white led. in ook, the data bits ‘1’ and ‘0’ are transmitted by turning each led on and off, respectively. once we would like to use the ook modulation, the data frame structure must be carefully designed to avoid any flickering effects. thus, the data frame structure must prevent that the emitting leds do not stay too much time turned off. this would correspond to the transmission of many ‘0’s in a row, which could make the human eye realize that the led lamp is switching. from the point of view of the quality of provided light (in order to achieve a bright white light), it is necessary to combine the three wavelengths (red, green and blue) in such intensities that, for the human eye response to each of these wavelengths, together it will be perceived as white color [11]. to create a communication protocol to ensure the required system performance and overcome the technology constraints, a 32 bits dataframe was designed. these 32 bits are divided in three control fields, one for synchronism and two for the identification of the cell (id). this sequence is followed by a fourth block that is for the payload, as it shown in the figure 8. fig. 8. dataframe structure it was considered a network comprising a single access point (mobile terminal) and several nodes that periodically generate data, at different rates. time synchronization is required for successful communication between nodes. nodes must quickly report the results to the receiver. here, the first five bits are used for time synchronization. the same sequence [10101…] is imposed simultaneously to all the emitters. each color signal (rgbv) must carry, also, its own id-bit. so, the next three bits give the id of the row and the other three the id of the column where the node is in the network. cell’s ids are encoded using a binary representation for the decimal number. for instance, an id_bit [001 010] for the r12 (red emitter at now 1 and column 2 location) light spot will be sent inside the message whereas in case of g2,3, an id_bit [010 011] will be send by the green led. with this information, the method will give an exact, unique answer, i.e., the location in the cluster and its position inside f. rodrigues et al. | i-etc, vol. 5, n. 1 (2019) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt the unit cell. the allocated time slots depend on the used topology and on the node packet generation rate. therefore, in a time slot, each node has a packet to transmit. the tested 32-bit packet includes: synchronization, node address and payload data. to the first five bits a synchronization header [10101] in an on-off pattern have been assigned, the next six [rrr;ccc] for the binary node address and the last ones for the message. in figure 9, an example of the digital optical signals codification (codewords), used to drive the leds, is illustrated. it corresponds to the simultaneous transmission of the four nodes of a unit cell number, where the nodes are labeled r12, g13, b22 and v23, corresponding to the modulation of the red, green, blue and violet emitters of the leds located respectively at position 12, 13, 22 and 23 (further developed in chapter 4). thus, r12, g13, b,2 and v23 are the transmitted node packets, in a time slot, from this cell in the network. 0,0 0,5 1,0 1,5 2,0 2,5 c 1,2 payload datadsync. [data transmission][10101] [r r r c c c] v 2,3 b 2,2 g 1,3 r 1,2 c od ew or d (r g b v ) time (ms) fig. 9. representation of one original encoded message, in a time slot with r12, g13, b22 and v23 as the designed emitter prototype hardware is the limiting element, regarding the bit rate achievable, it was determined that the transmission bit rate would be 5 kbps/channel. this means that the bit frequency is 5 khz. from the bit frequency, we can assess that the dataframe frequency is: , (1) as the data frame frequency is nearly 160 hz, and bit transitions will necessarily occur in the 3 control fields (sync. row and col. id) the minimum frequency in the emitter can easily be duplicated or even tripled. thus, flickering effect does not appear at frequencies above 200 hz, this issue hardly becomes a problem. correct handling of the light quality requires a more complex approach. the intensity of light that the human eye can see in a certain wavelength, i. e. in a certain color, can be managed in two ways. it can be done through the effective brightness of that light source or, due to the memory effect of the retina, through the amount of time that the source is emitting in a time period (duty cycle). this period just must be short enough to the frequency to overcome the flickering threshold. although in public illumination the dimming control is not a feature that should be implemented, in principle, the light intensity should be proper, as well as the resulting color. as at each emitter just one channel is used for data transmission at each time, the major problem could be the observation of some shift of tonality, especially due to intensity variation, caused by dataframes, whose binary content can have more or less bits ‘1’ or ‘0’. this effect results in a larger or smaller percentage of time that the respective led is on or off. to overcome this constraint, the option could be a combination from both solutions, creating thus a hybrid solution. that means that the brightness of the leds perceived by the human eyes can manipulate by controlling the amount of the led driving current as well as the fraction of time that those leds are on, in each period, at the same time. by studying the response of each led, from the point of view of the light produced in function of the current, it is necessary to determine how much current, in a dc regime (that is when the leds are immutable and permanently connected) that each should receive to produce a luminous intensity such that, in conjunction with the remaining wavelengths (rgb), provides the desired tonality of light. in the led corresponding to the wavelength that is being used to transmit data in each transmitter, the time that the led is off, i. e. transmitting the '0' logic level, must be compensated with an increase of current so that the following equation is verified: , (2) idc – driving current needed by each led to produce white color light, when it is not taking the role of data emitter. ibit1 – driving current needed by data emitter led, when it sends bit ‘1’, in order to achieve the right brightness to produce white color light. tframe – period of the dataframe. once again, taking advantage of the memory effect of the human eye’s retina, this adjustment of the driving current of the led may suffer some delay, if the sampling window used for this purpose represents short periods of time, thus simplifying the process. to evaluate the value that the current ibit 1 can take, it is necessary determine, in addition to the size of the sampling window, which are the extreme cases that the time ratio in which the led is connected is maximum and minimum. for this it is necessary to analyze the structure of the dataframe. figure 10 shows the block diagram that represents a dynamic control system of the biasing current of the led emitter, through the ratio of bits with the logic value '0' or '1', to stabilize the produced light intensity. f. rodrigues et al. | i-etc, vol. 5, n. 1 (2019) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt fig. 10. bias current dynamic control once the sampling window was properly defined, that means, a sampling windows whose number of bits is such that it results in a sampling period suitable for the bias current calibration of the led, it will be possible to size the counter block, which function is to count how many bits with the logic value '1' exist in each sampling window. this counter block can consist of two binary counters, one that receives the bitstream of data and increments its value to each bit '1'. the other counter receives a clock signal with the same frequency of the binary rate and counts the total amount of bits. its function is to make auto-clear to the first counter whenever the number of transmitted bits corresponding to the size of the sampling window that has been achieved. the next block is responsible for converting the calculated count to each sampling of the bitstream into a voltage signal. the signal shall describe a curve such that the output value decreases as the number of bits '1' increases and reflects the boundary conditions, i. e. the cases where the number of bits is maximum and minimum, to offer a proper offset. in this sense, the block can be constituted by a lut look up table and a dac. the dac converts the binary signal into a voltage signal as intended, but the lut offers the possibility of easily modeling this linear signal so that the response is adequate and allows to make the resolution of the controller independent of the size of the sampling window. the transconductance amplifier block has the function of transforming the received voltage signal from the previous block into a current suitable for the utilized led. for this, the amplifier must be sized to get a gain according to the maximum current that the led can receive and must be as linear as possible. finally, the led driver consists just in a switch, possibly a mosfet, that switches the led according to the value of the bits that it receives at any moment. when the logic value is '0' the driver must enter in the cut status, turning off the led, and when the logic value is '1' it must enter in the drive status, allowing the led to be biased with the current previously determined, thus producing adequate light. iii. network topology to meet the application of the vlc i2v communication technology and, a topology has been developed for the network, in order to. figure 11 shows an overview of the elements in a i2v vlc network and how they can interact between them. fig. 11. i2v network topology with full elements along the roads, street lamps are distributed in a square topology, for data transmission and lighting purposes. in this model it was used commercially available violet (v: 400 nm) and white rgb-leds. the network coverage space is therefore a two-dimensional plane, where the location of the cells is defined by the lighting lamps of the treads lanes, being assumed that their distribution is regular and equidistant, creating a foursquare pattern of clusters of four, as shown in figure 11 [12]. each cell is defined by an emitter which can be the red, blue, green or violet. it is characterized by a set of three attributes, namely, the cell line id, the cell column id and the transmission wavelength. the cell lines and column ids are transmitted in the frames sent by each emitter. a. 4.1. coverage area of each cell the coverage area of each cell consists of a circular area around it, where the radius is equivalent to the distance between the emitters. in this way, nine areas of coverage are created in a regular pattern, where each of these subzones is served by a certain number of emitters and therefore has a certain capacity [13] [14]. figure 12 represents a cluster of emitters with the distribution of the nine subzones. fig. 12. subarea delimitation in a cluster example thus, under the assumption that only one of the rgbv leds is modulated at each corner, it is presented in table 2, the nine possible allowed subzones defined inside the cluster. table ii allowed subzones defined inside the cluster f. rodrigues et al. | i-etc, vol. 5, n. 1 (2019) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt footprint regions #1 #2 #3 #4 #5 #6 #7 #8 #9 overlap r g b v r g b r b r b v b v g b v g v r g v r g if the signal comes only from one led, the position of the led is assigned to the device’s reference point. if the device receives multiple signals, i.e., if it is in an overlapping region of two or more leds, it finds the centroid of the received coordinates and stores it as the reference point. so, inside the cell, nine reference points are considered. thus, the overlap region is used as an advantage to increase the accuracy in position estimation because more overlapping region means more reference points. for wider areas the cluster pattern will be repeated as can be understood by the illustration of figure 13. fig. 13. cell cluster repeat pattern this uniformity in the distribution of cells and therefore in the resulting coverage of the subzones gives the network the ability to determine the speed at which vehicles are moving. this functionality is of high importance for the level of intelligent traffic control and management. with the definition of these attributes to the cells, the static network can be easily managed (north will be assumed as a reference). main features that characterize the adopted topology are related to the modulation wavelength, cell line id and cell column id. the wavelength of the transmission channel changes alternately from west to east between red and green or blue and violet, and from north to south between red and blue or green and violet. the cell line id takes values between 0 and 7 that starts in the west and increments for east, repeating in a cyclical manner. the cell column id takes values between 0 and 7 that starts to the north and increments to the south, repeating in a cyclical way. figure 14 intends to illustrate a road section that includes a crossover and that has vlc network coverage. here it is shown the id of each modulated emitter (line and column), as well as the wavelength (r, g, b or v) that will transmit the modulated signal. the different clusters are also identified in the area covered by the network. fig. 14. road network topology overview b. position and speed determination figure 15 illustrates the example of a two-way road (one for each direction) covered by a i2v vlc network. assuming the case of a vehicle moving from west to east (left to the right), it is known that at its displacement it will go through subzone 7, 6, 1 or 5, 4, 3, 4, 1 or 5 and 6, in this order, in a repetitive manner. the speed at which the vehicle moves can be extrapolated by assuming that between at least two emitters the speed of the vehicle is uniform and that its position in relation to the width of the road is also constant between the emitters. although it is expected that the vehicle moves as close as possible to the right side of the lane it may not happen. therefore, it is also important to determine the position of the vehicle on the road. this may be useful to infer if any user is driving in opposite direction. fig. 15. straight road coverage with subarea delimitation as so, it can be defined coefficients that represent the percentage of each subarea in each of the trajectories. consider ao, bo, co, do and eo as the coefficients for the orange trajectory (meaning asubzone 7, b-subzone 6, csubzone 1, d-subzone 4, e-subzone 3) and ab, bb, cb, db and eb as the coefficients for the blue trajectory. inside cell, the traveled path through the subzones inside the cell will follow equation: f. rodrigues et al. | i-etc, vol. 5, n. 1 (2019) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt , (3) where: ao – time of stay inside subzone 7 when travelling along the orange trajectory l(zonex) – travelled distance inside each subzone inside the cell distemitters – distance between consecutive cells. the network registers the time when each vehicle enters and exits each subarea. the time period while the vehicle is located inside each cell (∆tcell) can also be evaluated through the network. based on the example of the figure 15, related to c1 and assuming uniform speed between the emitters, this period of time can be extrapolated by the difference between t the time when the vehicle is under the coverage of b12 and the moment it undergoes on the coverage of v13. assuming a distance between emitters as 25 meters (spacing typically used between public lighting lamps), then, the speed of the vehicle (v) inside the cell is: , (4) where: v – speed of the vehicle inside the cell, (∆tcell – period of time inside the cell, analysing the case of the orange trajectory: , (5) then: , (6) assuming that the speed is uniform between cells, it is possible to extract the coefficients from each subzone and map these coefficients in a trajectory that represents the position of the vehicle in the road. this will give information on the position of the vehicle on the track. in the previous example it was analyzed a hypothetical vehicle moving from west to east along the orange trajectory. the speed can be determined at the level of the handover between cells and based on that speed and the amount of time the vehicle takes in each subarea, it is extrapolated that the position of the vehicle corresponds to the orange trajectory. in this example, conveniently, the speed was determined by the difference between the entry in the v13 domain to entry into b12 domain, assuming that the vehicle arrives first to b12 than the v13. this may not be the case, as it would be in the event of the vehicle moves in the opposite direction. this means that at the level of the processing done on the network, it is necessary to verify the coherence of the available measured data. it will be necessary evaluate the difference between the arrival times between two adjacent cells along the road (for instance, v13 and b12). this will infer about the driving direction. in other words, it is necessary for the network to check the direction of the displacement before calculating the speed. the question of the displacement direction, together with the position of the vehicle in the track, allows the network to determine whether the vehicle is doing an overtaking maneuver or to is just moving in the opposite way of the direction of the track. this capability can be a valuable feature since it allows the network to identify dangers and send this information to the users of the network. in this case the network is aware of a dangerous condition and this information can be re-transmitted to all the vehicles under the coverage of that area. in a scenario where the track is shared with autonomous vehicles and human driving vehicles, this feature can be especially interesting. iv. conclusions this work has focused on the development of a vlc network – visible light communication, for vehicular communication, where it is intended to provide a road infrastructure with the ability to communicate with the vehicles moving on it and analyze the way they do it. the final goal for this technology is that road traffic reaches a stage where communication between vehicles, infrastructure to vehicles and vehicles to the infrastructure is completely integrated and inter-operational. however, in this paper the focus has been the communication of the infrastructure to vehicles i2v. vlc receiver specifically designed for the application concerned was studied. for this purpose, a photodiode of the pinpin type was been used and the principles and characteristics to be taken was analyzed, such as dimensions, in order to receive the used wavelengths, in the range of red, green, blue and violet. background illumination of the photodiode, when applied by the front or behind, favors or inhibits the absorption of certain wavelengths. it was also notice that for our system, backlighting is only advantageously applied to the front, in order to better discriminate the different levels of the signals received. vlc emitter is the basic element in the structure of the network and the common point between the communication system and the lighting system. the structure of the data frame used in communication was defined. mode of operation of the emitter and the structure of the data frame are intrinsically connected. it has been studied the influence that the binary sequences to transmit have in the quality of the light produced by the emitters whose function of f. rodrigues et al. | i-etc, vol. 5, n. 1 (2019) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt illumination cannot be sacrificed. a bias current dynamic control system of the leds has been dimensioned, in order to maintain the correct brightness and tonality of the light produced, regardless of the data transmitted at every single moment. alternatives have also been studied to this dynamic control system based on a commitment between the complexity of the solution and the robustness of the system to harmful phenomena to the quality of light, such as the existence of long chains of bits to '1' or '0'. regarding to network topology, it was concluded that the placement of public lighting lamps is the basis for this topology determination and based on this starting point, the solution found has gone through the definition of clusters of four emitters square-shaped, corresponding to the repeating cycle of spectral resources, i.e. of the four wavelengths used for transmission, which correspond to four emission channels. have been seen that the repeating cycles of these wavelengths in the network create subzones where, in each of them, there is a specific combination of channels which they cover. these subzones follow a pattern through which it is possible to realize the direction of the movement of the vehicles, their speeds and the position they occupy on the road. it was defined the algorithm that allows the network to extract this information and thus to make intelligent traffic management. acknowledgments this work was sponsored by fct– fundação para a ciência e a tecnologia, within the research unit cts – center of technology and systems, reference uid/eea/00066/2013 and by project ipl/2018/lan4cc/isel references [1] y.-s. kuo, p. pannuto, k.-j. hsiao and p. dutta, "luxapose: indoor positioning with mobile phones and visible light," acm, 2014. [2] t. inc., "thorlabs," [online]. available: https://www.thorlabs.com/newgrouppage9.cfm?objectgroup_ id=3482&pn=ccs200. [accessed 30 9 2018]. [3] m. a. vieira, m. vieira, p. vieira and p. louro, "optical signal processing for a smart vehicle lighting system using a-sich technology," spie, birmingham, 2017. [4] m. viera, p. louro, m. a. viera, i. rodrigues, v. silva, a. fantoni and j. costa, "enlarged spectral sensitivity outside the visible spectrum in tandem a-sic:h pi'n/pin photodiodes," in sensors and applications in measuring and automation control systems (book series: advances in sensors: reviews, vol. 4 ), international frequency sensor association (ifsa), 2016, pp. 77-100. [5] p. louro, m. vieira, j. costa and m. a. viera, "on-off keying transmitter design for," proceedings of spie, san francisco, california,, 2018. [6] m. vieira, p. louro, m. fernandes, m. a. vieira, a. fantoni and j. costa, "three transducers embedded into one single sic photodetector: lsp direct image sensor, optical amplifier and demux device," in advances in photodiodes intech, 2011, p. chap. 19. [7] m. a. vieira, m. vieira, v. silva, p. louro and m. barata, "optoelectronic logic functions using optical bias controlled sic multilayer devices," mrs, 2013. [8] m. viera, m. a. vieira, p. louro and p. vieira, "finegrained indoor localization: optical sensing and detection," vbri, 2018. [9] r. w. hamming, "error detecting and erros correcting codes," bell syst. tech., 1960. [10] m. a. vieira, m. vieira, v. silva, p. louro and j. costa, "optical signal processing for data error detection and correction using a-sich technology," phys. status solidi, 2015. [11] s. muthu and j. gaines, "red, green and blue ledbased white light source: implementation challenges and control design," ieee, 2003. [12] m. vieira, m. a. vieira, p. vieira and p. louro, "coupled data transmission and indoor positioning by using transmitting trichromatic white leds and a sic optical mux/demux mobile receiver," spie, 2017. [13] t. komine and m. nakagawa, "a study of shadowing on indoor visible-light wireless communication utilizing plural white led lightings," kluwer academic publishers, 2004. [14] k. cui, g. chen, z. xu and r. roberts, "line-of-sight visible light communication system design and demonstration," csndsp, 2010. f. rodrigues et al. | i-etc, vol. 5, n. 1 (2019) id-2 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt applications for a-si:h tfts: modelling and simulation applications for a-si:h tfts: modelling and simulation p. lourenço2,3, a. fantoni1,3, m. fernandes1,3, j. costa1,3, m. vieira1,2,3 1. isel/adeetc instituto superior de engenharia de lisboa, instituto politécnico de lisboa, área departamental de engenharia eletrónica e telecomunicações e de computadores, rua conselheiro emídio navarro, 1, 1959-007 lisboa, portugal; 2. faculdade de ciências e tecnologia, fct, universidade nova de lisboa, departamento de engenharia eletrotécnica, campus da caparica, 2829-516 caparica, portugal; 3. cts-uninova, campus da caparica, 2829-516 caparica, portugal. plourenco@deetc.isel.ipl.pt afantoni@deetc.isel.ipl.pt mfernandes@deetc.isel.ipl.pt jcosta@deetc.isel.ipl.pt mv@deetc.isel.pt abstract hydrogenated amorphous silicon thin film transistors have been used as switching elements in liquid crystal displays and large area matrix addressed sensor arrays. later, these devices have also been used as analogue active elements in organic light emitting diode displays. however, this technology suffers from bias induced meta-stability. this issue introduces both threshold voltage and subthreshold slope shifts over time when gate bias is applied. such instabilities jeopardize long term performance of circuits that rely on these components. nevertheless, hydrogenated amorphous silicon thin film transistors present an exponential transfer characteristic when operating on subthreshold region and their typical power consumption is under 1 µw. this low power characteristic makes these devices ideally suited for low power electronic design. this work demonstrates, through transient analysis of a wellestablished simulation model for hydrogenated amorphous silicon, the viability of thin film transistors technology to perform both analogue and digital functions. hence, these structures may be used in both application fields. to this end, two different sets of analyses have been conducted with hydrogenated amorphous silicon based thin film transistors. the first set considers a driving circuit for an active matrix of organic light emitting diodes, biased in a way to minimize the “memory effect” (increasing shift on threshold voltage) due to long term operation. the second set of analyses were conducted upon the implementation of complementary output universal gates, namely nor/or and xnor/xor elements. keywords: thin film transistors, hydrogenated amorphous silicon, amoled driving circuit, universal gates. i. introduction the first thin film transistor (tft) has been developed in princeton, usa, at the rca laboratories (1962) and cadmium sulphide was the selected semiconductor for the active layer. this happened right after the first presentation, by mohamed atalla and dawon kahng, of the metal-oxidesemiconductor field effect transistor (mosfet) at the solidstate device conference, held at the carnegie mellon university, in 1960 [1]. nevertheless, commercially successful tfts were only obtained in the 1980s, when hydrogenated amorphous silicon (a-si:h) was included as the active layer semiconductor [2]. a tft is essentially a mosfet (see figure 1), only the semiconductor is deposited on an insulating substrate as a thin film layer for the former, as opposed to the bulk semiconductor body/substrate in conventional mosfet. moreover, bulk mosfets operate in the inversion mode and typical tfts engage in the accumulation mode [3]. for comparison purposes, the semiconductor film thickness in tfts ranges from 30-100 nm, though for single-crystal bulk silicon devices, the substrate depth usually is the silicon wafer thickness, spanning [4] from ~100 μm-1 mm. figure 1 illustrates, in a) a standard mosfet, in b) the silicon-on-insulator (soi) technology often used devices and, in c), the tft structure. the resemblance is evident, namely between c) (tft) and b) (partially depleted soi), for both operate with a floating semiconductor channel (i. e. the channel material does not extend over the whole substrate width). figure 1 soi typical mosfets, a) and b), and tft, c), structures. hence, tfts possess low leakage and good device isolation, better latch-up immunity, protection against radiation, low parasitic capacitances, reduced substrate noise and their architecture design simplicity requires no doped drain and source regions [5]. a-si:h tfts main disadvantage is their low charge carriers’ mobility. amorphous semiconductors owe their low mobility to the transport of charge being dominated by thermally activated transport of localized charge carriers. moreover, the mobility of carriers is directly related to the maximum current handling capacity and switching speed of the device. this may be concluded by comparing the i/v curves depicted on figure 2 and figure 3. here, the simulated dc transfer curves of an off-the-shelf si based mosfet and an a-si:h based tft, obtained when considering identical drain currents, are presented. i-etc: isel academic journal of electronics, telecommunications and computers vol. 6 , n. 1 (2020) id-6 http://journals.isel.pt mailto:plourenco@deetc.isel.ipl.pt mailto:afantoni@deetc.isel.ipl.pt mailto:mfernandes@deetc.isel.ipl.pt mailto:mfernandes@deetc.isel.ipl.pt mailto:jcosta@deetc.isel.ipl.pt mailto:mv@deetc.isel.pt figure 2 i/v curves of an off-the-shelf mosfet. as may be observed, the a-si:h tft graphic denotes a less abrupt transition on the subthreshold linear region of the i/v curve. this is due to the lower charge carriers’ mobility, which compromises the maximum driving current and switching capability for these devices. nevertheless, using tfts to drive application devices such as sensors and display arrays, where the required operating current and frequency are of some microamperes and a few kilohertz, this technology will be able to perform adequately. figure 3 i/v curve of an a-si h based tft for increasing vds. a-si:h tfts have been intensively used in driving liquid crystal displays (lcds) [3]. the commercial success of these display arrays has proportionated the study and development of evermore efficient tfts throughout a myriad of applications in multidisciplinary fields such as bioelectronics, optoelectronics and more. contemporaneous useful application areas for these devices are [6]: • advanced large display arrays; • sensor arrays; • rfid tags; • other disposable electronics. in this article, the simulation of a-si:h tft devices has been conducted in the automated integrated circuit modelling spice (aim-spice) simulator [7], which has been configured with asia2, an a-si:h tft level 15 spice model [8]. this model consists on a set of equations and corresponding parameters that enable the analysis of the dc operating point, ac small signal, transient and steady state of a-si:h based tft devices. namely in this model, the drain to source current is given by: 𝐼𝑑𝑠 = 𝐼𝑙𝑒𝑎𝑘𝑎𝑔𝑒 + 𝐼𝑎𝑏 (1) where, 𝐼𝑎𝑏 = 𝑔𝑐ℎ 𝑉𝑑𝑠𝑒 (1 + 𝐿𝐴𝑀𝐵𝐷𝐴 ∗ 𝑉𝑑𝑠 ) (2) 𝑔𝑐ℎ = 𝑔𝑐ℎ𝑖 1 + 𝑔𝑐ℎ𝑖 (𝑅𝑆 + 𝑅𝐷) (3) 𝑔𝑐ℎ𝑖 = 𝑞𝑛𝑠 𝑊 ∙ 𝑀𝑈𝐵𝐴𝑁𝐷/𝐿 (4) 𝑛𝑠 = 𝑛𝑠𝑎 𝑛𝑠𝑏 𝑛𝑠𝑎 + 𝑛𝑠𝑏 (5) 𝑛𝑠𝑎 = 𝐸𝑃𝑆𝐼 ∙ 𝑉𝑔𝑡𝑒 𝑞 ∙ 𝑇𝑂𝑋 ( 𝑉𝑔𝑡𝑒 𝑉𝑎𝑎𝑡 ) 𝐺𝐴𝑀𝑀𝐴 (6) 𝑛𝑠𝑏 = 𝑛𝑠𝑜 ( 𝑡𝑚 𝑇𝑂𝑋 𝑉𝑔𝑓𝑏𝑒 𝑉0 𝐸𝑃𝑆𝐼 𝐸𝑃𝑆 ) 2∙𝑉0 𝑉𝑒 (7) 𝑛𝑠𝑜 = 𝑁𝑐 𝑡𝑚 𝑉𝑒 𝑉0 𝑒𝑥𝑝 (− 𝐷𝐸𝐹0 𝑉𝑡ℎ ) (8) 𝑁𝑐 = 3.0 ∙ 10 25 𝑚−3 (9) 𝑉𝑒 = 2 ∙ 𝑉0 ∙ 𝑉𝑡ℎ𝑜 2 ∙ 𝑉0 − 𝑉𝑡ℎ𝑜 (10) 𝑡𝑚 = √ 𝐸𝑃𝑆 2𝑞 ∙ 𝐺𝑀𝐼𝑁 (11) 𝑉𝑔𝑓𝑏𝑒 = 𝑉𝑀𝐼𝑁 2 [1 + 𝑉𝑔𝑓𝑏 𝑉𝑀𝐼𝑁 + √𝐷𝐸𝐿𝑇𝐴2 + ( 𝑉𝑔𝑓𝑏 𝑉𝑀𝐼𝑁 − 1) 2 ] (12) 𝑉𝑔𝑓𝑏 = 𝑉𝑔𝑠 − 𝑉𝐹𝐵 (13) 𝑉𝑑𝑠𝑒 = 𝑉𝑑𝑠 [1 + (𝑉𝑑𝑠 𝑉𝑠𝑎𝑡𝑒⁄ ) 𝑀] 1 𝑀⁄ (14) 𝑉𝑠𝑎𝑡𝑒 = 𝛼𝑠𝑎𝑡 𝑉𝑔𝑡𝑒 (15) 𝛼𝑠𝑎𝑡 = 𝐴𝐿𝑃𝐻𝐴𝑆𝐴𝑇 + 𝐾𝐴𝑆𝐴𝑇(𝑇𝐸𝑀𝑃 − 𝑇𝑁𝑂𝑀) (16) 𝑉𝑔𝑡𝑒 = 𝑉𝑀𝐼𝑁 2 [1 + 𝑉𝑔𝑡 𝑉𝑀𝐼𝑁 + √𝐷𝐸𝐿𝑇𝐴2 + ( 𝑉𝑔𝑡 𝑉𝑀𝐼𝑁 − 1) 2 ] (17) 𝑉𝑔𝑡 = 𝑉𝑔𝑠 − 𝑉𝑇 (18) 𝑉𝑇 = 𝑉𝑇𝑂 + 𝐾𝑉𝑇(𝑇𝐸𝑀𝑃 − 𝑇𝑁𝑂𝑀) (19) 𝐼𝑙𝑒𝑎𝑘𝑎𝑔𝑒 = 𝐼ℎ𝑙 + 𝐼𝑚𝑖𝑛 (20) 𝐼𝑚𝑖𝑛 = 𝑆𝐼𝐺𝑀𝐴0 ∙ 𝑉𝑑𝑠 (21) 𝐼ℎ𝑙 = 𝐼𝑂𝐿 [𝑒𝑥𝑝 ( 𝑉𝑑𝑠 𝑉𝐷𝑆𝐿 ) − 1] 𝑒𝑥𝑝 (− 𝑉𝑔𝑠 𝑉𝐺𝑆𝐿 ) 𝑒𝑥𝑝 [ 𝐸𝐿 𝑞 ( 1 𝑉𝑡ℎ𝑜 − 1 𝑉𝑡ℎ )] (22) p. lourenço et al. | i-etc, vol. 6, n. 1 (2020) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt 𝑉𝑎𝑎𝑡 = 𝑉𝐴𝐴 𝑒𝑥𝑝 [ 𝐸𝑀𝑈 𝑞 ∙ 𝐺𝐴𝑀𝑀𝐴 ( 1 𝑉𝑡ℎ − 1 𝑉𝑡ℎ𝑜 )] (23) 𝑉𝑡ℎ = 𝑘𝐵 ∙ 𝑇𝐸𝑀𝑃 𝑞⁄ (24) 𝑉𝑡ℎ𝑜 = 𝑘𝐵 ∙ 𝑇𝑁𝑂𝑀 𝑞⁄ (25) in previous equations, all capitalized acronyms are parameters that, once acquired from the actual physical device, may be input into the model to better describe and to assure realistic analysis of the outcome from simulations. table 1 presents these acronyms along with the used values in our simulations and a brief description of each parameter. table 1 aim-spice model asia2, level 15 parameters description. alphasat saturation modulation parameter 𝟎. 𝟔 cgdo gate-drain overlap capacitance per meter channel width 0 𝐹/𝑚 cgso gate-source overlap capacitance per meter channel width 0 𝐹/𝑚 def0 dark fermi level position 0.6 𝑒𝑉 delta transition width parameter 5 el activation energy of the hole leakage current 0.35 𝑒𝑉 emu field effect mobility activation energy 0.06 𝑒𝑉 eps relative dielectric constant of substrate (𝜀𝑟 𝑜𝑓 𝑆𝑖𝑂2) 3.9 epsi relative dielectric constant of gate insulator (𝜀𝑟 𝑜𝑓 𝑆𝑖𝑂2) 3.9 gamma power law mobility parameter 0.4 gmin minimum density of deep states 1023 𝑚−3𝑒𝑉 −1 iol zero bias leakage current 3 × 10−14 𝐴 kasat temperature coefficient of alphasat 0.006 1 ℃⁄ kvt threshold voltage temperature coefficient −0.036 v ℃⁄ lambda output conductance parameter 0.0008 1 𝑉⁄ m knee shape parameter 2.5 muband conduction band mobility 0.001 𝑚2 𝑉𝑠⁄ rd drain resistance 0 𝛺 rs source resistance 0 𝛺 sigma0 minimum leakage current parameter 10−14 𝐴 tnom temperature measurement parameter (𝑇𝐸𝑀𝑃 = 300 °𝐾) 27 ℃ tox thin-oxide thickness 10−7 𝑚 v0 characteristic voltage for deep states 0.12 𝑉 vaa characteristic voltage for field effect mobility (determined by tail states) 7.5 × 103 𝑉 vdsl hole leakage current drain voltage parameter 7 𝑉 vfb flat band voltage −3 𝑉 vgsl hole leakage current gate voltage parameter 7 𝑉 vmin convergence parameter 0.3 𝑉 vt0 zero-bias threshold voltage 0 𝑉 this model has three main categories of parameters, the geometrical and technological related, the localised and defect states related, and the simulation curves optimization related [9]. in the first category are included, the drain (rd) and source (rs) resistances, the overlap capacitances between gate and drain (cgdo), and gate and source (cgso), the insulating oxide thickness (tox), the dielectric permittivities of both substrate and insulator layers (eps and epsi, respectively) and the leakage current when at zero dc bias (iol). in the simulations presented throughout this report, the aim asia2 model default values were used for the first category of parameters: • cgdo; • cgso; • rd; • rs; • tox; • eps; • epsi; • iol. the second category of parameters takes into account the distribution of localized (deep and tail) states and impurities of the a-si:h channel, which include the undisturbed by external influences (voltage or radiation) fermi level position (def0), the minimum density of deep states (gmin) and their characteristic voltage (v0). the values assumed in the simulations were (aim asia2 model default values): • def0 𝟎. 𝟔 𝑽; • gmin 𝟏𝟎𝟐𝟑𝒎−𝟑𝒆𝑽−𝟏; • v0 𝟎. 𝟏𝟐 𝑽. for the third category of parameters, these are mainly extracted from a physical device through a process included in aim-spice software and by the experimental interpretation of the curves obtained by the analysis of a given device. the simulations performed in this report have used the default values present in the aim asia2 model, except when otherwise specified. these parameters include: • 𝜶𝒔𝒂𝒕 : the saturation modulation parameter is temperature dependent and defines at which point gate saturation voltage takes place within the above threshold region. equation (16) relates the saturation modulation (alphasat) and its temperature coefficient (kasat), together with the difference between the nominal (tnom) and the external environment (temp) temperatures of the device. equations (17,18,19) incorporate the influence of the zero-bias threshold voltage (vt0) and its temperature dependence, with the corresponding coefficient (kvt), into temperature dependent threshold voltage (𝑉𝑇 ) and, ultimately, into the effective gate voltage (𝑉𝑔𝑡𝑒 ). here, vmin is an algorithm convergence parameter and delta defines the amplitude of the forward subthreshold region, as the transition width parameter. finally, equation (14) describes the effective drain source voltage (𝑉𝑑𝑠𝑒 ), where the sharpness of the knee p. lourenço et al. | i-etc, vol. 6, n. 1 (2020) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt transition between the linear and saturation regions is given by m, the knee shape parameter. • 𝑽𝒂𝒂𝒕 : the characteristic voltage for charge mobility in the field effect activated channel, which is temperature and tail states density dependent. equation (23) relates this parameter with the characteristic voltage for charge mobility (vaa), affected by the influence of the gradient between the nominal (𝑉𝑡ℎ𝑜) and external (𝑉𝑡ℎ) temperatures on the field effect mobility (characterized by the ratio between its field effect activation energy (emu) and power law gate dependence (gamma) parameters). the drain leakage current (𝑰𝒍𝒆𝒂𝒌𝒂𝒈𝒆) grows exponentially with the negative increase of gate-source voltage. according to [10], the current in the poole-frenkel region results from an accumulation of holes in the vicinity of the gate-drain overlap, which facilitates a conduction path in reverse direction. equations (20, 21, 22) define this drain current by incorporating the influence of corresponding minimum current (sigma0), temperature and activation energy hole current (el), and hole currents gate (vgsl), drain (vdsl) and zero-bias (iol) voltages parameters. the aboveand sub-threshold drain current (𝑰𝒂𝒃) considers previously referred equation (14), the effective drain source voltage (𝑉𝑑𝑠𝑒 ), a conductance parameter (𝑔𝑐ℎ ) and an output conductance parameter (lambda) that corresponds to the channel length modulation. the channel conductance (𝑔𝑐ℎ ) depends on the intrinsic channel conductance (𝑔𝑐ℎ𝑖 ) and on the drain (rd), and source (rs) resistances. on its turn, the intrinsic channel conductance (𝑔𝑐ℎ𝑖 ) is influenced by the geometrical dimensions of the channel (w and l) and the mobility in the conduction band (muband) affected by a parameter (𝑛𝑠) that, ultimately, depends on the density of localized states in the channel (𝑁𝐶 ). ii. pixel driving circuit research on active matrix organic light emitting devices (amoleds) technology for advanced displays has been an area of intense study. this technology is able to provide thin and lightweight devices with wide viewing angle, fast reaction times and operating at low power. one of the main constraints concerning amoleds is their lifetime when on normal operation. figure 4 shows the tft modelled behavior of threshold voltage over time and with temperature, of an sio2 gate insulated a-si:h tft, on an inverted staggered configuration [9]. figure 4 tft threshold voltage over time and with temperature. one must keep in mind that figure 4 depicts a steady state bias stress applied to the tft’s gate, drain and source, which is not the case when on normal amoled operation. when tfts operate as such, their gate is pulsed with a very low duty-cycle (0.01% to 1%) which, consequently, results in much longer time periods of operation for a given threshold voltage shift. moreover, the threshold voltage shifts with time predicted by figure 4, are in a much smaller scale than the ones verified in oleds (0.7 v shift for a period of 3 hours at -10 v which is equivalent to the shift verified under identical forward biasing) [11]. nevertheless, there have been reports stating that the driving scheme is a key element, when performance improving for these devices is intended. amoleds can be driven either by a direct current (dc) or alternate current (ac) schemes. the former driving mode has been more extensively used than the latter, for it was initially considered one of the major advantages and features for these devices [12]. however, research on ac driving mode has also been conducted and promising results reporting an extended lifetime, have been published by several researchers. at the same time, it has been reported that a constant current driving mode achieves longer lifetimes than a constant voltage one [11]. in this work, a basic ac driving circuit is presented, and its operation is characterized through simulation. specifically, this circuit operates as a pulse driving mode, combined with a reversed bias component to improve amoled lifetime. these two driving mode components should be able to assure better performance and accelerated recovery from p. lourenço et al. | i-etc, vol. 6, n. 1 (2020) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt degradation. the basic pixel addressing circuit is depicted in figure 5. figure 5 basic pixel driving circuit. this circuit is essentially formed by four elements. transistor t1, which is addressable through vsel on scan line, transistor t2 is a current source element, the oled element and capacitor cs. current source behaviour-like is assured by biasing transistor t2 in saturation and this constant current drives the oled element. the circuit operates as follows: current source t2 is controlled by data voltage vdat. when t1 is turned on through vsel, vdat is transferred to the gate of t2 and stored across cs which assures that the oled pixel current/brightness is kept almost unaltered until the next frame period. the oled pixel structure can be modelled by a diode in parallel with a capacitor, as represented in figure 6: figure 6 oled pixel equivalent circuit. capacitance of an oled small organic molecule and its associated polymeric diode depends on the area of the element and is approximately [12] 25 nf.cm-2. thus, considering the assumed dimensions of 110 x 330 µm for one element of the pixel structure, the aperture capacitance will be ~9 pf. figure 7 oled driving circuit. to compensate the influence of this capacitance and provide the ac operating mode capability, an extra capacitor, crev=10 pf, and a reversing polarization, vrev, are included. for previously mentioned and still not defined component cs, 1.5 pf has been the selected value, resulting on the circuit presented in figure 7, which has been designed with the help of ltspicexvii, a spice based physical simulator [13]. a. circuit operation the purpose of this work is to demonstrate through simulation that an amoled might be ac driven by a-si:h thin film transistors (tfts). to this end, the aim-spice analogic circuit simulator [7] with a precise a-si:h tft model has been utilized, after netlist importing from ltspicexvii. considering the circuit of figure 7, where m1 and t2 are asi:h tfts with widths and lengths of 33x11 µm. if the magnitude and duration of vrev pulse are large and long enough, the oled pixel is reversely biased and its current is maintained, except for the duration of vrev pulse when the current through the structure is decreased almost to zero. thus, ac mode operation is guaranteed and simulation resulting waveforms are shown in figure 8-10. figure 8 vdat and vrev waveforms. p. lourenço et al. | i-etc, vol. 6, n. 1 (2020) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt vdat is a train of 200 µs pulses with 1 ms period that emulates incoming data line. vrev and vsel are time synchronized pulses, both with 1 ms period and duration 60 µs; the former assures reverse/forward biasing to t2 and the latter mimics pixel addressing through the scan line. figure 9 current through vrev (confirming ac mode operation). figure 9 shows the current that flows through the voltage supply vrev. this bias is connected to the oled pixel “cathode” end, thus current flowing through vrev is the same current going through the oled pixel. observing the graphic of figure 9, one is able to notice that the mean current value going through the oled pixel over time is zero. this confirms oled pixel ac mode operation which, combined with a reversed-bias voltage, is expected to improve oled lifetime [11]. transient response at the diode’s anode (voled) is depicted in figure 10. here, it is noticeable the reverse bias operation and that voltage amplitude remains almost unaltered until the end of the frame period, thus providing stability to oled luminance. figure 10 – voled waveform (voltage level is maintained throughout frame period). b. four oleds ac mode driving this work refers to a basic driving circuit for amoleds, thus the next step considered simulating an 1x4 matrix of oleds. the resulting four-pixel circuit is merely a repetition of the unit circuit and all characteristics and principles of operation remained the same. figure 11 shows the 4-pixel driving circuit designed with the help of ltspicexvii. figure 11 1 x 4 amoled driving circuit. next, a similar procedure as previously performed has been followed. namely, exporting generated netlist to aimspice, transistors m1-m4 and t1-t4 have been configured with asia2, an a-si:h tft level 15 spice model, and correspondent widths and lengths, and transient analysis has been performed. it is worth mentioning that, in this circuit, vdat1 through vdat4 are each delayed by 50 µs to emulate different timings on the data line for each pixel. the resulting waveforms at the “anode” end of each oled are presented in figure 12. figure 12 pixel1 to pixel4 waveforms. once more, one can observe that voled amplitude is maintained almost constant for the remaining of the frame period after reverse biasing is applied, assuring insignificant decrease on amoled pixels luminance. c. digital elementary circuits any logic gate, thus any logical function, can be reproduced by a number of universal logic gates [5]. these universal elements are either nor or nand gates. hence, any p. lourenço et al. | i-etc, vol. 6, n. 1 (2020) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt technology able to fulfil the logical requirements of one of these universal gates, ultimately is a technology capable of reproducing all logical functions. tfts fabricated upon asi:h based technology, may also be associated to reproduce the logical functionality of these universal logic gates. in this work we will be presenting first, an a-si:h based tft association that performs as a logic nor/or gate, followed by another tft circuit which logical behavior is of an xnor/xor gate. both these circuits were designed and simulated as mentioned before in previous section. namely, circuits were designed with the help of ltspicexvii [13], the generated netlist was exported to aim-spice, all tfts have been configured with asia2, an a-si:h tft level 15 spice model [8], and considering 23 µm and 400 µm for each device’s length and width, respectively. because logic operation was the intended functionality, it is necessary that all tfts operate in saturation. this requires the i-v curve knowledge for optimal operation point determination. referring to figure 3, where it is shown the asi:h tft drain to source current, as the gate to source voltage is iterated from -10 v to 30 v with increasing values of drain to source voltage, one may consider that saturation occurs for vgs>10 v, for all simulated values of vds. then, by using ltspicexvii, a nor/or gate has been designed. this circuit is depicted in figure 13, where va and vb, and out represent the logic gate inputs and output, respectively. inputs rise and fall times were assumed to be 100 µs, and capacitors c1 and c2 have been included to smoothen or and nor switching fluctuations. these capacitances were set empirically through a series of trial and error iterating attempts. figure 13 nor/or gate circuit implementation. to verify gate’s functionality, inputs va and vb were set according to figure 14 (two square waves with different periods), with maximal amplitude of 15 v and raise/fall times of 100 µs. figure 14 va and vb inputs. once more, transient analysis was executed for a time period of 15 ms and the results obtained are presented in figure 15 and figure 16. figure 15 presents the voltage levels obtained at the gate node of m3 (see figure 13), which is the outcome result of applying the conditions depicted in figure 14 at va and vb input nodes. similarly, figure 16 shows the or functionality that was obtained at the drain node of m3 (see figure 13) under the same input conditions. next, one more logic gate has been designed, simulated and analyzed. this time the evaluated element was a xor/xnor gate and the correspondent circuit implemented in aimspice is depicted in figure 17: figure 15 nor waveform. figure 16 or waveform. p. lourenço et al. | i-etc, vol. 6, n. 1 (2020) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt figure 17 xnor/xor gate circuit implementation. the design of a xor/xnor gate requires both inverted and non-inverted inputs, hence the inclusion of inverters m6 and m7 on the above circuit schematic. also, inverter m5 provides the complementary output functionality. the same procedure, as before with the nor/or gate, has been followed. netlist from ltspicexvii has been imported into aim-spice and all tfts have been configured with asia2, a level 15 a-si:h model. at the input nodes, va and vb, were applied the same periodic square waves depicted in figure 14. again, transient analysis was conducted for a time duration of 15 ms and the results obtained are presented in figure 18 for xnor logic operation and in figure 19 for xor functionality. figure 18 xnor waveform. xor response shows better waveform stability and wider voltage differentiation between highand low-level states, than xnor response waveform. nevertheless, the agreement between the presented waveforms and xnor/xor truth tables is evident, hence one may say that a-si:h based tfts are able to perform as a building block of more complex digital circuitry. figure 19 xor waveform. tfts based on a-si:h technology may operate as the building blocks of digital and analogic circuitry, but their low mobility presents serious constraints for applications requiring fast operation. once the frequency of operation starts increasing, these devices are not able to respond adequately, giving way to a drastic decrease in performance. to illustrate this behavior, the circuit implementation depicted in figure 17 has been simulated with higher operating frequencies and the obtained results are presented next. previous waveforms depicted on figure 18 and figure 19 correspond to a maximum pulse frequency of 500 hz at the input nodes. figure 21 shows the waveforms obtained at the gate (xnor) and drain (xor) of m5 tft, for an increase of an order of magnitude in the maximum operating frequency (5 khz) performed at the vb input, as presented in figure 20. figure 20 va and vb (maxim frequency = 5 khz) inputs. it is evident the difference in performance between figure 18 and figure 19 and the correspondent waveforms obtained in figure 21. the frequency increase induced a degraded response of the circuit, namely at the xnor functionality. figure 21 xnor and xor waveforms at 5 khz maximum operating frequency. p. lourenço et al. | i-etc, vol. 6, n. 1 (2020) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt maximum operating frequency was further increased to 10 khz, resulting in the waveforms depicted in figure 22. again, there is a difference in performance when xnor and xor functionalities are compared. although xor functionality is still clearly present, xnor waveforms denote lower response performance which limits the intended functionality. this distinction between xnor and xor performances may be minimized by decreasing the ohmic value of r1, nevertheless this happens at the expense of a reduction on the performance of xor functionality (ringing increase after low-high edge transition). looking at equations (1), (2), (3) and (4), one may see that the current iab is directly related to the channel conductance gch which, on its turn, depends directly on the width (w) and inversely on the length (l) of the channel. all the other influential parameters are either intrinsic of the technology or dependent of the manufacturing process. typically, the channel length (l) is set as a constant for a given fabrication process, which leaves only the width (w) of the a-si:h channel as a designing variable, together with the gate and drain bias voltages, to try to improve the tft device performance. figure 22 va and vb inputs, and xnor/xor waveforms at the maximum operating frequency of 10 khz. as an end note for this section, we would like to add that previous circuits (nor/or and xnor/xor) are able to provide, within the frequency limits previously mentioned, complementary output functionality for they implement both logic gate and its logical complement functions. d. fanout the fanout definition refers to the number of load gates, of identical design, that might be connected to the output of a driver gate. as a-si:h tft based logic gates require essentially no quiescent current to drive similar devices, in terms of static dc characteristics, the number of connected gates to one output is virtually infinite. however, when taking into consideration the propagation delay time, associated to logic level switching across a network of connected devices, such is not true anymore. when a driver gate changes states, it does so by charging/discharging the load capacitance. this functionality may be modelled by a current source (the driver) connected to a capacitive load (the connected device), as depicted in figure 23a). as more devices are connected to the gate driver, as shown in figure 23b), the load capacitance that must be charged/discharged increases. this places a limit on the number of connected gates to the same driver device, for the rc time constant of the circuit increases, which extends in time the charging/discharging cycles. figure 23 capacitive load with a) single and with b) several connected devices. the output voltage, vout, is then given by [14]: 𝑉𝑜𝑢𝑡 = 1 𝐶 ∫ 𝐼𝑑 𝑡 −∞ 𝑑𝑡 = 𝐼𝑑 𝑡 𝐶 (26) load capacitance c is the summation of all (n) individual input gate capacitances and the driving current relates directly to the driver conductance. hence, we may consider the switching time as: 𝑡 = 𝑛(𝑊 × 𝐿)𝑙𝑜𝑎𝑑 ( 𝑊 𝐿 ) 𝑑𝑟𝑖𝑣𝑒𝑟 (27) this is so because the driving device current is controlled by the ratio, w/l (see equation (28), where µn is the mobility when at saturation, cox is the capacitance created by the insulator layer (sio2), vgs is the gate to source voltage and vt is the threshold voltage), and the loading gate capacitance (see equation (29)) is directly related to the cross section, (w*l)load, of each connected device [15]. 𝐼𝑑𝑟𝑖𝑣𝑒 = 1 2 𝜇𝑛 𝐶𝑜𝑥 𝑊 𝐿 (𝑉𝑔𝑠 − 𝑉𝑡 ) 2 (28) 𝐶𝑜𝑥 = 𝜀𝑜𝑥 𝑊𝐿 𝑇𝑂𝑋 (29) iii. conclusions amoleds operating in a constant current driving mode have longer lifetimes than when operating in a constant voltage driving mode. by applying a reverse bias component together with current pulse driving, enabling oled pixels to operate in ac driving mode, the expectations are to compensate for tft’s threshold voltage deviation when on prolonged operation and, consequently, improve the lifetime of these devices. in the pixel driving circuit section of this article, it has been demonstrated through analysis and simulations that amoled devices may be ac driven by a-si:h tfts, and consequently lengthen their lifetimes. moreover, this may be achieved without compromising any oled’s luminance for their voltage level remains almost constant after selection and for the rest of the frame period. next followed the digital elementary circuits section where it has been explored the feasibility of developing digital p. lourenço et al. | i-etc, vol. 6, n. 1 (2020) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt circuitry with a-si:h based tfts as the building blocks. any logic gate behavior may be reproduced by a number of universal gates. nand and nor gates are universal gates for they can implement any boolean function by themselves (i. e. without the need of any other gate type). hence, any technology able to implement one of these universal gates would be capable of reproducing any digital function. in this latter section, it has been demonstrated the capability of asi:h tft technology to design complex digital circuits for it has been implemented an or and a xor complementary output gates. both gates logical behavior has successfully been validated through simulation and transient analysis. however, these circuits performance is highly affected by the operating frequency. the low charge mobility presented by these devices leads to poor performance for switching frequencies above 10 khz. moreover, we have included some background information about the fanout associated to a-si:h tfts and both perspectives, static dc and dynamic, have been approached. we have also derived the equations to obtain the maximum number of connected gates to a given tft driving device. acknowledgements this work has been supported by portuguese national funds provided by fct – fundação para a ciência e a tecnologia through grant sfrh/bd/144833/2019. references [1] d. kahng, “a historical perspective on the development of mos transistors and related devices,” ieee trans. electron devices, vol. 23, no. 7, pp. 655–657, jul. 1976, doi: 10.1109/ted.1976.18468. [2] p. g. le comber, w. e. spear, and a. ghaith, “amorphous-silicon field-effect device and possible application,” electron. lett., vol. 15, no. 6, pp. 179– 181, 1979, doi: 10.1049/el:19790126. [3] a. sharma, c. madhu, and j. singh, “performance evaluation of thin film transistors: history, technology development and comparison: a review,” int. j. comput. appl., vol. 89, no. 15, pp. 36–40, 2014, doi: 10.5120/15710-4603. [4] m. r. marks, z. hassan, and k. y. cheong, “ultrathin wafer pre-assembly and assembly process technologies: a review,” crit. rev. solid state mater. sci., vol. 40, no. 5, pp. 251–290, 2015, doi: 10.1080/10408436.2014.992585. [5] t. d. takayasu sakurai, akira matsuzawa, fullydepleted soi cmos circuits and technology for ultralow-power applications takayasu sakurai, akira matsuzawa, takakuni douseki google books. [6] r. street, technology and applications of amorphous silicon, vol. 37. berlin, heidelberg: springer berlin heidelberg, 2000. [7] “aim spice.” http://www.aimspice.com/ (accessed jan. 24, 2020). [8] “aim-spice reference manual.” http://homepages rpi.edu/~sawyes/aimspice_tuto rialmanual.pdf (accessed jan. 24, 2020). [9] h. aoki, “dynamic characterization of a-si tftlcd pixels,” ieee trans. electron devices, vol. 43, no. 1, pp. 31–39, 1996, doi: 10.1109/16.477590. [10] l. zhu, “modeling of a-si : h tft i-v characteristics in the forward subthreshold operation,” current, 2005. [11] d. zou, m. yahiro, and t. tsutsui, “improvement of current-voltage characteristics in organic light emitting diodes by application of reversed-bias voltage,” japanese j. appl. physics, part 2 lett., vol. 37, no. 11 suppl. b, pp. 9–12, 1998, doi: 10.1143/jjap.37.l1406. [12] s. yujuan, z. yi, c. xinfa, and l. shiyong, “a simple and effective ac pixel driving circuit for active matrix oled,” ieee trans. electron devices, vol. 50, no. 4, pp. 1137–1140, 2003. [13] “lt spice xvii.” http://ltwiki.org/ltspicehelpxvii/ltspicehelp/ht ml/lt_spice_overview.htm (accessed jan. 24, 2020). [14] a. agarwal and j. h. lang, foundations of analog and digital electronic circuits. elsevier inc., 2005. [15] s. m. venugopal, “flexible active matrix displays and integrated amorphous silicon source drivers,” 2007. p. lourenço et al. | i-etc, vol. 6, n. 1 (2020) id-6 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt a primer on understanding google earth engine apis a primer on understanding google earth engine apis rui s. reisab, nuno datiaab, m. p. m. patobc aisel instituto superior de engenharia de lisboa, instituto politécnico de lisboa bnovalincs, fct – universidade nova de lisboa cinstituto de biofı́sica e engenharia biomédica, fc-ul ruisreis@hotmail.com {datia,mpato}@deetc.isel.pt abstract— this article introduces the rationale behind the usage of the google earth engine, and the advantages it offers, as an alternative to handle large volumes of georeferenced data using the existing tools we know as geographic information systems on premises. google earth engine is an efficient development framework that presents itself in two basic flavors: one online integrated development environment which uses the browser javascript’s engine; and two apis that can be deployed on either a python or a nodejs environment. after presenting a limited number of use cases, representative of the google earth engine design patterns, and building a prototype class using both variants, we conclude that both platforms are merely proxy apis to the google earth engine and do not have any measurable performance difference. however, since they run on fundamentally diverse contexts — a javascript’s engine on an internet browser, that integrates seamlessly with google maps, and a python environment — it is argued that their utility depends on the user requirements instead of being true alternatives. keywords: google earth engine, javascript, python, code editor, georeferenced data, multi-spectral data. i. introduction google earth engine [1] (gee) is, primarily, a distributed parallel computing platform. it is designed around a functional language pattern, even though supported on an object model, and a map reduce [2] distributed workload paradigm. leveraging the sheer computing power delivered by the google infrastructure and a multi petabyte georeferenced data repository, gee is an efficient development framework to handle all tasks related with selecting, computing calculations and displaying georeferenced data. a. georeferenced data working with georeferenced data is a grueling task considering the large volumes of information, the complexity and diversity of storage formats. using, as an example, remote sensing multi-spectral data gathered by instruments on board of satellites, it is easy to understand the complexity of obtaining, interpreting and making calculations using this kind of data: • multi-spectral data is arranged in bands that store the reflectance measurements on a range of wavelengths. for instance, the level-1c instrument’s aboard the sentinel 2 [3] constellation read 13 reflectance bands and 3 additional data quality bands. for the sentinel 2 each coordinate, a pair longitude and latitude values, represents a 10m2 area and is associated to a vector of 16 values, one for each of the reflectance and quality bands; • this data is organized using a set of rules that are, generally, specific to each satellite. seldom an interpretation layer must be used to transform the source format to one of the standard (or “de facto”) file formats, so that it can be used by one of the existing libraries (e.g. gdal [4]); • the calculations require a lot of resources for storing and processing. until recently, many researchers opted almost exclusively to use calculated products, which are datasets with multi-spectral calculated data, mostly in the form of indices, like ndvi [5]. these were published by organizations (profitable or not) like copernicus1 or vito2, and the availability of these datasets is delayed in time, considering the actual date of retrieval; • the usage of calculated products might reduce the complexity of the data, for instance a ndvi dataset has a single value for each coordinate, but the information volume is still very large. if we take a single day of data for a 3.245km2 area in portugal of ndvi gathered by the proba-v [6] instruments, where each coordinate represents a 300m2 area, we will get an approximately 1gb [7] geotiff [8] file. however, besides the storage requirements, to make additional calculations on this data, an adequate tool must be used. consider using, for example, the gdal [4] library embedded in an integrated geographical information system (gis) tool, qgis [9]. using this on premises3 setup, and the ndvi dataset described previously, the calculation of the arithmetic average of the ndvi value, on every coordinate, for two approximate areas of 300km2 and 170km2, took close to 5 minutes (using a computer with an intel core i5-6200u, 8gb ram and a 256gb ssd) [7]. the multi-spectral data is very sensitive to the presence of clouds and atmospheric aerosols. this means that multispectral data is potentially sparse due to the varying weather conditions and pollution. methods like maximum value composite [10], that require handling several of the previously described datasets, in order to obtain significant ndvi values, will require even larger amounts of storage and computing resources [7]. 1https://www.copernicus.eu/en 2https://vito.be/en 3the software is installed and runs on computers on the premises of the person or organization using the software, rather than at a remote facility. i-etc: isel academic journal of electronics, telecommunications and computers vol. 6 , n. 1 (2020) id-4 http://journals.isel.pt https://www.copernicus.eu/en https://vito.be/en so, the main challenges to handle remote sensing multispectral data are: • the quantity of storage resources needed; • the data transformation into convenient formats; and • the computing power to enable efficient calculations on these significant volumes of complex data. b. google earth engine the gee is a recent cloud platform built to handle large volumes of georeferenced data, using google’s storage and computational resources. it tackles the challenge of understanding the complex organization of remote sensing data, while unifying its representation around a set of common formats supported by the api4. this data is stored in a large database that contains some static datasets but, most importantly, live datasets, from sources that produce new data periodically, which are ingested by gee on a regular basis, namely remote sensing multi-spectral satellite data. in short, the repository contains, to date, a catalog of some 600 datasets from 50 different sources (devices and organizations) [11]. remote sensing data is gathered from 30 satellites, or satellite constellations. all this, according to google [12], represented a volume of more than 20pb5 in 2018. the api exposes an extensive set of operations that can be used to explore the public repository or other user defined datasets, that are kept in a private assets area. there is a common set of objects that structure vector based data and georeferenced bitmaps which are organized in the repository as collections. most functions work on these collections: filtering, sorting and computing calculations over their data. primitive data types (e.g. numbers, strings, etc...) and sets (e.g. lists, dictionaries, etc...) are also supported by the api. the paradigm is fully functional, since a call to gee is self contained, but it’s built around a set of objects that wrap each of the data types, and expose methods that operate on them. the main development environment is the code editor, which is a browser based tool that interacts with gee platform using the javascript’s engine and is able to use other cloud based tools from google, especially google maps and google charts. this technology mix makes it a valuable tool whenever user interaction with a rich visual interface is needed, specifically map overlays and prototyping. gee can expose applications using, what is called, earth engine applications6. however, gee also exposes an api using python or nodejs libraries7 that leverage the development of applications that do not fully depend on the google infrastructure and leapfrog some of the difficulties of using the browser based code editor, which will be addressed further. 4application programming interface 51pb ≡ 250bytes 6more information can be found at https://www.earthengine. app/ 7at the time of writing, gee is on the verge of a major update. it is not clear if the nodejs is still a priority. all of these are particularly important given the fact that google states, “earth engine is not subject to any servicelevel agreement (sla) or deprecation policy” [13], which might force the developer to use asynchronous resubmitting strategies to properly make use of gee. this article was wrote based on the experience of using the gee framework as a development platform [7]. it will pinpoint the challenges of using code editor, as well as the advantages of making use of it to produce eye catching visual information, and the quick development of prototypes, as opposed to using the python library to leverage the usage of gee in a rich development environment8. c. motivation the learning curve of gee began with exploring the code editor and the javascript api in a particular use case scenario [7]. several issues aroused while using the code editor, but two were prominent: • the complexity of cross domain scripts that include, in the same control flow, both local and distributed processes and data structures; • the need to extract data that might be used in other applications (e.g. microsoft excel) in such a way that the whole process could be streamlined and automated. it soon became clear that code editor is a valuable tool to produce maps and overlays that might be used to illustrate results obtained, using georeferenced data, but clearly inadequate to integrate gee data or functionality in more elaborate scenarios. even though the code editor is a comprehensive tool that is adequate to explore the framework and interact with the google cloud applications ecosystem, there are some obstacles using gee in some use cases, specially those that require the extraction of information to be reused over time: • the execution of long running tasks in the browser’s environment does not allow a proper progress notification system, the console output is integrated in the graphical user interface and there is no way to integrate the execution path with another control flow, namely using a callback mechanism; • the browser’s javascript engine does not distinguish between a badly written piece of code and a proper long running process, which results in the user being prompted to confirm if the execution should be aborted; • finally, and most important, the data extraction using the browser is challenging, and it’s integration in a data flow isn’t feasible, though the platform supports batch processes that can persist data using cloud storage (e.g. google drive or google cloud). there is also the issue of concurrency. sometimes it is necessary to harness the temporary unavailability of storage and processing quota due to concurrency issues between users’ processes. in code editor it is not easy to control this behavior, 8google recently has announced a third path, the colaboratory hosted jupyter notebooks which inherently uses python. r. reis et al. | i-etc, vol. 6, n. 1 (2020) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt https://www.earthengine.app/ https://www.earthengine.app/ specially while interacting with google maps. so, whenever the platform fails to respond to user queries, or the user quota is exhausted, an exception is thrown and handled by the code editor and thus the process control flow is interrupted. though somewhat over simplifying, we can state that using the python api obviates most of this issues, but loses the benefits of integrating with google maps. google made an effort to keep both apis syntactically equivalent [14]. most differences are related to the grammar of each supporting language, javascript and python, but the bottom line is that they are inherently similar. d. testing environment in the next sections we will browse some of the main gee api concepts in order to provide insights on this framework. to use gee the user must have a google account and request access to the platform9. the code editor does not require any additional setup. using the internet browser, after being authenticated using the google account, the user navigates to the site using a url10. using the python environment requires some additional steps [14] that include the installation of the python package earthengine-api. all experiments were performed using the google chrome browser (version 81.0.4044.122) to access the code editor, and a python environment (version 3.7.6, 64 bit) including the gee earthengine-api package (version 0.1.219). the host was a computer running microsoft windows 10 enterprise (version 1709, build 16299.1686), using an intel core i7-8550u cpu, 16gb ram and a 500gb ssd. e. organization the rest of the paper is organized as follows: section 2 will elaborate on the several issues faced during the previous work that led to the usage of both the code editor and the python api in different scenarios. in this section we also exemplify a few design patterns of the gee, common to all apis, and explain the interaction between the local development environment and the remote gee. in order to demonstrate all these caveats, in section 3 we introduce two simple implementations of a same class interface using both environments. section 4 presents the conclusions of this work. ii. basic concepts gee is a distributed parallel computing platform dedicated to store and process georeferenced data, which explores the most adequate programming paradigms to handle very large volumes of data. each operation is mostly self contained in a functional paradigm pattern, and parallelism is implemented via map-reduce [2] mechanism. both development environments, javascript and python, make an effort to hide the complexity of the object model and underlying processes. 9https://signup.earthengine.google.com/ 10https://code.earthengine.google.com/ this section is not supposed to be a gee reference guide, neither a tutorial. it is, instead, an overview of the api, highlighting its most relevant patterns and data organization structures. each different aspect we wish to highlight uses an object of the gee model to be used as a showcase. though both apis are almost syntactically equivalent, all the following use cases will be presented using the python environment stressing the differences whenever they occur. nonetheless, the gee user guide singles out some differences[14] (see table i). table i: some common syntax differences between javascript and python (source gee user guide). description javascript python function definition function fun(){} def fun(): variable definition var a = ”value” a = ”value” logical operators and() or() not() and() or() not() multi-line method chain fa() .fb() .fc(); fa()\ .fb()\ .fc() dictionary keys {“key”: “value”} or {key: “value”} {“key” : ”value”} boolean true false true false null values null none comment \\ # in both environments, gee objects are referenced through an ee namespace, using a dot notation. in codeeditor it is implicit and points to an existing object that supports the api, while in python it references a package explicitly using import ee. a. initializing gee any interaction with gee will have to be preceded by the establishment of a secure context which will provide user authentication and session data. a python application that uses gee will always have to include, beforehand, the call in listing 1 in order to authenticate the user and establish a connection to the gee platform. listing 1: gee session initialization. import ee ee.initialize() this is a major difference compared to using javascript in the code editor. the code editor is a google integrated cloud r. reis et al. | i-etc, vol. 6, n. 1 (2020) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt https://signup.earthengine.google.com/ https://code.earthengine.google.com/ application, the user authentication an initialization procedures are transparent. b. primitive types handling numbers this section will address the number data type, but most considerations will apply to other gee primitive type wrappers. the ee.number object encapsulates a numeric value and exposes a set of functions that act upon it. for instance, let’s consider the encapsulation of the pi constant. one could represent the value of pi using a gee object and keep it in a variable number. by the way, we could also represent the calculation of the cosine of this value, by setting a variable called cosine (see listing 2). notice the symbolic nature of both variables, stressing that, at the local python environment (or for this matter, javascript), they are only javascript object notation (json) representations that maybe used to compose complex gee operations, with no intrinsic value. it is this representation that will be submitted to the gee for execution. listing 2: cosine calculation of the pi constant in gee using ee.number import math number=ee.number(math.pi) cosine=number.cos() print(cosine) ee.number({ "type": "invocation", "arguments": { "input": 3.141592653589793 }, "functionname": "number.cos" }) in this section, whenever print() is used to output a result to the console, the code listings will be divided in two regions, split by a single line: the top region contains the python code; the bottom region reflects the output to the console. printing the content of the cosine variable does not trigger any call to the gee platform. from the generated output, it is self explanatory that the variable contains gee’s internal representation of a call to a cos function using, as single parameter, the numeric constant. every object in gee exposes a getinfo() method. whenever this is called on an instance of a given object, the representation contained at the local environment is submitted to gee, for evaluation, and the result is returned to the caller in the form of a json encoded object (see listing 3). listing 3: executing and obtaining the value of cosine in gee result=coseno.getinfo() print(result) -1.0 note that result contains a python native object that is the result of the evaluation of the representation contained in cosine on the gee platform, that is, cos(π) = −1. in short, at the local environment level, representations of gee can be composed to translate increasingly complex operations. the execution and evaluation of these operations occurs in the gee platform explicitly when getinfo(), or some other function with similar behavior, are invoked, returning the operation result as a json encoded object. c. non primitive types encoding dates a calendar date is wrapped by an ee.date object, which also exposes some functions that act upon it. one of these functions, fromymd(), acts as a constructor and encodes a timestamp given its year, month and day numeric representations. listing 4, once again, underlines the symbolic proxy nature of the gee object representation in the local python environment. listing 4: setting a date in variable today today=ee.date.fromymd(2019,5,4) print(today) ee.date({ "type": "invocation", "arguments": {"year": 2019, "month": 5, "day": 4 }, "functionname": "date.fromymd" }) using the advance() function, varying the offset parameter, it is possible to represent yesterday and tomorrow (see listing 5). listing 5: setting variables yesterday and tomorrow. yesterday=today.advance(-1,"day") tomorrow=today.advance(1,"day") print(tomorrow) ee.date({ "type": "invocation", "arguments": { "date": { "type": "invocation", "arguments": { "year": 2019, "month": 5, "day" : 4 }, "functionname": "date.fromymd" }, "delta": 1, "unit": "day" }, "functionname": "date.advance" }) evaluating the variable tomorrow (see listing 6), using getinfo(), the operation result value is returned. listing 6: evaluating tomorrow print(tomorrow.getinfo()) {’type’: ’date’, ’value’: 1557014400000} the result of an ee.date object does not translate to a primitive local timestamp, instead a dictionary is returned that contains the numeric value that translates to the number of milliseconds since midnight of the first of january 1970, also known as an unix epoch. r. reis et al. | i-etc, vol. 6, n. 1 (2020) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt d. distribute using map operations lists lists are encapsulated in ee.list objects. consider the representation of a list containing the previous three timestamp variables in listing 7. listing 7: list dates dates = ee.list([yesterday, today, tomorrow]) exploring parallel processing, the workload could be distributed using the map-reduce paradigm. all objects that support iteration, including ee.list, expose a map function which has a single parameter, that will be a reference to another function, which has a specific set of features: (i) it will accept the item to process as parameter; (ii) returns the result of the process; (iii) must be self contained in the sense that it may not depend on anything that is not in the local scope of the function. so, consider the function mapper() in listing 8, built around these constraints. note that it receives element as a single parameter. it does not have the context of the parameter type, so casting will be necessary. finally it returns the numeric representation of the day. this last particular feature is important to illustrate that the data type of the result of the map function might be different from the elements in dates. in this case, originally, the list has ee.date members and the mapper function returns ee .number instances. listing 8: a parallel processing map function. def mapper(element): return ee.date(element).get("day") distribute=dates.map(mapper) note that distribute is just a representation at the local level. in order to process the implicit operation and obtain the result, getinfo() might be used (see listing 9) to obtain the evaluated list. listing 9: result of distribute evaluation. print(distribute.getinfo()) [3, 4, 5] e. reducing and aggregating using image collections georeferenced bitmaps11 are kept in gee using collections ee.imagecollection of images ee.image which, in turn, may either translate to a single image or sets of multispectral bands. those that are stored in the gee repository are singled out by unique string identifiers. images might contain more than a set of values for each represented coordinate. these values can be part of calculations using discrete or aggregation operations. the calculation may affect a single coordinate or all of them. all these operations are gathered around the concept of “band math” [15] in gee. 11also know as raster images aggregation operation’s parallelism and distribution are supported on reducers. demonstrating this feature, consider listing 10 which filters sentinel 2 [3] data, for the month of may 2019, bounded by the portuguese mainland geometry and excluding all images obscured by more than 20% of clouds. listing 10: reducer operation on a given region. images=ee.imagecollection("copernicus/s2")\ .filterdate("2019-05-01", "2019-05-31")\ .filterbounds(portugal.geometry())\ .filter(ee.filter.lt("cloudy_pixel_percentage", 20)) band1=images\ .first()\ .select("b4") print(band1.reduceregion(reducer=ee.reducer.minmax() ,\ geometry=portugal. geometry(),\ scale=10,\ maxpixels=1e9).getinfo()) {’b4_max’: 6186, ’b4_min’: 392} the variable band1 represents the eldest image (first in the collection), the band “b4” contains the reflectance values for the red wave length. the reducer obtains the minimum and maximum values observed in the set of coordinates bounded by the geometry of the portuguese mainland, using a scale of 10m2, capped by a maximum of 109 coordinates. another perspective is to aggregate values across a set of bands to compute another band. the aggregate operation acts on every set of values in the scope of each coordinate (see listing 11), resulting in a new set of values as a band. listing 11: reducer operation resulting in a band. band2=images\ .select("b4")\ .reduce(ee.reducer.median()) print(band2.getinfo()) {’type’: ’image’ ,’bands’: [{’id’: ’b4_median’ ,’data_type’: {’type’: ’pixeltype’ ,’precision’: ’double’ ,’min’: 0.0 ,’max’: 65535.0} ,’crs’: ’epsg:4326’, ’crs_transform’: [1.0, 0.0, 0.0, 0.0, 1.0, 0.0]}]} in this case, gee uses some syntactic sugar [16] that eases the burden of writing the whole expression. reduce(ee. reducer.median()) can be rewritten simply median(). also note that the result of listing 11 is a single band, named after the original band concatenated with the reducer name, in this case “b4 median”. iii. building a comparison prototype the main objective is to distinguish between: • the code editor which is an integrated development environment, internet browser based, that uses javascript as the support language; and r. reis et al. | i-etc, vol. 6, n. 1 (2020) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt • a python local development environment using the gee api package. so, more than just comparing two apis, this section will differentiate two development platforms. note that there is not a proper efficiency measure between both platforms. the only true difference, that can be discarded for this matter, is the overhead of the local processing. in the case of the code editor it would be difficult to single out the performance of a script running in the context of the internet browser without proper instrumentation of the code. let’s keep in mind that the relevant processing, and thus the performance indicators, depend totally on the google’s infrastructure. fig. 1: sample of the gee code editor profiler after running a script even though it does not make sense to benchmark any of these scenarios, there is a profiling tool in the code editor that may give us some insight on the usage of the google’s infrastructure for a specific script execution, regarding user processing and memory quota. see figure 1 for an illustration of the profiler output after executing a script in code editor. in figure 1, the columns represent:(i) the “compute” cost of the operation; (ii) the “peak mem[ory]” usage of the operation; (iii) the “count” of running instances of the operation; and finally (iii) the “description” of the operation. a. the prototype class a prototype class, named territory, was designed so that the difference between both support languages, using the same feature set, can be exemplified. the structure of the territory class is show in figure 2. this class will wrap partial access to a dataset of the territory boundaries of countries and regions which exists in the gee repository12. the following features will be provided: (i) a list of all country codes contained in the dataset by using the countries method; (ii) a list of all regions13 in the database for a given country code, implemented by the regions method; finally 12specifically the lsib: large scale international boundary polygons, simplified published by the united states department of state 13the concept of region is prone to be equivocal. for instance, in portugal the dataset establishes three regions: mainland, madeira and azores. (iii) the geometry of the boundaries for a given region returning a ee.geometry representation, using the geometry method. fig. 2: unified modeling language (uml) class diagram for the territory class listing 12: territory class implemented using javascript exports.territory = function() { var me = { region_field: "country_na", country_field: "country_co", db: ee.featurecollection("usdos/lsib_simple/2017 "), countries: function() { var operation = this.db .iterate(this.foreach_record(this. country_field), ee.list([])); return operation .getinfo(); }, regions: function(country) { var operation = this.db .filter(ee.filter.eq(this.country_field, country)) .iterate( this.foreach_record(this.region_field), ee.list([])); return operation .getinfo(); }, geometry: function(country, region) { var operation = this.db .filter( ee.filter.and( ee.filter.eq(this.country_field, country), ee.filter.eq(this.region_field, region))); return operation; }, foreach_record: function(field) { return function(record, list) { return ee.list(list) .add(record.get(field)) } } } return me; } note that the first two methods will return native representations of a list of strings as defined in each of the apis support languages but the latter will return a json proxy r. reis et al. | i-etc, vol. 6, n. 1 (2020) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt representation of an ee.featurecollection object (see listing 2 and listing 3 to illustrate the distinction). there is also a conceptual gap between the uml class diagram and the implementation using both supporting languages. neither one supports the visibility scope, and the private members cannot be strictly enforced using the javascript’s syntax (see the implementation in listing 12). the visibility scope is only supported partially, but not constrained, in pyhton using the double underscore notation convention (see listing 13). listing 13: territory class implemented using python import ee class territory: region_field = "country_na" country_field = "country_co" @property def db(self): return ee.featurecollection("usdos/lsib_simple /2017") def countries(self): operation = self.db\ .iterate(self.__foreach_record(self. country_field), ee.list([])) return operation\ .getinfo() def regions(self, country): operation = self.db\ .filter(ee.filter.eq(self.country_field, country))\ .iterate(self.__foreach_record(self. region_field), ee.list([])) return operation\ .getinfo() def geometry(self, country, region): operation = self.db\ .filter(\ ee.filter.and(\ ee.filter.eq(self.country_field, country) ,\ ee.filter.eq(self.region_field, region))) return operation def __foreach_record(self, field): def __iterator(record, list): return ee.list(list).add(record.get(field)) return __iterator the iterator method works similarly to a map function (see listing 8), except it accepts a list and returns another list, instead of a single element. observing both listing 12 and listing 13 it becomes clear that the relevant differences occur due to the diverse syntactical rules, but the object model is the same. b. referencing and reusing in both environments a process was developed in order to output both: the full listing of country codes and the list of region names for portugal. to reuse the class in javascript one could include the class definition in the same script but that would lead to repeating the same code whenever needed with all the associated drawbacks. however, the javascript browser based platform is well built and supports the development of reusable code by enabling the linkage of libraries using two major syntactical contributions to the javascript grammar: (i) a library script maybe referenced using a function (named requires) which has a single parameter, the location of the library file, in the user’s code repository in the code editor; (ii) a predicate (named export) which exposes javascript variables and functions that will be referenced using the previous linked library. for instance, considering a script named bar that implements a function named baz: export.baz = function()... a client script would use this library using: foo = requires("bar") foo.baz() in listing 12 the primitive export is used to expose the territory class, while in listing 14 the reference to a file is done using the function require. the return for this function will be a reference to the file kept in a lib variable. the reference to the territory class is then possible by prefixing it with the lib variable. assuming the class in listing 12 is placed in a file named territory, in the user private assets in the code editor (named users/geeprimer/lib), the main function will be implemented using the script in listing 14. listing 14: main process script in javascript var lib = require("users/geeprimer/lib:territory"); function main() { print("starting..."); var territories = new lib.territory(); print("obtaining country codes:"); var countries = territories.countries(); print(countries); print("obtaining portugal’s regions:"); var regions = territories.regions("po"); print(regions); } main(); the same behavior in python uses the builtin constructs. this sets a first meaningful difference between both apis (see listing 15). suppose the class in listing 13 is placed in a file named territory.py in the same path as listing 15, the main function will be implemented using the script in listing 15. the same behavior might be extended to the standard packaging model used by python developing new modules. another relevant difference, already described in section ii, is the need in python to use the initialization procedure that is implicit when using javascript in the code editor. listing 15: main process script in python import ee from territory import territory r. reis et al. | i-etc, vol. 6, n. 1 (2020) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt def main(): print("starting...") ee.initialize() territories = territory() print("obtaining country codes:") countries = territories.countries() print(countries) print("obtaining portugal’s regions:") regions = territories.regions("po") print(regions) main() c. console output and execution flow both apis implement a console to support basic user interaction. the main script process uses the console to present the user with two human readable lists: one corresponding to the country codes; and another with the region names for portugal’s territorial boundaries. a function print, with the same syntax in both languages, is used to output text to the console. fig. 3: console output of the main script execution in a python environment. the output console in python is the standard output stream (see figure 3). each of the print calls is presented, incrementally, to the user as soon as the data is available and the process flow maybe interrupted without losing the output already produced (neither the execution context). also note that both territory class methods used (countries and regions) return the gee operation result by using getinfo() (see listing 13). in python this is essential to evaluate the operation result in gee as shown in listing 2 and listing 3. the code editor lives in a html document context and the user console is rendered as a html object (see figure 4). the code editor behavior is quite different from the python’s standard approach: (i) the console output is rendered once at the end of the script execution, it does not behave incrementally like a regular text console; (ii) if the process is interrupted all the execution context and console output are fig. 4: console output of the main script execution in the code editor. lost, and during the script execution there is no way to present the user with any kind of progress information; and (iii) the output is rendered using presentation rules that are adequate to a rich graphical user interface (note the expanded panel which exposes the region names list output in figure 4). these behavioral differences are mostly visual and are sideeffects of a deeper distinction between running the same script using a python environment and the javascript’s engine on the internet browser: (i) there is no difference to the internet browser’s javascript engine between a long running script and a poorly written snippet of javascript code, the protection mechanism will prompt the user if the process takes too long to execute; and (ii) controlling the execution flow and recovering from a timeout error is challenging because the script is executing in a single blocking shared thread. finally the code editor makes some assumptions concerning the use of the print function. if the territory class methods countries and regions are modified so that the return value (see listing 12) is the operation instead of using operation.getinfo(), the content of the console in code editor would be the same as the one shown in figure 4. what happens is that the code editor assumes that the user wants to dump the operation value instead of its json representation and implicitly uses the getinfo() method. this happens whenever a gee object is passed to the print function. however, it must be highlighted that this assumption is related to the console output in the code editor. the correct implementation of the territory class is the one in listing 12. d. extracting data one of the most challenging issues related to using the code editor is the way it provides functionality to generate data that can be reused by other processes or integrated in a proper workflow. there is an export object that supports exporting data to a cloud based storage (either google drive or google cloud storage) or to the assets folder of the gee user in the code editor. it is possible to persist data in the form of an image, a video (sequence of images) or a collection of georeferenced data using a set of standard formats like comma separated files (csv ) or other dedicated formats like geojson. r. reis et al. | i-etc, vol. 6, n. 1 (2020) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt in the code editor, the procedure begins with creating an export task using the export object (see listing 16). listing 16: creating an export task in code editor var lib = require("users/geeprimer/lib:territory"); function extract() { var territories = new lib.territory(); export.table.todrive( {collection:territories.db, description:"exportregions", selectors: [ territories.country_field, territories.region_field]}); } extract(); once created, the new task will be presented in the “tasks” list in the user interface (see figure 5). fig. 5: tasks list in the code editor showing the newly created task. when the user starts the task, using the “run” button, a dialog is presented where the original settings for the export task maybe redefined (see figure 6). pressing the “run” button in the dialog the task will be started. fig. 6: tasks list in the code editor showing the newly created task. in this case, once generated, a csv file with two columns with all the country code and corresponding regions’ names will be written to the user’s google drive. even though this process makes sense to most users, whom usage requirements are fulfilled by the code editor, it isn’t easy to integrate it in a process workflow neither is it practical to extract data to be used by other systems. note that, in the example, the user interaction is needed for the process to be executed. if the same data was to be part of an execution flow, the script in listing 17 would extract the same data in a data variable, that could then be used as the input to some other function in python. listing 17: extracting the same data using python import ee from territory import territory def extract(): ee.initialize() territories = territory() data = territories\ .db\ .select( [territories.country_field, territories.region_field], none, false)\ .getinfo() return [feature["properties"] for feature in data[ "features"]] data = extract() note that, if the requirement is to generate a file with some kind of standard format, that could then be reused in a data flow, the resulting structure kept in the data variable could now be stored in a file using one of the many libraries in python, for instance a “pickle” file [17]. listing 18: creating an export task in python import ee from territory import territory def extract(): ee.initialize() territories = territory() return ee.batch.export.table.todrive( collection=territories.db, description="exportregions", selectors=[ territories.country_field, territories.region_field]) task = extract() # execute the task task.start() # the execution status of the task task.status() {’state’: ’ready’, ’description’: ’exportregions’, ’creation_timestamp_ms’: 1587978085189, ’update_timestamp_ms’: 1587978085189, ’start_timestamp_ms’: 0, ’task_type’: ’export_features’, ’id’: ’tnmjecykyc6axmmq2inhxaod’, ’name’: ’projects/earthengine-legacy/operations/ tnmjecykyc6axmmq2inhxaod’} export tasks, like the example in listing 16, might be of use in some other integration schemes: (i) some of the datasets generated by a gee process can be considerably large and it would be inefficient, and sometimes impractical, to reuse using the python api to serve as an intermediate input to some other process; (ii) a dataset resulting from a gee application can have a single objective to be another item in the user assets in the gee repository which might be reused by the user in other gee processes or shared with other gee users; or (iii) a prototype designed in the code editor may not have r. reis et al. | i-etc, vol. 6, n. 1 (2020) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt any other objective other than generate a file to be delivered to a compatible cloud storage. while listing 18 reproduces the same export task that was created in the code editor (see listing 16), there are, however, two differences: (i) the export task is triggered by a call to task.start; and (ii) the execution status can be obtained by calling task.status. the python api fully supports the export tasks feature and allows a gee application to implement asynchronous batch processes. e. imagery the google maps platform is part of the user interface of the code editor and it is fully integrated. points, lines and polygons maybe drawn directly in the google maps pane, and imported to a script, creating variables that describe the geometry drawn by the user. using javascript it is also possible to render georeferenced bitmaps or geometries (set of vectors) as layers in the google maps interface. listing 19: render the territorial boundaries of portugal’s mainland as a layer in google maps using javascript var lib = require("users/geeprimer/lib:territory"); function overlay() { var territories = new lib.territory(); var mainland = territories.geometry("po", " portugal"); map.centerobject(mainland); map.addlayer(mainland, {color: "blue"}, "portugal’ s mainland", true); } overlay(); consider the script in listing 19, it will draw a layer in the google maps pane with the territorial boundaries of the portuguese mainland. the geometry is drawn in blue and the result would be similar to the one in figure 7. fig. 7: google maps pane in the code editor showing portugal’s mainland boundaries in blue. reproducing the same using the python environment is not possible because there is no integration with google maps which is an internet browser based tool. however, it is possible to draw just the portugal’s mainland boundaries using the script in listing 20. using the url representation in url, an internet browser may be used to visualize the image or the resource referenced by the url, and maybe streamed into a local file using a standard python library. listing 20: draw the territorial boundaries of portugal’s mainland in a bitmap and obtain it’s url using python import ee from territory import territory def overlay(): ee.initialize() territories = territory() geometry = territories\ .geometry("po", "portugal") image = geometry\ .draw("blue")\ .getthumburl( {"dimensions":"1024x768",\ "region": geometry\ .geometry()\ .bounds()\ .getinfo(),\ "format": "png"}) return image url = overlay() however, if we observe the generated image, it is noticeable that it is skewed and distorted. this apparent anomaly is the result of drawing the geometry projected in a flat surface instead of the earth’s surface. the rendering of this data has to consider a reference system and geodetic datum [18], in order to make sense. to overcome this issue, we would have to use more features out of the api scope. one option would be to use google colaboratory [14], which is a browser based interface that runs in the context of a remote python environment. it can be used to support a small number of features similar to google maps. in short, the distinction between both platforms is summarized in table ii. table ii: summary of the main differences between the javascript in the code editor environment and using the api in a python environment. javascript code editor python internet browser based tool, no deployment or setup needed. there is a setup process that the user must follow in order to use the api. authentication and initialization are transparent to the user prior to using gee, a initialization step is always needed. code libraries can be used using dedicated grammar contributions and the code editor code repository builtin constructs used to import packages and code is organized using python’s standard rules. r. reis et al. | i-etc, vol. 6, n. 1 (2020) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt table ii (continued) javascript code editor python the console is presented to the user only at the end of the script. if interrupted, the whole context of the execution, up to then, is lost. the console is a true character stream. the execution context, if the flow is interrupted, is kept. data may be extracted in the form of files, in batch processing, and requires human interaction. the supported formats are limited. export task feature is fully supported and can be triggered, and monitored, using code. however, data can be extracted using conventional patterns and can be streamed using other formats supported by python. program flow and error recovery is limited due to the fact that the execution environment is an internet browser. the interaction with gee is similar to any other api in python. integration with google maps is tight, and produces high quality imagery with ease. there is no limitation obtaining georeferenced data, producing imagery requires more effort when compared to the code editor. iv. conclusion gee is a distributed processing platform, leveraged on the google infrastructure, and a multi petabyte repository of georeferenced data, incrementally updated, presented to the user through an unified object model that hides the complexity of the sources, both supported on an api that is available using either javascript or python. the features gee offers significantly ease the burden of working with multi-spectral data compared to a typical on premises use of gis tools. some of the main platform concepts and design patterns were illustrated with code samples that present an overall perspective of the gee as well as exposing some of the syntactical and semantic differences between javascript and python apis. any of the apis developed by google are merely proxies to the gee. the engine’s performance is not warrantied by any sla from google, and is affected by the number of concurrent users, and the usage level, at any time. therefore, there is not a meaningful execution time difference between the internet browser hosted code editor and the python api using a local environment. the code editor has a profiling feature that exposes an insight on the execution cost and memory usage that is of limited utility for this matter. a prototype was implemented, of a class interface, using both javascript and python. the prototype was used to distinguish between all the unique features in the different execution environments. we conclude that, while the code editor is adequate to prototype gee applications that interact with google maps to produce imagery and map overlays, the python environment api is more flexible and is able to integrate seamlessly in a typical software architecture. references [1] noel gorelick, matt hancher, mike dixon, simon ilyushchenko, david thau, and rebecca moore. google earth engine: planetary-scale geospatial analysis for everyone. remote sensing of environment, 202:18–27, 2017. [2] jeffrey dean and sanjay ghemawat. mapreduce: simplified data processing on large clusters. in osdi’04: sixth symposium on operating system design and implementation, pages 137– 150, san francisco, ca, 2004. [3] esa. sentinel 2 user guide. https://earth.esa.int/ web/sentinel/user-guides/sentinel-2-msi, 2019. accessed on 2020-04-25. [4] gdal. gdal geospatial data abstraction library. https: //www.gdal.org/, 2019. accessed on 2020-04-25. [5] nasa. measuring vegetation (ndvi & evi). https: //earthobservatory.nasa.gov/features/ measuringvegetation/measuring_vegetation_ 2.php, 2018. accessed on 2020-04-25. [6] vito remote sensing. product types proba-v. http: //proba-v.vgt.vito.be/en/product-types, 2019. accessed on 2020-04-25. [7] rui s. reis, célia gouveia, nuno datia, and m. p. m. pato. modelo preditivo de recuperação da vegetação afetada por incêndios florestais. in inforum 2019 atas do 11o simpósio de informática, page 461–472. nova.fct editorial, 2019. [8] ogc. ogc geotiff standard. https://www.ogc.org/ standards/geotiff, 2020. accessed on 2020-04-25. [9] qgis. qgis. https://qgis.org/en/site/, 2019. accessed on 2020-04-25. [10] brent n holben. characteristics of maximum-value composite images from temporal avhrr data. international journal of remote sensing, 7(11):1417–1434, 1986. [11] google. earth engine data catalog. https://developers. google.com/earth-engine/datasets/, 2019. accessed on 2020-04-25. [12] google. share your analyses using earth engine apps. https://medium.com/google-earth/shareyour-analyses-using-earth-engine-apps1ac29939903f, 2018. accessed on 2020-04-25. [13] google. earth engine data catalog. https://developers. google.com/earth-engine/, 2020. accessed on 202004-26. [14] google. google earth engine guides, python installation. https://developers.google.com/earthengine/python_install, 2019. accessed on 2020-0425. [15] google. google earth engine guides, band math. https://developers.google.com/earthengine/getstarted#band-math, 2020. [online; accessed 2020-01-15]. [16] p. j. landin. the mechanical evaluation of expressions. the computer journal, 6(4):308–320, 01 1964. [17] python software foundation. pickle — python object serialization. https://docs.python.org/3/library/ pickle.html, 2019. [online; accessed 2020-04-25]. [18] wikipedia contributors. world geodetic system — wikipedia, the free encyclopedia. https://en.wikipedia.org/ w/index.php?title=world_geodetic_system& oldid=938668810", 2020. [online; accessed 2020-02-03]. r. reis et al. | i-etc, vol. 6, n. 1 (2020) id-4 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt https://earth.esa.int/web/sentinel/user-guides/sentinel-2-msi https://earth.esa.int/web/sentinel/user-guides/sentinel-2-msi https://www.gdal.org/ https://www.gdal.org/ https://earthobservatory.nasa.gov/features/measuringvegetation/measuring_vegetation_2.php https://earthobservatory.nasa.gov/features/measuringvegetation/measuring_vegetation_2.php https://earthobservatory.nasa.gov/features/measuringvegetation/measuring_vegetation_2.php https://earthobservatory.nasa.gov/features/measuringvegetation/measuring_vegetation_2.php http://proba-v.vgt.vito.be/en/product-types http://proba-v.vgt.vito.be/en/product-types https://www.ogc.org/standards/geotiff https://www.ogc.org/standards/geotiff https://qgis.org/en/site/ https://developers.google.com/earth-engine/datasets/ https://developers.google.com/earth-engine/datasets/ https://medium.com/google-earth/share-your-analyses-using-earth-engine-apps-1ac29939903f https://medium.com/google-earth/share-your-analyses-using-earth-engine-apps-1ac29939903f https://medium.com/google-earth/share-your-analyses-using-earth-engine-apps-1ac29939903f https://developers.google.com/earth-engine/ https://developers.google.com/earth-engine/ https://developers.google.com/earth-engine/python_install https://developers.google.com/earth-engine/python_install https://developers.google.com/earth-engine/getstarted#band-math https://developers.google.com/earth-engine/getstarted#band-math https://docs.python.org/3/library/pickle.html https://docs.python.org/3/library/pickle.html https://en.wikipedia.org/w/index.php?title=world_geodetic_system&oldid=938668810" https://en.wikipedia.org/w/index.php?title=world_geodetic_system&oldid=938668810" https://en.wikipedia.org/w/index.php?title=world_geodetic_system&oldid=938668810" a survey on the 5th generation of mobile communications: scope, technologies and challenges 1 a survey on the 5th generation of mobile communications: scope, technologies and challenges m. sousaacd, p. vieirabd, m. p. queluzad, a. rodriguesad adepartamento de engenharia electrotécnica e de computadores, instituto superior técnico, portugal bdepartamento de engenharia electrónica e telecomunicações e de computadores, instituto superior de engenharia de lisboa, portugal ccelfinet, consultoria em telecomunicações, lda. dinstituto de telecomunicações, lisboa, portugal marco.sousa@celfinet.com pvieira@deetc.isel.pt [paula.queluz, ar]@lx.it.pt abstract— the 5th generation (5g) of mobile communications will impact the costumers quality of experience (qoe) by addressing the current mobile networks usage trends and providing the technological foundation for new and emerging services. additionally, 5g may provide a unified mobile communication platform, with multiple purposes, leveraging industries, services and economic sectors. in this paper, a 5g tutorial is presented, including the 5g drivers, main use cases, vertical markets and a current status of the standardization process. furthermore, several 5g key enabling technologies are presented, concerning the radio access network (ran) and core network (cn) perspectives. finally, a brief outline over the internet of things (iot) concept and current research topics is presented. keywords: 5g, embb, mmtc, urllc, sdn, nfv, mmimo, mwaves, mec. i. introduction mobile wireless communication networks have been experiencing enormous advances throughout its successive generations. starting with 1st generation (1g), which was an analog technology only for voice calls in the 1980s, industry moved on to 2nd generation (2g), that entered into the digital domain in the 1990s; besides voice, it also supported short message services (smss). the breakthrough of the 3rd generation (3g) was the global access to mobile data services, video streaming, web browsing, e-mail, etc, in the 2000s. the introduction of the 4th generation (4g) eliminated the circuit switching domain to embrace an all internet protocol (ip) network, enhancing data services, in the 2010s. finally, the 5th generation (5g) will be disruptive in the sense that while previous generations had the purpose of connecting people, 5g will connect not only people but also the physical world (things). also, the 5g will provide the technological foundation for a wide range of new applications and services, headed for an increasingly connected world. in the economy field, the concept of general purpose technologies (gpt), identifies technologies whose adoption introduce changes that redefine work processes as well as the rules of competitive economic advantages [1]; the printed press, the internet, and the computer, are a few winning examples. according to [1] from ihs markit, the 5g has the potential to enter the exclusive group of gpts, as a technological breakthrough. in this paper, a state of the art about the 5g is presented. as mentioned, 5g is significantly different from previous generations and will have a significant impact in society and economy. in that sense, besides the 5g technological foundations, also the new services/applications and economic sectors, which are likely to adopt or benefit from the 5g, will be considered. this paper is organized as follows. in section ii, the development scope of the 5g is presented; it includes the technology drivers, the main 5g use cases and the regulatory perspective of 5g. in section iii, an overview of possible technologies to facilitate the 5g networks, is presented; it includes key technologies for both radio access network (ran) and core network (cn). section iv, is dedicated to the internet of things (iot), in light of being one key difference between 5g and legacy networks. finally, in section v conclusions are draw. ii. 5g scope this section provides a wide scope analysis of 5g networks, starting with the main drivers for its development and foreseen use cases. since the mobile network operators (mnos) business approach will be also transformed along with 5g deployment, new vertical players are anticipated, such as the automotive industry, which has shown a strong interest in taking advantage of the upcoming networks. the efforts with standardization and regulation, from entities such as the 3rd generation partnership project (3gpp) and international telecommunication union (itu), are crucial for the first 5g commercial deployments, and will be also overviewed in this section. a. 5g drivers the 5g networks drivers can be loosely classified as costumer or industry associated. with respect to mobile network costumers, there are several key trends. firstly, the number of mobile subscriptions is growing at 6% annual rate [2], reaching 7.8 billion in the third quarter of 2017 (q3), and the respective grow rate for mobile broadband subscriptions is around 20% year-on-year [2]. besides the mobile subscriptions up rise trend, the mobile data traffic generated by each subscriber is also increasing. latest statistics account 65% annual increase i-etc: isel academic journal of electronics, telecommunications and computers iot-2018 issue, vol. 4, n. 1 (2018) id-1 http://journals.isel.pt in mobile traffic between the 3rd quarter of 2016 and 2017 [2], and the momentum is expected to continue and reach a 8-fold increase in 2023. currently, more than 50% of the mobile traffic is video streaming [2], being the leading application in traffic generation. the 4g networks are not fully developed yet, for instance self-organizing networks (son) features, as covered in [3] [4] can improve network performance, also mnos are starting to evolve their long term evolution (lte) networks to long term evolution advanced (lte-a), enhancing the 4g network performance. nonetheless, even considering such solutions, 4g networks will not be able to cope with the subscribers increasing demand. furthermore, considering services such as video streaming, where the uprising trend in video resolution will generate more content in 4k or even 8k, or considering the appearance of new services, like virtual reality and augmented reality, its feasibility with the current mobile networks is limited. overall, such services require high throughputs, and constitute the first main 5g use case, the enhanced mobile broadband (embb) [5]. on another perspective, the fourth industrial revolution known as “industry 4.0” is starting. according to [6], industry 4.0 is defined as the digitalization process of the manufacturing sector; it consists on embedding sensors for monitoring all products and equipment, using cyber physical systems, and in applying cognitive analysis to all collected data. from the technological point of view, 5g will be an important facilitator in the full realization of the industry 4.0 concept. accordingly, the second main use case, defined by itu, is the massive machine type communications (mmtc) [5]. in essence, it is characterized by a large number of connected devices, collecting and sending non-delay-sensitive data [5]; it can also be seen as one of the applications of iot [7]. lastly, itu also defined the ultra-reliable low latency communications (urllc) [5] requirements. this will establish the technological foundations for the self-driving [8] endeavor. there is a strong economic interest from the automotive industry in developing not only the autonomous vehicles, but also a connected and intelligent infrastructure. it requires both vehicle-to-vehicle (v2v) communications and vehicleto-infrastructure (v2i). along this way, the transmitted data is as time sensitive as intolerant to failures and errors. satisfying the above mentioned requisites entails in the development of wireless communication systems with high availability, ultra reliable and with very low latency. similarly to the automotive industry, other sectors may also develop new services based on urllc, such as wireless tele surgery (wts) [9] in the health care sector. b. use cases figure 1 presents the main 5g applications in a three dimensional space, where each dimension corresponds to one of the main 5g requirements (embb, mmtc and urllc). these requirements are deeply analyzed in the following subsections. 1) enhanced mobile broadband: the embb essentially deals with the human centric use cases, aiming to deliver fig. 1: usage scenarios of imt for 2020 and beyond [5]. multimedia content and data services. it depends on the ability to provide high throughputs, especially when considering high quality video streaming or 3d video services. in a recent survey [10], the authors identified that network throughput and security are the factors that mostly condition the subscribers expectations about the upcoming 5g networks which is aligned with the embb requirements. in [11], the authors classified embb as the first 5g “killer app”, based on a costumer survey where the experienced bottlenecks, of smartphone users, were identified as one of the factors that most impact the quality of experience (qoe). itu has defined minimum requirements for the main network capabilities [12]. nonetheless, depending on the usage scenario, some capabilities are more relevant than others. this relative importance is depicted in figure 2. fig. 2: relative importance of key capabilities in different usage scenarios [5]. considering the embb, the key capabilities are: network energy efficiency, area traffic capacity, peak data rate, user experienced data rate, spectrum efficiency and mobility. in [11], and considering the ran, the authors identified three pillars to address the requirements of embb, densify networks, deliver higher spectral efficiency and the usage of new spectrum bands. in this matter, the development and standardization of the 5g new radio (nr), that is an ongoing process, will provide the technology to deliver the performance requirem. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt ments for embb. the main key aspects of the 5g nr are the shared access scheme, ortoghonal frequency division multiplexing (ofdm) based or non-orthogonal multiple access (noma), new system architectures enabling cloud radio access network (c-ran) and network slicing, new spatialdomain processing techniques including massive multipleinput multiple-output (mmimo) and 3d beamforming, and also the use of millimeter wave (mmwave) bands. 2) massive machine type communications: the services belonging to the mmtc are such that machine-type devices are used for monitoring, sensing and metering. on a 5g perspective, the network is expected to support a large number of these machine-type devices, typically requiring low throughput and sparse communications. also, it is required to support a high connection density of these devices [13]. basically, 5g mmtc encompasses the iot use cases which are non delaysensitive, using a standardized technological solution. the 3gpp group has already standardized the narrowband internet of things (nb-iot), in lte release 13 and lte release 14, to provide wide-area mmtc connectivity for iot [14]. besides this licensed band standard, other proprietary solutions, such as sigfox and lora, were developed in unlicensed bands [15]. nonetheless, these solutions are struggling in a scenario where the number of devices significantly overpass the resources. as these solutions rely on orthogonal transmission principles, in recent years non-orthogonal strategies have been also proposed to accommodate more users than the more classical orthogonal approaches [13]. the 5g embb and the mmtc require two completely different communication system designs. not only the embb service category is heavily focused on the downlink communication, while the mmtc service is focused on the uplink, but also the embb requires high packet sizes and throughputs, whereas the mmtc requires low values of these parameters [16]. consequently, the mmtc advocates a set of technologies that are quite different from those of the embb use case. the first main difference is, as mentioned before, the medium access scheme, in order to support the highest number of connected devices. while the orthogonal medium access tightly sets the available resources according to the number of supported users, a non-orthogonal medium access allows some degree of resources overloading [16]. moreover, the grantbased access, used in lte, requires a good prediction of uplink requests and additional signaling, which is not ideal for the mmtc scenario. in that sense, a grant-free solution is expected to enhance the requirements feasibility of mmtc [16]. both solutions imply a complexity increase from the base station in order to simplify the devices complexity and achieve the stipulated requirements. technologies such as sparse code multiple access (scma), compressed sensing based multi user detection (cs-mud) and continuous phase modulation (cpm) can be strong candidates to enable massive access. 3) ultra-reliable and low latency communications: the urllc supports the applications that are latency sensitive as remote control and autonomous driving [17]. another emerging application is tactile internet [18]; it allows humans to control real and virtual objects wirelessly, and hence, equivalent to human touch, visual and auditive perceptions would be transmitted seamlessly via data networks. this application requires 1 ms end-to-end latency [19]. other services for 5g urllc require latencies in the range of 1-10 ms. currently, lte networks are characterized by latencies in the range of 30 to 100 ms [13]. this end-to-end latency values are due to the best-effort policies, typically used in the backbone network. in order to comply with the latency requirements of the 5g urllc, changes have to be conducted not only in the backbone network but also in the ran. the use of software defined networks (sdns), network function virtualization (nfv), and the concept of network slicing, can establish dedicated connections for urllc services [17]. likewise, the use of multi-access edge computing (mec) can reduce the latency even more [20]. on the ran, given that a large portion of the latency is introduced by the control signaling [13], the communication overhead for urllc has to be reduced. to accomplish this, the packet and frame structure, and the scheduling schemes, have to be revisited [17]. the use of polar codes for large-sized packet and of sparse vector coding (svc) for small-sized ones are technological enablers to achieve the latency requirements. c. new players even though the number of mobile subscribers has been rising, the mnos revenues have been flatting out in the past years [21]. with the anticipated capabilities of 5g, mnos can achieve a much larger growth, addressing key challenges in digitalization of manufacturing, automotive and other industries, also known as vertical markets. in this ecosystem, mnos can become, besides network developers, service enablers or even service creators [21]. in this new reality, the role of mnos, might evolve from business-to-consumer (b2c) to business-to-business (b2b) providers. in the following text, an overview of main vertical markets, in a 5g perspective, is presented. 1) automotive: the automotive industries have been adopting several connectivity technologies pursuing the long term goal of autonomous driving. in that sense, the automotive industry is clearly interested in the possibilities that 5g might enable. the 5g can improve the vehicle-to-everything (v2x) communications [22] and also the in car “infotainment” [23]. other new use cases such as platooning [24] or teleoperated vehicles [25] might be developed with the support of 5g networks. 2) manufacturing: within the manufacturing industry, and as pursued by the industry 4.0, 5g networks will be able to provide the underlying unified communication platform to fulfill the digital transformation [26]. the 5g can be a key enabler for remote assistance and robot control, logistics tracking or process automation within factories [23]. throughout, efficiency improvements and automation, all over the supply chain and product’s life cycle, are expected. 3) agriculture: even more traditional sectors, such as agriculture, are undertaking profound technological innovations. with iot, agriculture can move to smart farming and precision agriculture [23]. the baseline is to monitor crop yields, moisture levels and/or terrain, providing more data to support m. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt assertive farming decisions. the 5g can provide a robust network for iot and also for remote control of the farming machinery [23]. 4) energy and utilities: regarding the energy and utilities environment, the energy industry is facing some challenges; the increased electricity consumption associated with some degree of uncertainty in modeling the demand, and considering several sources of energy generation (renewable sources), contribute with additional threats. in this complex scenario, emerging concepts as smart metering and smart grids [27], have been investigated. the smart metering can be considered part of the mmtc applications whereas the smart grid concept might be more challenging. the smart grid concept requires communication between sensors, control systems, energy generation and storage to monitor and to optimize the grid in realtime. the 5g can support this use case, specially where fiber based access is not cost effective [28]. d. regulation the regularization and the full standardization of 5g is a lengthy process, involving several entities and organizations, private and public ones. around 2014, the international telecommunication union radiocommunication sector (itur) started working in the definition towards the 5g technology performance requirements, as outline in figure 3. fig. 3: timeline for imt-2020 in itu-r [29]. within itu-r, the 5g is known as international mobile telecommunications 2020 (imt-2020). in 2016, itu-r produced the first draft of the imt-2020 performance requirements. the submitted proposals, for the new radio interfaces, to be included in imt-2020 specification, will be evaluated by independent external groups. the whole process is expected to be due by 2020, with the approval of the specifications [29]. meanwhile, the 3gpp is developing its proposal for the imt-2020 call from itu-r. the final specifications for the imt-2020 are planned to be completed with release 16, by the end of 2019, as shown in figure 4. the 3gpp group has released an early drop of release 15 containing the non-standalone (nsa) 5g radio specifications. the full release 15 will include the standalone version of the 5g nr by mid 2018. while the release 15 focus on the embb, release 16 is expected to evolve more in depth the mmtc and the urllc specifications. fig. 4: 3gpp timeline towards 5g [30]. iii. technologies in the previous section, the 5g environment and the three major use cases were considered. several technologies were identified as key enablers to meet the 5g requirements. in this section, these technologies are presented considering ran and cn domains, respectively. moreover, the different network architectures that are being developed within the 3gpp group, are also considered. a. architecture regarding the upcoming 5g network architecture, there are two distinct architecture approaches: the nsa 5g architecture and the standalone 5g architecture. the 3gpp group, in a early drop version of release 15, focused their effort mainly towards the specification of the nsa 5g network. the complete release 15 should include also the standalone 5g architecture specification. the left side of figure 5 presents the nsa architecture, that should be used on the first commercial deployments of 5g; the right side of the figure 5 displays the standalone option for 5g. fig. 5: proposed 5g architectures in 3gpp release 15 [31]. according to the specifications, the nsa 5g architecture, should have a 5g base station (bs) anchored in a lte network. a user equipment (ue) is expected to be connected to the lte bs, using this link for user and control plane, and while connected with a 5g bs for user plane only. also, the 5g bs is connected with the lte evolved packet core (epc). considering the standalone option, the 5g ran is connected m. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt only with the 5g core network (5g-cn) providing both user and control planes, forming a full 5g network. b. radio access network taking into account the 5g ran, there are several expected evolutions compared with current networks, that are described next. 1) massive mimo: traditional multiple-input multipleoutput (mimo) systems, improve the signal-to-noise ratio (snr) in uplink due to the diversity gain, mitigating the fading effects and increasing capacity due to the achieved spatial multiplexing gain [32]. there are two major mimo types: single-user mimo (su-mimo) and multi-user mimo (mumimo). whereas su-mimo allocates the time-frequency resources to a single user, mu-mimo simultaneously serves several users in the same time-frequency resource using beamforming [33]. finally, mmimo is a form of mu-mimo [33], where the number of bs antennas is much larger [34] than in mu-mimo. the mu-mimo requires a highly reliable channel state information (csi), which is unlikely and complex in practical scenarios [32]. additionally, in [35] the authors showed that when considering a large number of antennas at the bs mu-mimo allows the simplest sort of precoding on the forward link and processing on the reverse link. thus, mmimo deliver high throughputs (high spectral efficiency) reliably, both in the uplink and downlink channels, and in fast-changing propagation environments, without the effects of uncorrelated noise and fast fading. it only remains the inter-cellular interference due to pilot contamination [35]. even though, work is being developed to mitigate the pilot contamination using: protocol based methods [36], blind methods [37], precoding [38], genetic algorithm (ga) based allocation scheme [39] or the use of dual pilot sequences [40]. hence, mmimo will be a key enabler for the 5g use cases. moreover, combining the high antenna gains (mmimo) with large available bandwidths (mmwave), 5g network throughput barriers can be pushed even further. 2) millimeter waves: the limited available spectrum, and its cost, have always been a concern for mnos. in that sense, efforts such as spectrum refarming, through the re-purposing of the terrestrial tv spectrum [34], or by refarming the 2g bands to other technologies, have been implemented. also, spectrum sharing techniques [41] [42] have been proposed. nonetheless, some open issues remain to be solved [34], as how to increase the available bandwidth at microwave frequencies. alternatively, there is a great available bandwidth at the mmwave, ranging from 3 to 300 mhz [34]. due to the associated high pathloss for higher frequencies, the use of mmwave towards 5g networks was initially considered for short-range indoor locations [43]. more recently, it has been tested in outdoor non-line-of-sight (nlos) environments, with satisfactory results [44] [45]. moreover, combining mmwave with other techniques, such as beamforming, can enhance the outdoor mmwave performance [46]. besides some regulatory constraints, key technical constraints prevent the use of mmwaves. the high pathloss, the assumption of needed line-of-sight (los), the infeasible nature of some hardware components due to technical limitations and the high doppler shift, remain as key challenges [47]. more recently, all these challenges have been subject of important developments. regarding the high pathloss, it can be mitigated with the use of smart antennas, with high gains [45] [48]. the los requirement can be solved by exploiting the multipath reflections [45]. considering the hardware limitations, new antenna array designs have been proposed [49], but there are still some challenges, as the linear relationship between power consumption and the sampling rate, in the analog to digital converters, which is even more critical at millimeter bands [49]. finally, the doppler shift is also mitigated when considering directional antennas [48]. aside from the mmwave applications for the ran, backhaul and relay solutions [50] are also available. 3) others: some other 5g ran technologies, not included in the previous items, can be considered. the ran random access scheme proposed in the 3gpp release 15, is ofdm based. nonetheless, new waveform candidates are expected in future releases. for instance, non-orthogonal waveform enables higher spectral-efficiency compared with the orthogonal ones [51]. some of the non-orthogonal access schemes are the noma [52], the scma [53], the multi-user shared access (musa) [54], the pattern division multiple access (pdma) and successive interference cancellation amenable multiple access (sama) [55]. all these schemes achieve higher throughputs, or spectral-efficiencies, than orthogonal frequency-division multiple access (ofdma) based schemes. other key technology towards 5g is full-duplex (fd) wireless systems, which enables a radio transceiver to receive and transmit simultaneously in the same frequency and time frame [56]. it can potentially double the system capacity and increase the spectrum utilization efficiency. the main challenge of these systems is self-interference (si) mitigation [57] [58]; in fact, although feasible, some practical imperfections still limit its performance [56]. aiming to target the spectrum shortage problem and the problem of spectrum under-utilization (spatial and/or temporal), the dynamic spectrum sharing (dss) concept has been investigated [59]. another, key technology towards 5g is network slicing [60]. the concept was initially proposed for the 5g-cn [61] and was extended for end-to-end (e2e) network slicing. in [61], the authors evaluate the impact of network slicing in the ran, namely traffic differentiation, efficient management mechanics to setup and operate new slices, among others. c. core network moving on to the 5g-cn, there are also new technologies and concepts being introduced. 1) cloud radio access network: the current ran architecture in mobile networks contributes to the network resources underutilization [62], in the sense that the network load at different locations and time instances varies due to the user movements; hence, some bss can be overloaded while m. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt others are idle (e.g., office locations vs. residential areas during the day). consequently, the bs processing power is being underutilized in some bss and used at the maximum in other bss [63]. to overcome this challenge, the concept of c-ran was proposed [64], which consists in physically separate the baseband units (bbus) from the remote radio units (rrus), by centralizing the bbus into a shared bbu pool. under a c-ran architecture, depicted in figure 6, all computational resources are concentrated in a cloud platform which includes the bbu pool; the bs contains only the rru to receive and transmit the radio signals to the cloud platform [62]. fig. 6: schematic of c-ran architecture [65]. the bbu pool will serve a particular area providing all the processing to the macro and small cells, allowing bss to share computational resources. the distance between the cloud platform and the bss can be up to 40 km where the distance limitation comes from the propagation delay [64]. the fronthaul communications are analyzed in dept in [66]. another advantage of the c-ran architecture is the network energy efficiency [64], as a result of a more efficient use of the computational resources. additionally, as the bss became less complex, containing only the rru, the deployment and maintenance cost of new bs is expected to drop [64]. the cran concept will take a key role towards the 5g networks. 2) multi-access edge computing: another important concept towards 5g is the mec. recent years have seen a paradigm shift towards decentralized computing and cloud platforms, with its realization in mec, in mobile communications. this aims to push computing, storage and control resources to the edges of the network, thus near the final users. in the mec architecture, presented in figure 7, the latency experienced by the users is reduced, while enabling a more efficient usage of the network backhaul and of the core network [67]. mec results from a strong synergy between information fig. 7: mobile edge computing architecture [68]. technology (it) and mobile networks domains, which is a current trend, in the sector. thus, the mec, can leverage the development of a wide range of new services and applications, especially if information such as contextual information and location awareness, is used to deliver, in realtime, a customized mobile broadband experience to users [69]. also, as mec offers an open radio network edge platform, mnos can monetize it, by allowing third-parties to access the storage and processing capabilities; this may facilitate service enhancements, or even new services, towards not only mobile subscribers but also to enterprises and vertical segments. this should be a secure cloud platform that, through application programming interfaces (apis), shares access to third-parties [69]. similar to mec, there are several related concepts, such as mobile cloud computing (mcc), local cloud, cloudlet and fog computing. mcc takes advantage of mobile computing and cloud computing to enable virtualized computing and storage resources [68], for mobile end-users. it provides all resourceintensive computing executed in clouds without a device with a powerful configuration. also, it extends the battery life and storage of the end-users devices. local cloud is managed by internal or external sources and are intended to provide services exclusively to a group or institution [68]. in [70], the authors proposed a cooperative scheduling algorithm to manage the local cloud resources together with internet cloud resources. cloudlet is an emerging paradigm, where a small-box data center is deployed at one wireless hop away from mobile devices [68]. it is the middle entity between the mobile device and the cloud, with data routing and security functionality [71]. thus, cloudlet aims latency and resource sensitive mobile applications. it can be deployed in public places such as hospitals, shopping centers, etc. fog computing is also known as edge computing, supporting ubiquitous connected devices [68]. in this concept, processing is carried out in the local area network or in the iot gateway. in fog computing, one could retrieve data from several sensors and act accordingly, without having to access a remote cloud platform [68]. as a concluding remark, mec is highly complementary with c-ran [72]. not only the collocation of both technologies allows a more cost effective investment for mnos but also the mec apis can provide access to ran information m. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt which, otherwise, would not be trivial, enabling other services. 3) software defined networks/network function virtualization: mnos have high expectations on these two technologies, sdn and nfv, as they promise to reduce network costs, improve the network scalability and flexibility and provide the base ground for a more dynamic and efficient network. firstly, sdn is a centralized networking paradigm where there is a key separation between control and user data [73] [74], as seen in figure 8. fig. 8: sdn architecture overview [73]. moreover, as the network control functions are centralized in one or more sdn controllers, it enables to simplify the data forwarding of applications and network services [73]. the communication between control and data planes is managed through a communication protocol, being openflow the most used [75]. this is an open source protocol which is controlled by open networking foundation (onf) [76], and allows to access the flow tables [75] that control the traffic routing in switches and routers. nonetheless, other protocols exist, such as the internet engineering task force (ietf) forces [77] protocol. as already stated, a key element in a sdn architecture is the sdn controller. it should be not only a platform for deploying sdn applications but also a sdn application development environment [78]. there are several controller implementations, such as the opendaylight [78] or the onos [79]. overall, sdn applications interact with the sdn controllers using, as interfaces, representational state transfer (rest) and javascript object notation (json); the sdn controller uses communication protocols such as openflow [75]. thus, sdn presents itself as an alternative to standardized networking protocols, by providing the same role as a centralized software application. regarding the nfv [80], it is basically the process of replacing dedicated hardware with software instances which run on cloud environments or general purpose servers [73]. within this concept, each conventional network function (nf) runs in a virtual machine (vm) or even in multiple vms. each implementation of a nf, using vms, is called virtual network function (vnf). these instances are deployed and executed in a network function virtualization infrastructure (nfvi). this is composed by physical resources (computing, storage and networking) which are used through a virtualization layer by the vnfs. ultimately, and using as example the lte epc, each core entity (e.g., mobility management entity (mme), serving gateway (s-gw), etc.) could be virtualized as vnfs, where each type of vnf, by forming a common pool, can be scaled independently and according with the network requirements and resources [73]. towards 5g, not only nfv but also sdn will be key technologies. many components of a 5g network can be turned into vnfs, allowing accelerated service deployment, when compared to the traditional hardware deployment, network flexibility, scalability and capacity. above all, vnfs enable network slicing, where the same physical infrastructure can be used to provide multiple and independent logical networks, where the users experience similar conditions as having a dedicated physical infrastructure [81]. additionally, even though sdn and nfv are independent technologies, both can be enhanced when used together due to their complementary traits. explicitly, nfv can serve an sdn architecture by virtualizing components such as the sdn controllers or the forwarding data entities. also, sdn can serve nfv by allowing programmable and dynamic network connectivity between vnfs [73]. this synergy between sdn and nfv is called software defined network virtualization (sdnv) [82], and is an emerging research area. both sdn and nfv have the potential to redefine the evolution of network architectures and the potential to be key enablers for the 5g networks. iv. internet of things although iot [7] has been an active research topic in the past years, the concept of a network of smart devices has been around since the mid 80s. concretely, mark weiser’s paper about ubiquitous computing [83], in 1991, set some cornerstones of the actual iot concept. the iot concept of a network of connected physical objects, cyber-systems and sensors everywhere, combined with recent technological advances, opens up the range of iot applications in different environments, which are being proposed and developed. besides, the applications mentioned in section ii-c, the concept of smart home [84] and the smart cities are other examples of iot applications. extensive work has been developed in creating applications towards enhancing the user lifestyle, especially through the retrieval and analysis of relevant city information [85]. also, the concepts of intelligent transportation system (its) and smart healthcare have been receiving important contributions by the iot community [86]. a. design requirements to fully implement the proposed iot use cases, including the 5g mmtc and urllc scenarios, there are some main key design principles. a low device (e.g., a iot sensor) cost will be a key enabler for mass-market iot applications within the 5g mmtc use case. the iot network deployment cost, including both capital expenditure (capex) and operating expense (opex) must be minimized in order to provide massive iot through a feasible economic perspective [86]. in this scenario, m. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt the usage of cellular networks, compared to dedicated lowpower wide-area (lpwa) networks, might be preferred. also, a key point is energy efficiency, mainly on the iot device side. considering that these devices are battery powered and are expected to have a time span of years, without human intervention, energy efficiency is crucial. thus, the development of lightweight protocols and scheduling optimization are important research areas [86] [87]. other design principle, specially towards smart metering, is the coverage requirement. some smart meter locations are in deep indoor scenarios, which require enhanced coverage [88]. another key topic in the iot design is the security and privacy of personal data [89]. in [89], the author describes extensively the main concerns in providing security in a iot environment. b. architecture the iot generic architecture is organized in three layers: the application layer, the transport layer and the sensing layer [90]. the sensing layer is composed by all cyber-physical objects and sensors. all these devices may collect any kind of data, which is then delivered to the application layer through the transport layer. in the application layer, the data is aggregated and analyzed using intelligent computing, in order to extract valuable information or to trigger actions towards the sensing layer devices [90]. a general iot architecture is presented in figure 9. fig. 9: general architecture of iot [90]. nonetheless, since it is expected that iot networks will contain millions of devices, large scale iot (scalable and flexible architecture) is required. thus, in [91] the authors proposed a self-configuration peer-to-peer architecture. this architecture type, provides automatic discovery mechanisms, enabling the absence of human intervention in the configuration phase. other architectures have been proposed, such as the sdn [92] and the cloud computing based [93]. c. communication technologies iot networks have been widely investigated on the last years, resulting in several iot communication technologies, which solve the transport layer on the iot general architecture. there are three types of technologies: long-range, short-range and cellular networks [86]. long-range networks, or lpwa technologies, are among the most popular iot approaches. one of them is lora [94], which is a physical layer protocol targeting low-cost, low-power and long-range communications [86]. lora architecture is a star shaped network where each device has a direct connection with a lora gateway. also in the long-range iot, sigfox [95] technology offers end-to-end iot connectivity. sigfox relies on ultra-narrowband (unb) communication technologies, well adapted to a wide range of conventional iot use cases, that rely on sparse and low throughput communications requirements [96]. others, as dash7 [97] or weightless [98] are also promising long-range solutions. on the short-range iot, one of the most used is the bluetooth [99]. even though bluetooth was standardized by the institute of electrical and electronics engineers (ieee) as a communication technology for replacing wires in mobile devices, it has evolved to other applications, especially in iot smart home. the main drawback of bluetooth is the restriction of only one-to-one communication. in that sense, the bluetooth smart mesh working group was proposed to standardize a new mesh architecture aimed at iot use cases. zigbee, which is developed on top of the physical and data link layers defined in ieee 802.15.4, is also a shortrange communication base for iot [86]. the main difference of zigbee compared with bluetooth, is the range. whereas bluetooth operates around a 50 meters range, zigbee can provide service through hundreds of meters. also in the short-range communication domain the wireless local-area network (wlan) technology is widely used. initially, it was designed to support high bandwidth communications between devices. as it does not verifies the modern iot network requirements, ieee proposed a low power wlan, ieee 802.11ah [100], as an amendment to the legacy standard. wlan based iot applications include parking metering, autonomous lightning, smart security, smart home thermostats, etc., [101]. while iot essentially aims to interconnect a great number of devices and extract value from all collected data, a new paradigm has recently emerged to enhance even more iot, called cognitive internet of things (ciot) [102]. the ciot arises from the application of cognitive abilities to the iot network. within the scope of ciot, heterogeneous smart objects inter-operate through a cognitive centralized entity. this adds an intelligent decision making layer to conventional iots networks [102]. compared with iot, ciot is a new and hot research topic. v. conclusion the next generation of mobile communications, 5g, is around the corner. besides mnos, several economic sectors glimpse at 5g as a key promoter for socio-economic development. clearly, the mobile network costumers are enthusiastic concerning the upcoming 5g networks and associated services. m. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt the evolution of the current mobile networks to 5g is expected to be gradual, minimizing the capex for mnos. the first 5g commercial deployments are expected by the end of 2018, in a nsa solution, where 5g bss are anchored in the lte epc. in this paper, apart from an overview of the 5g ecosystem, several key technologies were pinpointed. concerning the 5g ran, mmimo and mmwave will be crucial to accomplish the embb requirements. nonetheless, these two topics are still open research areas. in mmimo systems, the channel estimation techniques, pilot contamination and the rich scattering environments dependence are still being investigated. also, the joint operation of mmimo and mmwave is being explored. the 5g-cn will constitute an advance of mobile core networks, being empowered by several it driven concepts. cran, mec, sdn and nfv are candidate solutions heading for the 5g-cn, where work is in motion to solve associated open issues. overall, these technologies are associated with practical constraints, regarding system inter-operation, orchestration, resource management, flexibility and scalability. the association of multiple technologies, from the ones mentioned, is again a research area of importance. also, e2e network slicing is a promising solution towards the development of the future 5g networks. as open challenges, dynamic slicing, end-to-end slice provisioning, slice isolation, are front runners. as a concluding remark, from the technological point of view, 5g can be considered an evolution of lte, specially in the early 5g phase. from a socio-economic point of view, it certainly can be seen as a revolution. vi. acknowledgements the authors would like to thank fct for the support by the project uid/eea/50008/2013. moreover, our acknowledgement concerning project mesmoqoe (n◦ 023110 16/si/2016) supported by norte portugal regional operational programme (norte 2020), under the portugal 2020 partnership agreement, through the european regional development fund (erdf). references [1] k. campbell, j. diffley, b. flanagan, b. morelli and b. o’neil . the 5g economy: how 5g technology will contribute to the global economy. economic impact analysis, ihs economics and ihs technology, 2017. [2] ericsson. ericsson mobility report. technical report, ericsson, 2017. [3] m. sousa, a. martins and p. vieira. self-diagnosing low coverage and high interference in 3g/4g radio access networks based on automatic rf measurement extraction. in proceedings of the 13th international joint conference on e-business and telecommunications, icete 2016, pages 31– 39, portugal, 2016. scitepress science and technology publications, lda. [4] m. sousa, a. martins and p. vieira. self-optimization of low coverage and high interference in real 3g/4g radio access networks. i-etc: isel academic journal of electronics telecommunications and computers, 3(1), jan 2018. [5] itu-r. imt vision framework and overall objectives of the future development of imt for 2020 and beyond. recommendation itu-r m.2083-0, itu, 2015. [6] mckinsey digital. industry 4.0 how to navigate digitization of the manufacturing sector. technical report, mckinsey, 2015. [7] itu-r. series y: global information infrastructure, internet protocol aspects and next-generation networks. technical report, itu, 2012. [8] x. krasniqi and e. hajrizi. use of iot technology to drive the automotive industry from connected to full autonomous vehicles. ifac-papersonline, 49(29):269 – 274, 2016. 17th ifac conference on international stability, technology and culture tecis 2016. [9] d. soldani and f. fadini and h. rasanen and j. duran and t. niemela and d. chandramouli and t. hoglund and k. doppler and t. himanen and j. laiho and n. nanavaty. 5g mobile systems for healthcare. in 2017 ieee 85th vehicular technology conference (vtc spring), pages 1–5, june 2017. [10] a. seetharaman and n. niranjan and v. tandon and s. devarajan and m. k. moorthy and a. s. saravanan. what do customers crave in mobile 5g?: a survey spotlights four standout factors. ieee consumer electronics magazine, 6(3):52–66, july 2017. [11] qualcomm and nokia. making 5g a reality: addressing the strong mobile broadband demand in 2019 & beyond. technical report, qualcomm and nokia, 2017. [12] itu-r. minimum requirements related to technical performance for imt-2020 radio interface(s). report itu-r m.24100, itu, 2017. [13] ji, h. and park, s. and yeo, j. and kim, y. and lee, j. and shim, b. ultra reliable and low latency communications in 5g downlink: physical layer aspects. arxiv e-prints, april 2017. [14] a. hoglund and x. lin and o. liberg and a. behravan and e. a. yavuz and m. van der zee and y. sui and t. tirronen and a. ratilainen and d. eriksson. overview of 3gpp release 14 enhanced nb-iot. ieee network, 31(6):16–22, november 2017. [15] u. raza and p. kulkarni and m. sooriyabandara. low power wide area networks: an overview. ieee communications surveys tutorials, 19(2):855–873, secondquarter 2017. [16] c. bockelmann and n. pratas and h. nikopour and k. au and t. svensson and c. stefanovic and p. popovski and a. dekorsy. massive machine-type communications in 5g: physical and mac-layer solutions. ieee communications magazine, 54(9):59–65, september 2016. [17] p. schulz and m. matthe and h. klessig and m. simsek and g. fettweis and j. ansari and s. a. ashraf and b. almeroth and j. voigt and i. riedel and a. puschmann and a. mitschele-thiel and m. muller and t. elste and m. windisch. latency critical iot applications in 5g: perspective on the design of radio interface and network architecture. ieee communications magazine, 55(2):70–78, february 2017. [18] m. simsek and a. aijaz and m. dohler and j. sachs and g. fettweis. 5g-enabled tactile internet. ieee journal on selected areas in communications, 34(3):460–473, march 2016. [19] itu-t. the tactile internet. itu-t technology watch report, itu, 2014. [20] j. liu and q. zhang. offloading schemes in mobile edge computing for ultra-reliable low latency communications. ieee access, 6:12825–12837, 2018. [21] ericsson. the 5g business potential. technical report, ericsson, 2017. [22] ma. vieira, m. vieira, p. vieira, p. louro. vehicle-to-vehicle and infrastructure-to-vehicle communication in the visible range. sensors and transducers, 218(12):40–48, dec 2017. [23] dotecon ltd and axon partners group. study on implications of 5g deployment on future business models. no berec/2017/02/np3, dotecon ltd and axon partners group, 2018. m. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt [24] 5gpp. 5g automotive vision. technical report, 5gpp, 2015. [25] huawei. 5g unlocks a world of opportunites. technical report, huawei, 2017. [26] 5gpp. 5g and the factories of the future. technical report, 5gpp, 2015. [27] 5gpp. 5g and energy. technical report, 5gpp, 2015. [28] arthur d. little. creating a gigabit society the role of 5g. technical report, vodafone, 2017. [29] itu. itu towards imt for 2020 and beyond. https: //www.itu.int/en/itu-r/study-groups/rsg5/ rwp5d/imt-2020/pages/default.aspx. accessed: 2018-05-29. [30] 3gpp. preparing the ground for imt-2020. https://http: //www.3gpp.org/news-events/3gpp-news/ 1901-imt2020_news. accessed: 2018-05-29. [31] nokia. 5g status in 3gpp and future direction past phase 1 of 5g. technical report, nokia, 2017. [32] f. jameel and faisal and m. a. a. haider and a. a. butt. massive mimo: a survey of recent advances, research issues and future directions. in 2017 international symposium on recent advances in electrical engineering (raee), pages 1– 6, oct 2017. [33] j. brady and a. sayeed. beamspace mu-mimo for highdensity gigabit small cell access at millimeter-wave frequencies. in 2014 ieee 15th international workshop on signal processing advances in wireless communications (spawc), pages 80–84, june 2014. [34] f. boccardi and r. w. heath and a. lozano and t. l. marzetta and p. popovski. five disruptive technology directions for 5g. ieee communications magazine, 52(2):74–80, february 2014. [35] t. l. marzetta. noncooperative cellular wireless with unlimited numbers of base station antennas. ieee transactions on wireless communications, 9(11):3590–3600, november 2010. [36] fabio fernandes and alexei e. ashikhmin and thomas l. marzetta. inter-cell interference in noncooperative tdd large scale antenna systems. ieee journal on selected areas in communications, 31:192–201, 2013. [37] memon, sajjad ali and chen, zhe and yin, fuliang. pilot decontamination in multi-cell massive mimo systems. in proceedings of the 2nd international conference on communication and information processing, iccip ’16, pages 227– 232, new york, ny, usa, 2016. acm. [38] j. jose and a. ashikhmin and t. l. marzetta and s. vishwanath. pilot contamination and precoding in multi-cell tdd systems. ieee transactions on wireless communications, 10(8):2640–2651, august 2011. [39] m. m. shurman and o. banimelhem and d. a. al-lafi and s. j. al-zaro. pilot contamination mitigation in massive mimobased 5g wireless communication networks. in 2018 9th international conference on information and communication systems (icics), pages 192–197, april 2018. [40] a. n. aljalai and c. feng and v. c. m. leung and r. ward. eliminating pilot contamination using dual pilot sequences in massive mimo. in 2017 ieee 86th vehicular technology conference (vtc-fall), pages 1–6, sept 2017. [41] m. d. silvius and r. rangnekar and a. b. mackenzie and c. w. bostian. the smart radio channel change protocol a primary user avoidance technique for dynamic spectrum sharing cognitive radios to facilitate co-existence in wireless communication networks. in 2009 4th international conference on cognitive radio oriented wireless networks and communications, pages 1–6, june 2009. [42] r. duan and m. elmusrati and r. jantti and r. virrankoski. capacity for spectrum sharing cognitive radios with mrc diversity at the secondary receiver under asymmetric fading. in 2010 ieee global telecommunications conference globecom 2010, pages 1–5, dec 2010. [43] y. zeng and r. zhang. millimeter wave mimo with lens antenna array: a new path division multiplexing paradigm. ieee transactions on communications, 64(4):1557–1571, april 2016. [44] m. r. akdeniz and y. liu and m. k. samimi and s. sun and s. rangan and t. s. rappaport and e. erkip. millimeter wave channel modeling and cellular capacity evaluation. ieee journal on selected areas in communications, 32(6):1164– 1179, june 2014. [45] w. roh and j. y. seol and j. park and b. lee and j. lee and y. kim and j. cho and k. cheun and f. aryanfar. millimeterwave beamforming as an enabling technology for 5g cellular communications: theoretical feasibility and prototype results. ieee communications magazine, 52(2):106–113, february 2014. [46] s. kutty and d. sen. beamforming for millimeter wave communications: an inclusive survey. ieee communications surveys tutorials, 18(2):949–973, secondquarter 2016. [47] m. tesanovic and m. nekovee. mmwave-based mobile access for 5g: key challenges and projected standards and regulatory roadmap. in 2015 ieee global communications conference (globecom), pages 1–6, dec 2015. [48] z. pi and f. khan. an introduction to millimeter-wave mobile broadband systems. ieee communications magazine, 49(6):101–107, june 2011. [49] j. zhang and x. ge and q. li and m. guizani and y. zhang. 5g millimeter-wave antenna array: design and challenges. ieee wireless communications, 24(2):106–112, april 2017. [50] j. du and e. onaran and d. chizhik and s. venkatesan and r. a. valenzuela. gbps user rates using mmwave relayed backhaul with high-gain antennas. ieee journal on selected areas in communications, 35(6):1363–1372, june 2017. [51] y. tao and l. liu and s. liu and z. zhang. a survey: several technologies of non-orthogonal transmission for 5g. china communications, 12(10):1–15, oct 2015. [52] y. saito and y. kishiyama and a. benjebbour and t. nakamura and a. li and k. higuchi. non-orthogonal multiple access (noma) for cellular future radio access. in 2013 ieee 77th vehicular technology conference (vtc spring), pages 1–5, june 2013. [53] h. nikopour and e. yi and a. bayesteh and k. au and m. hawryluck and h. baligh and j. ma. scma for downlink multiple access of 5g wireless networks. in 2014 ieee global communications conference, pages 3940–3945, dec 2014. [54] e. m. eid and m. m. fouda and a. s. t. eldien and m. m. tantawy. performance analysis of musa with different spreading codes using ordered sic methods. in 2017 12th international conference on computer engineering and systems (icces), pages 101–106, dec 2017. [55] x. dai and s. chen and s. sun and s. kang and y. wang and z. shen and j. xu. successive interference cancelation amenable multiple access (sama) for future wireless communications. in 2014 ieee international conference on communication systems, pages 222–226, nov 2014. [56] s. k. sharma and t. e. bogale and l. b. le and s. chatzinotas and x. wang and b. ottersten. dynamic spectrum sharing in 5g wireless networks with full-duplex technology: recent advances and research challenges. ieee communications surveys tutorials, 20(1):674–707, firstquarter 2018. [57] z. zhang and k. long and a. v. vasilakos and l. hanzo. full-duplex wireless communications: challenges, solutions, and future research directions. proceedings of the ieee, 104(7):1369–1409, july 2016. [58] s. hong and j. brand and j. i. choi and m. jain and j. mehlman and s. katti and p. levis. applications of selfinterference cancellation in 5g and beyond. ieee communications magazine, 52(2):114–121, february 2014. [59] c. yang and j. li and m. guizani and a. anpalagan and m. elkashlan. advanced spectrum sharing in 5g cognitive m. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt heterogeneous networks. ieee wireless communications, 23(2):94–101, april 2016. [60] ngmn. description of network slicing concept. requirements and architecture, work stream end-to-end architecture, ngmn, 2016. [61] i. da silva and g. mildh and a. kaloxylos and p. spapis and e. buracchini and a. trogolo and g. zimmermann and n. bayer. impact of network slicing on 5g radio access networks. in 2016 european conference on networks and communications (eucnc), pages 153–157, june 2016. [62] e. j. kitindi and s. fu and y. jia and a. kabir and y. wang. wireless network virtualization with sdn and c-ran for 5g networks: requirements, opportunities, and challenges. ieee access, 5:19099–19115, 2017. [63] t. cunha, a. rodrigues, p. vieira, a. martins, n. silva and l. varela. energy savings in 3g using dynamic spectrum access and base station sleep modes. in radio science conference (ursi at-rasc), 2015 1st ursi atlantic, may 2015. [64] a. checko and h. l. christiansen and y. yan and l. scolari and g. kardaras and m. s. berger and l. dittmann. cloud ran for mobile networks a technology overview. ieee communications surveys tutorials, 17(1):405–426, firstquarter 2015. [65] i. a. alimi and a. l. teixeira and p. p. monteiro. toward an efficient c-ran optical fronthaul for the future networks: a tutorial on technologies, requirements, challenges, and solutions. ieee communications surveys tutorials, 20(1):708– 769, firstquarter 2018. [66] j. liu and s. xu and s. zhou and z. niu. redesigning fronthaul for next-generation networks: beyond baseband samples and point-to-point links. ieee wireless communications, 22(5):90–97, october 2015. [67] y. mao and c. you and j. zhang and k. huang and k. b. letaief. a survey on mobile edge computing: the communication perspective. ieee communications surveys tutorials, 19(4):2322–2358, fourthquarter 2017. [68] n. abbas and y. zhang and a. taherkordi and t. skeie. mobile edge computing: a survey. ieee internet of things journal, 5(1):450–465, feb 2018. [69] t. taleb and k. samdanis and b. mada and h. flinck and s. dutta and d. sabella. on multi-access edge computing: a survey of the emerging 5g network edge cloud architecture and orchestration. ieee communications surveys tutorials, 19(3):1657–1681, thirdquarter 2017. [70] tianchu zhao and sheng zhou and xueying guo and yun zhao and zhisheng niu. a cooperative scheduling scheme of local cloud and internet cloud for delay-aware mobile cloud computing. corr, abs/1511.08540, 2015. [71] y. liu and m. j. lee and y. zheng. adaptive multi-resource allocation for cloudlet-based mobile cloud computing system. ieee transactions on mobile computing, 15(10):2398– 2410, oct 2016. [72] etsi. cloud ran and mec: a perfect pairing. technical report, etsi, 2018. [73] v. g. nguyen and a. brunstrom and k. j. grinnemo and j. taheri. sdn/nfv-based mobile packet core network architectures: a survey. ieee communications surveys tutorials, 19(3):1567–1602, thirdquarter 2017. [74] i. santos, p. vieira, r. borralho, m.p. queluz, a. rodrigues. emulating a software defined lte radio access network towards 5g. in 12th international conference on communications (comm), june 2018. [75] mckeown, nick and anderson, tom and balakrishnan, hari and parulkar, guru and peterson, larry and rexford, jennifer and shenker, scott and turner, jonathan. openflow: enabling innovation in campus networks. sigcomm comput. commun. rev., 38(2):69–74, march 2008. [76] onf. software-defined networking: the new norm for networks. white paper, onf, 2012. [77] a. doria, j. hadi salim, r. haas, w. wang, l. dong, and r. gopal. forwarding and control element separation (forces) protocol specification, 2010. [78] j. medved and r. varga and a. tkacik and k. gray. opendaylight: towards a model-driven sdn controller architecture. in proceeding of ieee international symposium on a world of wireless, mobile and multimedia networks 2014, pages 1–6, june 2014. [79] on.lab. driving sdn adoption in service provider networks. white paper, on.lab, 2014. [80] etsi. network functions virtualization: an introduction, benefits, enablers, challenges and call for action. white paper, etsi, 2012. [81] n. siasi and n. i. sulieman and r. d. gitlin. ultra-reliable nfv-based 5g networks using diversity and network coding. in 2018 ieee 19th wireless and microwave technology conference (wamicon), pages 1–4, april 2018. [82] qiang duan and nirwan ansari and mehmet toy. softwaredefined network virtualization: an architectural framework for integrating sdn and nfv for service provisioning in future networks. ieee network, 30:10–16, 2016. [83] m. weiser. the computer for the 21st century. scientific american special issue on communications, computers, and networks, september 1991. [84] m. al-kuwari and a. ramadan and y. ismael and l. alsughair and a. gastli and m. benammar. smart-home automation using iot-based sensing and monitoring platform. in 2018 ieee 12th international conference on compatibility, power electronics and power engineering (cpe-powereng 2018), pages 1–6, april 2018. [85] j. jin and j. gubbi and s. marusic and m. palaniswami. an information framework for creating a smart city through internet of things. ieee internet of things journal, 1(2):112– 121, april 2014. [86] g. a. akpakwu and b. j. silva and g. p. hancke and a. m. abu-mahfouz. a survey on 5g networks for the internet of things: communication technologies and challenges. ieee access, 6:3619–3647, 2018. [87] m. tavares and d. samardzija and h. viswanathan and h. huang and c. kahn. a 5g lightweight connectionless protocol for massive cellular internet of things. in 2017 ieee wireless communications and networking conference workshops (wcncw), pages 1–6, march 2017. [88] m. pennacchioni and m. g. di benedette and t. pecorella and c. carlini and p. obino. nb-iot system deployment for smart metering: evaluation of coverage and capacity performances. in 2017 aeit international annual conference, pages 1–6, sept 2017. [89] m. shackleton d. bastos and f. el-moussa. internet of things: a survey of technologies and security risks in smart home and city environments. iet conference proceedings, pages –(1), january 2018. [90] i. yaqoob and e. ahmed and i. a. t. hashem and a. i. a. ahmed and a. gani and m. imran and m. guizani. internet of things architecture: recent advances, taxonomy, requirements, and open challenges. ieee wireless communications, 24(3):10–16, june 2017. [91] s. cirani and l. davoli and g. ferrari and r. lone and p. medagliani and m. picone and l. veltri. a scalable and selfconfiguring architecture for service discovery in the internet of things. ieee internet of things journal, 1(5):508–521, oct 2014. [92] z. qin and g. denker and c. giannelli and p. bellavista and n. venkatasubramanian. a software defined networking architecture for the internet-of-things. in 2014 ieee network operations and management symposium (noms), pages 1–9, may 2014. [93] j. zhou and t. leppnen and e. harjula and m. ylianttila and t. ojala and c. yu and h. jin. cloudthings: a common m. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt architecture for integrating the internet of things with cloud computing. in proceedings of the 2013 ieee 17th international conference on computer supported cooperative work in design (cscwd), pages 651–657, june 2013. [94] lorenzo vangelista and andrea zanella and michele zorzi. long-range iot technologies : the dawn of lora. in future access enablers for ubiquitous and intelligent infrastructures, pages 51–59, 2015. [95] sigfox. sigfox. https://www.sigfox.com/en. accessed: 2018-06-15. [96] p. vieira, a. martins and t. cunha. introducing redundancy in the radio planning of lpwa networks for internet of things. in proceedings of the 13th international joint conference on e-business and telecommunications, icete 2016, pages 137– 144, portugal, 2016. scitepress science and technology publications, lda. [97] dash7-alliance. dash7 alliance mode specification, dash7 alliance std. http://www.dash7-alliance.org/ dash7-alliance-protocol-specificationv1-/ 1-ready-for-download/. accessed: 2018-06-15. [98] weightless. weightless. http://www.weightless. org/. accessed: 2018-06-15. [99] ieee. ieee bluetooth 802.15.1. https://standards. ieee.org/findstds/standard/802.15.1-2002. html. accessed: 2018-06-15. [100] ieee. ieee 802.11.ah. https://standards.ieee. org/findstds/standard/802.11ah-2016.html. accessed: 2018-06-15. [101] s. dimatteo and p. hui and b. han and v. o. k. li. cellular traffic offloading through wifi networks. in 2011 ieee eighth international conference on mobile ad-hoc and sensor systems, pages 192–201, oct 2011. [102] k. zaheer and m. othman and m. h. rehmani and t. perumal. a survey of decision-theoretic models for cognitive internet of things (ciot). ieee access, 6:22489–22512, 2018. m. sousa et al. | i-etc iot-2018 issue, vol. 4, n. 1 (2018) id-1 i-etc: isel academic journal of electronics, telecommunications and computers http://journals.isel.pt itec issue 13.pdf http://editorial.isel.pt/journals function and meaning: the double aspects of technology ������� ��� ���� � ��������������� ������������� ������������������������ abstract���������������� ��������������� ��� � ������������������� ���������� ��������� �������������������������������������"��#��$� % ���&� �������"�� ���'�(���������� ������ ���������������� ����������������������� ���������������������� ����������������������� )�������'��������������������� ������������� ������������'������ ���������� ������������ ���������� ����� ������������ �� �������� ������� �� ����� �������� ������ ����� ��� �� ����� ��� ���������������������������������������������� ����*��������+�������������������� ����'� ���� �� ������� ��� ���� ������ ������ ����� *��������� ������� ������ ��� ���� ����������� ���� ��������������� �������������������� ���������������������������� keywords��,��������������� ��������� �� ������"��������"��#��$� % ���*���������� "�� ��������������)������� 1. introduction: technology in the social world -����� ���� ������ ���� ���� �������� ��� �.������� ����� ��� ���������� ������'� ������ �������� �#������ ���� ��� ������ ��� ���� ���� �� ����� ��� �� �� ���� ���� ��� ���� ��'� ����� ���� ����������� �#���������� ����� ���� �� �� ���� ����� ����������� ���� ���� �� ��� ������ ������� ��� �����������������������������'�)������ �������������������������������������'������� ��������� ��������� ������������� ���������������� �������������� ����� ���������'������������� ������ �������������������������������������������������� �������#����������������������������������������� ���������������������������� ��� ��������� �����'��������#����������������������� �����#���'� ������ ������������� ������������������������������������������������������������������������� ���������'�/��� ����� ������������������������0�����'�&���������������������������������������� ���������������������'�(������ �����������1�������2����1�����0 �� �2������ �������' (������������������������ ���� ����� ��������������� ������� ���������������������� ���� ���� ���������0������������������������������������� �����������'�������������.�������� ������ ���������������������������������������������(�������� ���������������������.��������������� ���������������� ������� �������������������������� ����������������� ��������������� ����'� ������ ��� �� ����� ������� ���� ����� �������� ��� ���� ��������� ������� ��� ����� ���� �������� ��� ���������������� �� ���������� ���� ������� �������������������������������������.����� ���� ���������'� (�� ������� ���� .�������� ��� ���� ���� ��������� ��������� ��� ��������� ��� ������� �� ������� �������� ��� ���� ���������� ��� �������� ��� ���� ������ (� ����� ������ ���� ����'� (� ����� ���������� ��������������� �������������������������� ������������ ���'� special inauguration issue vol.1, nº1 id-2 (2012) afantoni stamp andrew feenberg | function and meaning: the double aspects of technology i-etc, volume 1, nº 1, special inauguration issue, (2012) paper id-2 2. the technologization of society "������ �� ������� ����������� ����������� ���� ���� ������ ��� ������ ��� �� ��������� ����� ��� ������� ���� �� ������ ����� ������������ ���� ��������� ��� � �������'� ����� ���� ��� ���� ����� ��� ������ ���� ����� �� ��� ��� ������ ��� ����������� ���� �������� ������ ���� ��� ������� ��������������� ��������� ����� �������'����������������������������������� ���� ����������� ������������������������������������ ��� ������ �����������������0 ����� ���� ������� ������ ���������� �����������������' �����������������������3����� ��������������� ������������������� �����3���������������� ��� �3������������� ������������������������ ������������ ��'�������������������������������� ��� ���������� ��� ���� 45��� ������'� ���� ����.��� ��� ����� ����������� ���6� �� ��� ���� ���� ��� �����������&� ����$� % ������*����������"�� �����*� ���������� �� ������������ ����� ���� ������������ �������������� ���� �������������� ��� �������� ������������������������� ��� ����3���� ������� ������ ������������ ��������'� (��������������3������ ����������������������� �������������� �������������� ������� ��������������� ����� �� ����� �����#� ��������� ��� ���� ��������� ������'� )��� ������ ��� �������������������������������������� �������������� �������� �����(������������������� � 6� ����'�7� �������������� ����������������� ���������������������������������������� � ���������'������ ����3������� ��������������������������������� �'������������������������� ������������������� ����3�������� �'��������������� ����������� ������������������� ����������� ������������������'�)�������������� �����������������������0��������������������1�������'2� ���� ����������� �� ��������(������������������� ������������������������� ���������� ������������������������� ���� ������������ ���������������� ���������������� ���� ������ ���������� ������� 6� ����������� ���� ����#�'������ ���� ����������� �������������������� ��� �������������������������'����� �������� �������������������� ������������������������������� ������� ��������������������������������������������������������� ��������������������' ��������� � ��� �. ������8��������&� ����������������������������� �������� ���������������� ���������� 1���������������� ��� �������� �������'2�&� ��� �������� ���� ���� ����� �������������� ����������� ��� ��� ������� ���� ��� ��� � 6� ��� ����� ������ ������� �����0 �� �� ��� ������ �� ���� ����#�'� "����������������������������������� �� ���������������� �� ����� �������� ���� �� ��� �����'� ���� ���������� ������� ��� 1����� ��2� ����� ���� ��������� ��� �� ������� ��� �������� ������� �� ������������������� ���� �����������������������������������������������3��� �������������������������� ��������' ���� ����������� ��� ����� ��� ����� &� ������ ����� ����������� ����������� ��� "��#� ������� ������������������� ������������.������������������������� ������������ ���������� � ����� ��# ������������������������������������������ �'������������ ����#���������������� ��� �� ��������� ������ ��� ������� ���� ���� ��� ����� ��� ��� �� ����� ������� ��� ����� ��� ��� ��� ���� ����� ��������� ����� �� ����� ������������� �� ��������# �����������'�"��#������ �����������������������1��������������2������������������� ������������������������������ isel academic journal of electronics, telecommunications and computers http://editorial.isel.pt/journals ������������� �0 � ��������� ��������� ����� �������� ����'�9����������������������� ����� ���� ���� ��6���������������� ��'� ������������ �������� ����. ���������&� ����$� % ����������3�������"��#���� ����.��������� ��������������� ��� ����� ���� ������ ����� �� ��� ������� ���������� �� ������ �� ��������� ���������������� ���� ����������������'���������������������������������0 ��������� �������������3���"��#����� ��� ���� ����������� ���� ���� ������ ��� ���� ��� ����� � ������ ��8��� ���� ����� ���� ���� ��� ������ ��������������� ���� ��������������'�$� % ��������������0��������������������������� � ������� ��� ��� �� ������ ���������+�� ����������� ������������ ������ �������� ��������� �������� ���� ������ ���������� ��� �� ���� ����'����� ���������������� ��� ������������ ����� ������ ���������������������������������������� 6� ���������� ������������������� ���������'� $� % ����������������������� �������� �����0 � �������������������������������� ��� "��#� ���� �3��'�������� ��������1�� ����������2������������������������������������ ����� ��� �������������������� �����'�$� ��������� ��� ���������� �������� ������������������� ����� ����� ����� ������������������������������������������������ ����������'�$� % �� �� ��� ����� &� ��+�� ��������� ��� ����� ���� � ���� ������ �������� ��� �� �������� �� �������� ���������� � ��������� ��� ��� '� *�� ������� ����� ���������� ��������3��� �� ����� ������� ���� ��������� ���� ������ ���� ������� ������������ ��� �������� �������'� *��������������������������� ����.��������������������� ������������������ �������� �� ����������� ��������������� ��������������� ������� �� �'���� ������:;����������������� ��� �������� ��������������������������� ����������������� ��������������� ��������� ����������0��� ��������� ��������� ��� ��������������������� ��� �� ���������� �����'����$� % ���������<:=� �1&������������������������� ����3�� ��������������������������������>�������������� � 6� �������� ����� ������?���������� ��������������� 6� ������������������� ���� ����������� ��� ���� �� ������� �������� ��� �������� � ��� �� ���� ��� ���� � ���� �� �� ������� ��� ���� ����� ��� ������'����������������� 6� ����������@� ����+��� ���������������� �������������������������� ���������� ���������������aa����0 ������� ���� �aa��� ����������������������������#����������'2a ������ ����� ���� ������� ������������ � ���� ���� ����� ���� �������������� ��� ���� ���������� 6� ��'��������������������� ����������� ������� ����� 6� �������������� ������� ��� �� ���������� �� ��� ����'� ������ ��������� ������������� ��������� �.������ ���������� �� ��� ��� ������� ��� ���� ���� ��� ��� ������� ���� �� ����'� )��� �������� ��� ����� ���� ������� ������������������� ��� ������ ��������������� ������� �� �������������������������� ��� ���������������� ��'�7������� ������������������������������� ������������ ������ �������� � ��������������������0��������'�$� % ����������� �������������������#�������������� �� ������������� �� �������0 ���������������������� ���'� _______________ a the passivity of the experimenter to which lukács refers is only apparent: the experimenter actively constructs the observed object but, at least in lukács’s view, is not aware of having done so and interprets the experiment as the voice of nature. while lukács does not criticize the epistemological ������������ � �� � ���� �� � ������� �� ����� � ��� ��� �� ����� � � ���� �� �� �� ���� ��� andrew feenberg | function and meaning: the double aspects of technology i-etc, volume 1, nº 1, special inauguration issue, (2012) paper id-2 ������������� � ���������'� *��������� ������� ����� ���� ������� ������ ��� �� ���� ��� ������ ���� ���� ���������� ����������� ���������'�9���������������������������������������� ���������������������� �������������� ������� ������ �������������������������#����������������������������������� �������������� ��� ���������'�/ 6� ������������������������� ����#����������� ������������ ������� ����������'� ������ �� ����#�����3������� ���� ���� ������ ���� ����������� ���a������ ���� �������'� (�� ����� ����� �� ������� �� �������� �������� ����� ����� ��� � ��� ��� �� �� ���a �#������������������������ �� ������������������������������������������������3��� ������ ������������������������� �� �������������������������������� '�(����������������� ������� ���������������������������������'�(���������������������� ����.�������������$� % ����������� ���*���������� ����������������������� ��� �������'����������� �� ����������� ������ �����0 a�� ��� ����������������������������� ���������������� b �� ��������������������������������� ���������������������������'�'�������������������������� �������������������� �� ������������������������b �� �������������� ������� ��������������������������������������������b �� ���� ����.������������������������������ �����������������������0 �������� ���������� �����b �� ���� ���������� ���� ����������� ����� ��� ��� ����� ����������� ��� ���� ��������� ������� ��� �� ��� ���������������' ����� �������� ������ ��!��� "��#��$�%�� ��"��#�� ����� ��� ���� � ������� ��� "�� ���+�� ����.��� ��� ���a������������ �� ����'� ����������������������� ��$� % ������*���������������������������������.���������������� ��� "�� ���+�� �������������� �� ���� ��� ���������� ��� � ������'� ���� ������� ������� ��� ����� �����������������������#������������������������ ���������������������� ��� �������� �������� ����������� ����������������������������������������������������������'������������� �������� "�� ���+�� ����.�������������� �����1�� ������� ���������������2���������������������������� ������� ���� � 6� ��� ��� ������� ��� ������� ������ �������� ������������� ���� ����� �# ���� ���� ��������� �����������#�������'�� (�� ������� ��#� ��� one-dimensional man�� "�� ���� �������� 1����� ��������������� ��������������� �����0 ����6� ��>�����������������?��prior ������������ ���������������3������� �������� �0 ��� ���������6� ���������������������������� ���������������������������� �����0 � �����������''''(�� ��� ��� ������ ���� �������� ���� ���� ��� �� �������� � 6� ������� ��� �� ��� �0 � ������� ����� 6� �aa������������ ��� �������������������������������� ����''''2'<4=�"�� ���+�� ������ �������#�����0������������������������� �������������������������������������������� ��� ������� � �����0 a�� ��� ��� ������������ ��� ��� ��� ��� ����������'� 9���������� ��� 6���� ���� ���������������������������� �����.��������������� ���������������� 6� ������������������������ ���� ��������� ��������������'� /����� ������ ��� � ����� ���� ������ ����� ������� � ����� ������ ���������� ������������� �������������� ������������������������� ������ ������������������� ��� ���� ��������������� ��� ������ � 6� ���� ��� ���� ������ ���� ������ �� ��� ����� ������� ���� ��� ���� ����������������� �����0 ��������������'� isel academic journal of electronics, telecommunications and computers http://editorial.isel.pt/journals &��� ��� ����������� ��� �0 ����� ��� ��� ��� ���� ���6� �� ��� ����������� ��� ���������c� �� ������� �������� ����� ��� ��������������������������������������������� ������������������ �������� ��������������0�'���������������������� ����������������� ������������������������� ������������������ �� ������������������������������ ������������������ �'������ �������� ��������� ���� ���������� ����������� ��� ����� ����� �� � �� ������� ��� ���������3������� ���� �#������� ����������������#���������������� ������������������������������������������������� ������3��������������� �����������'� � �����0 a�� ������� ��� ��������� ��� �������� ��� ���� �������� ��� ������ �� ���� ���� ������� ������.���������������� ����������� 6� �������������� ������� ��� �� �� �������� ���������������'�/����� ������������������������������� ������������������������������ ����������������������������� ��������������������������� ������� ������� �� ������������ ������� �������� ��������������'�������������������������� ����������������� ������������� �������� ��� �������� ����������������������������������������������������������' "�� ���+������� ����������������0 �����d���������������������������������$� % �� ����*�������������������������������������������������������������������������������������� ���������������� ������������������� ��������������������������'�"�� ���� ������������� ��� �� ������� ������������������� ����� ��� ���� ����� � � �� � �� ���� ��� ����������� � ��� �� ��� �� ��� artistic imagination direct the construction of a sensuous environment, only if the work world loses its alienating features and becomes a world of human relationships, only if productivity becomes creativity, are the roots of domination dried up in the individuals. no return to precapitalist, pre-industrial artisanship, but on the contrary, perfection of the new mutilated and distorted science and technology in the formation of the object ����� � ���������� � �� ���� ���� � �������� ��� �������� ���� ������ �� ������� ��� condition—not of an ��������+��� isolated from real existence…but that harmony between man and his world which would shape the form of society.”<e= �� ������� ����� ����������� ��������� ���� ��.������ ���� ���� ������ ����� ��� �� �� �������� ��������������� ��������������������������������������� ������� ������ ����������� ������������������� �����������������'�)���"�� ������������������������������������������� �� ����������������������������������������������������'���1.���������������� �2�����# ������ �������������'�(���������������������� ����������������������������� ��� �������������� �3������ ����� ������������������� ������������������������������� �������.�����0� �������������'� �������������������������������������������� ��������1�������� �lebenswelt,” �� ���� ���� ����� ��� �#������ �� �� ���������� �������� � �������'� �������� �3��� ��� ������� ������ �� �� �� ��� ������� ���� ��� ��� � 6� ��� ��� ���� ������� �������� � ��� ������ ��������� ��� ����� ��� �� �����'� (������� ��� �� ������� ������ ��� �������������� ��� � 6� ��� ����� ��� ������� � �����0 a�� ��� ������������������������������ ������������������ ������� ��������������� ����� � 6� ������������������������������ ����������������������������������� 6� ����������� ��������� ������������ ������������������������������������� ���������0�����'� &����� ����� ������ ���� ������� ������ ��������� ��� ���� ��� "�� ���+�� ����������� ����.������ �������� ���������������'�*����������� ������������������������������ ������� andrew feenberg | function and meaning: the double aspects of technology i-etc, volume 1, nº 1, special inauguration issue, (2012) paper id-2 ��������� ��� ���������� �� ����� ���� ���������� ��� ���� ��� ���'� ���� ����������� ��� ���������� ���� ������� ���������� �������� ����� �� ���� ����� ������ ��� ���� ��� ������ ����� ��� ���� ����������������� ���� ��������� ��� ���� ����.��� ��������� ��� ���� ������� ��� � ��������� ������ ������������������'�� 3. meaning and rationality in recent philosophy of technology ���� ����������� ��� �� �������� ���� �� ��������� ���� ����� ����� .�������� ����� ���� ������"�� ����������������� ����������� ������������������ ����#������� �����3������ ����������������������������������������� ������� ������������� ��������������'� �� ���� )�������� ���� $����3�� ��������� ���� ������������� ��8��� ��� �� ���� ������ *���������� ����� ���������� ����� .�������� ��� ������ ��������'<f�g=� (�� ���� ������ ��� � *��������� ������ ���� �� 1����� ��������2� ��� �� �������'������ ������������ ��� �� �� ������� ��� �����3������������������.�����"�� ���+���������������� �������� ������������ ��������� ������������������ �������������� ������'�,������ ���������� ����������� ���������������� ��������������������������������� ������� ������������� �������6������� ������� �������������� ������������ ���������0 ��� ������������������ ������������'� ������������)������������������ ����� ���������*��������+���������'������ ��������������������������������� �� ���������������������������������������������� ��� ��������'��������������������� ��������� �������������������6� ���0����������� ������������� �������� ���� ���������� �������� ���������� ���� �� ����� ���� ��� ���� ��� ��������� ��� �� ���� ������������ �� ����'� )���� �������� ���� )�������� ��6� �� ���� ����������� � �������� ��� �� ����������������������������� ��������������������������������������������������������� �������� ��� ������'� ����� ����� ��6� �� ���� ������� ����� ������ ���� ������������ �������� �������� ��������������������� ������'�����������������$� % ���*�������������"�� ���� that there is����������0����� ����������������������������������� ������� ���������� ��� �� ������'������ ������������������ ����������������������������� ������������� �� ���������� �������� ����������������� �������������� �����#������ ������� ����'������������� )��������������������������� ������������������������������������������������ ��������� ��������������������� ������� ���0#������������������ ������'� �������+�� ���� )�������+�� ����.���� ���� ����� ��� ���� ������ ����� ������� ��� ����� ���� �������'� ����� ������ �������� ���� ��������������� ��� ���� ������� ��� ���� �#�������������������'�������������������� ����������� ���������������� 6� ��� ����� �� ����� ���� ����������� ����� ������������ ��� ������'� (������� ����� ����� �#����� �� ��� ������� �������������� ��� ����������� ���� ������ ������ ���������'������ ��� ���� �� ������ ��� ��#� ��� �������a���������������������'��� ������������������� ������ ���������� ��������� ��������� ����������������' ����������� ���������������������������������������������� �� ����� �����'����� ���������������� ��� �� ����� ���� ��� ���� ��� ����� ��� �� �������������� ����� ���� �� ����� ������������� ��� ������ ��� ��� �� ���� ����'� /��a��������������� �������� ����� ���� �������� ��� isel academic journal of electronics, telecommunications and computers http://editorial.isel.pt/journals ������3������� ����������������������������������������������������������# ������ ������ ���������������� ��� ��� ����'����� ���� ��������� ��� ����� ������3������ ��� ���� �� �������� ��� ��� ��� ������ ��� .�������'� (�� "�� ���+�� ����� ��� ��� ���� ������ ��� �������� ���� )��������� ��� ��� ���0 ������� ���� ����� ��� ��������� ���������� ��� ���� ����� ���������� 1���������2� ��� �� ���� ������ ���� ��������� ��� ��������������� ����� ��� ��� ������� �������������� ���� ����a ��������������������"�� �����������������3����� ������������' �����&�'(����� � ����� ���) "��'� � $����3�� �������� ���������� ����� ��� ���� ��� ���� ���� ����� ��� 1�������2� ��� 1�����'2� h������ ��� ��� ��0���� ����� ���� ������� ����� ��� �����'� ����� ��� �� � ���� ���� ��������� �����#��� ������������������� ������������������ �������������������������� ������������ ��� ����� ����#�'� �� �� � ���� ������ ����� ������ ������ ��� ����� ����� ���� �� �������������������������������� ����������������������������������������'� "������������ ������������������� ���� ������ ��������#������ ��������������� �������'����������������0������ ��������������������������������������� ������� ��������� �������� ����� ��������� ���������� �� ����� ��������� ��������������������������'� "�������� ���� ��� ���� ��� ���� ��� �������� ���� ��� �� ��'� ����� ���� ���� ������ ��� ������� �����1 �������'2� �� ��������������������� �������������������������������� ������ ��� ��������������������������������������0 �� ��������������'������� ����������������������� ����� ���������������������������������������� �������� ������'�������������������������� �������������� ����� ��� �������0 ����������������� ������������������������������'����� �� ������������������������������������������������ �����#��������������� ���������������� �����������������������������������������������������������������������#������ ���������� �����������������������#�������0 ��� �'��� ����������������������� ���� ����.��� ������������� ������������������������'� ���� ����������������� ����������������������'������ ��� ��6����0���������������������������� �� ��������� ������� ���������������������������������������������������������.�������'�h���������� ����������������������� ��������������6����0��� ������������������������������������������� �������� ���������'�)����� ������������������� �������������������������a������ ����������� ���� ���������������������������������������������������������������������'� ������������������������������������� ��������������������� ������������� ���� � ������'�)�������1���������i��2�������������������������� ������� ��������������������� ������ �����������������������'���������������� “what happens in such a transformation of meaning into value? as meaning becomes thematized as value, the manifold connections which operate in part ‘behind our backs’ and which, through informing and shaping our experience, predispose us to experience in a characteristic way, are transformed into premises. the validity of these value-premises, apart from the referential anchoring in the meaning which gave rise to the value, stands or falls with the rational evaluation of those premises. our inability to provide ������� �������� foundations for such premises, in abstraction from the meaning that gives them point, results in our inability to experience as binding in a nonarbitrary andrew feenberg | function and meaning: the double aspects of technology i-etc, volume 1, nº 1, special inauguration issue, (2012) paper id-2 way. that is, such values .�� values, that is, in isolation from meaningful practices, cannot claim us.”<j= (�� ���� ��������� ��� ���� ����� ������ ��� ����������� ������� ����� ����� ��� ���� �������� � ������������������������������ ��������� �����������������������0 �� ������������������ ����� �� ����������������������'������������������������������ ���������������� �b������ ��� �� �������3��� ���� ���� �3��'� )��� ���� �#�� ���� ��� ���� ��� ���������� �� ��� �� ������� ������� ���� ���� ��� ������ ���� �� ���� ������ ���� ���� ���� �����������'� ����� ���� ����� ���� �������� ����� �������� �������� ��� ��������� ��� ��� ���� �� ����� ��0���� ���� ������������ ��������� ��� �������� ��� ���� ����� ����� ���������� ��� ���� ����� �������� ����������'� (�� ����� �#������ �������������� ��������� ��������� �����������������3��'� ����������������������������������������������������������������� ����������������� ����� ������� ���� �� ����������� ����� ������ � ������� ��� ��������� ������ ��� �� ���������� ��� ����'� �� ����������� ���������� ������ ����� ������ ����#��� ���� 6���� ��� � ���� �� ��� ������ ����0 �������������������������������� ����������� ������������������'���������������������+�� � ����� ��� ��������� �� ��� �� ����� ���� ����������� ������� �� �������� ���� �������'� ���������������������������������� ������� ����������� ������� ������������������ ����� �� ����������� ����������������� �������� ����.��� �������������������������������������������� �����������'<;=����������� ����3�������������������� ����������������������������������� ���������� �� ���������� ���� ���� ��� ����� ������ ��� ������� ����� ���� ����������� ��� �� �������� ��� ���� �3��� ������������1�������2� ���� ������� ������ �������������������<k='������������������� ������������������������������������� �������������������� ������ �������� ������������������'� "����������������� ������������������������ �� �������������������������������� ����������� �� ��� ��������������� ��������������������������� ����#�����������3�����' �����*� �' ����% ��+�,���� ���''� )�������+���������������������'�*�� ����������� ����������������������������� ��������������������������3������������������ �����1�� ���������2���������������� ������ *�����������������'� ���*������������� �� ������������������������������������������#������� ������� ���������������������������������� 6� ������������� ������������������������������� ����������3��'��������1�����2�� ����������*�������������������������������������������� �������� �������� ������������������� �����������������������������������������'�(������� ����������������������� ���������� ��������������������������� ��������' )�������� �������� ����� ��� ����� � ���� ��� � ������� ����� ���� � .��������� ��� �������������������������������� ����������������������*�����������������'��� �������� ��� ���� ���� ������ �������� �� ������� ������ ���� ����� ������ ��������� �� �� ����� ����� ���� �����'� *������ �� ��� ��� ������� ����� ��� ���� �����#� ������������� ��� ������������ ������� ������������������������������������������'�1l��� ��m������������� ������������ ��������� ���� ���������������� ������� ������������������'�(�������� ���������������������� ������������������ ��� ������ �������� ��������� ���������� �� ��������������� ������������ ��� �� ���� ����� ������ ��� �� �� ���� ��� ���� ��6����� �������� ���� �� �� ��� �� ��� ��� ���� ������������������ ����#�2'<n= isel academic journal of electronics, telecommunications and computers http://editorial.isel.pt/journals ��������� �� ����� ��� ����� ����� ��� �� �� �� �������� ���0 ������� ����� ��� ��� ��������� ����� �'�)����������������� ������������������� ���'�(������������������������ ��� �#������ � ����'� ���� ������ ������ ������ ��� �� ������� ��� ������ ��� ��� �0 � ������ ��� ������� ������ ������ ��� ���� �� ������ ��� ������� �������� ������� 6���� ������ ������� �����'� � ������ ���� �������� �� �#�� ���� �� ������� �������� ��� �� )�������� ����� ���� 1���� �� ��������'2������������������������������������������ ���������������� ������������������3��� ��� �� �����������0 ������� ���� �3���� ���������� ����������������� ���'�1�� ��������2���� ����������1���������������#���������������� ���� ��������������2'<:5=� "������� ������� ����� ����������� ����� �� ��� �������� ������� ����� �#�� ���� ���� ������������������� ����������������'��� ������������� �� ��� ������������ ��������������� ���� 6� ��'��������.��������������������������������� �� ��1����� ��� ��������������������� ���������2'<::=������������������������������������� ����������������������������������'� ����������������� ������������������ ������#������ ���� ����#�������������������������� ����������������0 ��� ����������'�)������������������������������ ������������������ �� 6� ��������������6����0 �������������� �����0 ������'���1��� �� 2���� ������ ����������� ����������������������������������������������������� ��'�(����� �� ���� ��������� ����������� ��������������� ����������� ��������������������������� ������������������������������������� �����������������������������������������������������������'� $� �� ��������� )�������� ������ ��� ��������� ����� �� �������� ����� � ��������� ���������������� �������������������������������������� ������������'�)����������6� ��� ��������������� ����������������������������������� ������������������ ����������� ����� ���������������������� �������'�������� ������������� ���������������� ���� �������������� ���� �������������� �������������������������� �����������0 ����������������������������� ������ � ���� ���� ����� �� ����#�� ��� �������'<:4=�o��� ��� ���� ���� ��� ���� �� �� )�������� ����� ���� �� 1���� �� ���� � �����2� ��� ��� �� �� �������� ��� ���� �� ����� ��#���� ����� ����� ����� ����'���������������������������� ���������������������������� ������������� �������� ������������������'�/� ���������������������������� ������� ���������������� ������� ��� �������������� ������������ ��������������'<:e= 4. agency in the technical sphere "�� ���� ������� ����� ������������ ������ ��� �� �������� ����� �� ��� ���� �� ��� ������������ �� ���� �����'� /�� ���� ������ �� ������� ��� �� ��� ������� ������ ������ ���� ����� �� ���� ��������� ������� �� �������� ��� ����� ���� �� �������'���������� ���� ��8� ������ ��� ����� ������ �� ����� ���� ������ ��� ����� ���� ������ ������ ��� ������ �� ����� ���������� �� �� ����� ��� ���� ��� ������ ���� �������� ����� ��� ������ ����'������ ��� ���� ������ ����� ��� ���� �����������(�������'�9����������������(��������������������� ���������� ��������������� �� �� ������ ����� ���� �������� ��� ������ ����� ������� ��� ����� �������� ����� ��� �������� ���� ���������� ���� �� �������'� ����� �#������ ��� ���������� ���� ���������� ���� ������ ��� ����� �� ��� ������ ����� ������� �� �������� ���� ���� ����������� ����������'� &�� ��� ���� andrew feenberg | function and meaning: the double aspects of technology i-etc, volume 1, nº 1, special inauguration issue, (2012) paper id-2 ����������$� % �+����������� ������������� ��������0 �����������������#������� �������������� �� ����������*���������������������'�"�� ���+������������� ���������������������������� ���������������������� ���' �� �� ������ ���� ���� ��� ���� ��������� ������ ��� ���� (�������+�� ��������� ��������'� (�� ���� ��������� ��� ������ �� ��� ��� ��� ����� ��� ����� �������� ��� ���������� ��������� ���� ��� ��������� ��0 ���� ������������ ������� ���� ����������� ���� ����� ����� ��� ����������� �������'�(������������������������������������������������������������������� ������ �������������� ���������������� ������������'�)��������������������������6������ ��������� ��� ��� �� ������ ������ �������� ��� ���� ������� ���� ����� ������ ������ ������ � ��������������������������������������'�*����������������������������������������������� �������������������������������������������������'�*����� ��� ��������������������0 ������� ������ ������ �������������������� ������ ���������������'� ���� ������ ��� ��� ������� �������� ��� �� �� �������������� ���� �#�������� ��� ��� �� �� �� �� ���������� ��� ���� l������� -.�������� ������������ ��� ���� ������ ����� ��� ��������� ��������'� ��� ����� ����� ������ ������ ������ ��� �������� ������ �� ���� ������� �������������������������������������� ��������������������� ���������������������'����� �� �� ���������� ������ 1&�� ����� ������ ���� ��������� ���� ��� ��������� �����3��� ����� ��� ���������6���� ���� ����������� ������ ��������������������������� �����'2� (� �� ��������� �������� �������� �������� ��� �����3������ ��� ���� ������ �������� ��� �������������� ���������������� �� �����������'� ����#����������� ��� ��"�������������� ���� ���� ��������� ��� ������ ��� ���� ������ :nk5��� ����� ������ ���� (�������� ���� ������� ��� ���� �� �� '� (�� ���� ����������� ��������� ��� �� ��� ������������ ������� ����������� ��0 ���� ����� ���� ����� ��� ��� � ����������� ������ 1"�������'2� )��� ���� ������� ���� ����� .�� ��� �� ��� �� ������ ���� ��������� ��� ����� �� ������ ��� ������ ������ �����'� ����� ������ �������� ���������� ��� ���� ������� ��� �� �#������� ���� ��������� �������� ���� ������� ��� ��� ���������������������������������������������#'�/�� ���������������������������������� ���� ����������������������"����������������� � ��������������� �������������������������� ���� ���������' ����� ��� ����������������������������������� �������������� ������������ ����� ����������������������������������������������� ����������������������������(�������'����� ��������������� ������������������������������������� ������ ���������0 ������������3������ ���� ������� ���������� ����� ��� ���� �� �������� ��� �� ���� ����� ������'� ����� ���������� ����� ��� ����� ��������� �����������'� ��� �� ���� ����� ��� �� ����� �� ����#�� ����� �� ��������'� ����� ����#�� ���������� ���� ������ ��� �� �� ��� �0 � ������ ��� ������������� ���� �#������� ��� ��� � ��������� ������ ������ ��� ������ ������ ������� ������ ��� ������������� ���� ��� ��'� ������� ����� ������ ����#��� �� ��������� ��� ���� �� ������� ������� ���� ��� ������������ ������� �������� ���� ��������� ��.�����'� ����� ��� � ����� ����#��� ���� ��� ���� �� �������� 1�����������2��������� ����������������������� ���������� ���������'� ������������������������ ������������ �����������0��� �������������������������'� )��� ����� ���� ���� ���� � ��� ��� ����� �����'� ����� ���� ���� ������ ��� ������������� ������� ��� ������������������������ �'����������������#������������������������ �����-������ isel academic journal of electronics, telecommunications and computers http://editorial.isel.pt/journals ����/����� '���������������������������� �������������1������2����������������������������� ������������������������1���� �#��2� ������ �������� �0 ������������'������������� ��������� ���������� ���� ��������������������'�)�� ��������������������� ������������������ �������� ������� ���������������� �#�������� ������ �������� �������'�)��������������������������� �������������������������������� ������ �����������������������������'����������� ������� ����������� �� ���������� �����#�������������� �� �������������������� ��������' ������������������������������������������������������������������� ������ ������ ��� �������� ������ �'����� ��������� ������ ��� ������ ������ ������ ����� ���� �������� �������� ���������� �������� ������� ��������������������� ��������������������������� ������������0������������������������������������'�$���� ����������(�������������������� ������ �� �������������� ���� ����� �������������������� ����������������������� ������� �� ������������������������� ������������� �������������� �'� &������������������� ���������������������������� ������ ������������������������� ������� ���������������� �������������������� �� ����'���������� ��������� ������ ������ ������ ��� ���� ������� �� �������� ���� ���� ��� �������� ��� ��� �� ����� ����� �������'� ����� ������� ���� ������ ����� ��� ����� ���� ������� ����� ��������� ��� ������ ������ ��� �� ��'� ����� ������ �������������������������������� ���.��������� ��������������������������'� ��� �� ������� ��� ������ � ���������� �������� ������ �� � ���� ������������� ������� ��� �� �������������������� ��������� ��������������'<:f= ������ �#������� ��� ������ ������ ������ ��� �������� ������ �� ����� ����� ���� ��������� ��� ����.��� ������� �� ����� ��� ��������� ��� �� �������� ��� �� �� ��� ���� ����� ��� ������������������������������ ����������� ������������������������� ������ ����� �#�� ����'�7� ������������������������������� ������ �������������������������������� �������� ���� ������������������������������������� ��������������������1�� �������2������� ���� ��������������� ������ ���� �'� &�� ����� ����� �� �� �� �� ������ ��� ���� �� ��� ���� �������������� ��������� ��� ����� ��� ����� ���� (�������'� -������������� ��������� ����� ���� ��� �����0 ���� ������� ��� �� �������� ���� ���������� ���� ����� ��� �� ������� ��� ������������ ���� ���� ���� ������ �� ��� ���� ���������'�&�� ���������������������������� ����������������� ������������ ���� ����.��� ��'����������������������������� �������������������� ���� ������������� �� �#������ ��� ������� ���� ��� �� ���� ������ ��� �� �������� ��� �� �������� ��� ��� �� ������ ���������������#�� ����������������������������������������������' ������ ���� �#���������� ��� ���� �� ��� ���� �� ��� ��� ������� �������� ��� �#������ �� ����� �� ���������'� ���� ����������� ��� ���� ���� ��� �������� ������ �� ���� ���� �������������� ����������� ��� �� ��� ��� ��������� ����� ���������� ��� ��� �� ����� 0��� ����������'������������������������������������������������������������ �������������������� ����������� ��� �� ������� ��� ������������ ������ ������ ���� �� ����'� (�� *������������ ������ ���� ����� ���� ����� ����� �� ������� ������ 1�����2� ��� ���� ������ ��� ��������������� ������� ���� �� ������ ���� 0���� ������ ���� ���� ��������'� (���������� ��� ����� ���� 6���� ������� ������������������ ���������������������������� �����������������'�"�� ���+�� ������������������������������������ �������������������������� ���'������������������ �� andrew feenberg | function and meaning: the double aspects of technology i-etc, volume 1, nº 1, special inauguration issue, (2012) paper id-2 �������� �������������������������� ����� ���������������������������� ������������������ �������� ����� ��� ������� ����� ���� ������ ���� ���� ���6� ����� ��� ���� ����� ��0������ ����� ��������������������"�� �������� ����' 5. hermeneutics of technology ����� ����� � �� �� ��� ���� ����� ��� ���� �� ��� ������������ �� �������� ���� ���� ������������ ����� ����� .�������� ���� ��������� ������ ��� ���� ��������� ���������� ����'� ��������������� ���������������� ��������������������������������������������������������������� ������b� �� ��������� �������������� �������� ����� �����0 ���� ����.��� ��'� *��� ���� ��� ��� ����������� ����� ����� ������ ��� ���� �� ����� ����� ��� �� ��� ����c� ���� ����� �������� ������������������'� &�� ����������������������"��#���a&� ������������������� �������������������� ���� �������3������� ���� ����� ��� ���� �������� ������� ��� ��� ���� �������� �����0 ������ ��� ������������ ���� ���� �� ����� ���������� ��� ��'� ����� ������ �� �������� ��� ���� � ���� ���� ����������������������������� ���� ����.����#�������� ��������������������������� ������� ��� � ��� �� ���� �� �������� �������'� ����� ������� ���������� ���� �������� ��� �� �������� ��� ��������'� )��� ������ ��������� ��� �� ������ �� �� �� ���������� ���� ��� ��� �� ���� ����� ��� ���� �� ���������� ����������� �������� ����� ���'�����������a ���� ����������������������� ���� �� ������������������� ������������������������������������� ���� ���'�/��� ����������������� ����� ������ �� ������� ������������������������ ��������������� �������� �������' ������������������ ������������������������������������������������������3������ �� ������������� ������������������ ���� �� ��������� ���������� ������� ��� �� ����� ��� ���� �������� �� ���� ���� �'� (�� ����� ����� ���� ���������� ����� ��� �����#� �� ���� ������ ���� ��������� ��� ��� ������� ��������� ������ �a�#���� ����� ������ �� ���� ��� ������ �������� ��� ����������������� ����������������� ������ �������� �����#����������������'������������� ������������*� ������������#�����'�*���������������������������������������������������� ��� �������� �������� �������� ��� ���� �������� ��� �� ������� ������������� �� ���� ���� � ���� ���6� ��������� ������ ������ ����� �������������'�*� ������������� ��8� �������������� ������ ���������������� ���������������� ������������������������3������ �������������� ��� ����� ���� ������� �� �������� ������ �� ���������� �� �� ������ ��� �������'� (�� $� % �� �� �������� ��8� ����������������������������������� �� � ����������0������������������������ ���������������� ������������� ������'����������(������������ ������������������������������ �� ��������������������������� ���������������.������������������������������������������� ���'�(� ������������������������������������������������ ���� ���� ����� ����������������� ���� ��� ���������'� �� ���������� �� ��� �� �������� ��� ������� ��� ���� ������ ���� ���������� ��� �������� ��� �� �������� ���� ��� �#������ ���� ��� �������� ��� ��� ���������'� "���� ������ ���� ��� ����������� ����� ���������� � ���� ����� ��'� ��� �#������� ���� ����� ����� ��� ���� �� ���� ����������������������� ���������������������������������������������������������� ������ isel academic journal of electronics, telecommunications and computers http://editorial.isel.pt/journals ����������������������'����������� ��������� ������������������������ ����������������������� ����� ����������������������������� ���#������'�"������� ������ �� ����������� ������� ������������������������� �������������������������������'�&�������������� ����������� ��� ����������� ���������������� ��������������������� ������������������������������������ ���������� ���� �����#���'���������������� �����������������������������3������������� �� ��������#��b����������3������������+������������������������������0 ��������������'�(������� �� ��������� �����������������(������� �� ����������� ������������������������� ������� �� �����������' 6. heidegger’s hermeneutic (� ��� ���� � �� ���� ����������� ������� ��� ���� ���6� �� ���� *��������+�� ���� ����������� ��������� ���������������� ��� �� ��� ��� � ����� ��� ���� ������ ���� ����� �������'� ����������� ���������#��� ������������������ ��'�(��� ���� �������� �������������#������ ��� ��������������������������������������'������������� ����������������������������������� *��������� ��� ������� �� ��� ������� �� ��� �������������'� ����� � ������ ��������� ����� � ���� ��� ����� ���� ������ ������ ��������������� ��� ��� ���������� ��� ������ ��� �#������ ��� �������� ����'�(� ������������������������������ ���������������������������������� ������� ��� *��������� ��������� ����� �������������� �������� ������� �� ��� �� �� �����'� "� ���� ����������� ���� ����� �������������#��� �������������� ���������������������������� ��� ���� �������� ����������� ��*�����������������������' ���� 0���� ����� ��� being and time� ����� ���� �� ����� � ��� �������������� ��� � ����'� *��������� ��� ��� ���� ��� �� ����� ��� ��������� ��������� ��� ���� �������� ���� �������� ���������� ��� ���� �������� ������������������� ���������� ���� ������ ����� ����� ��� �� ��� ��� � ����'������������������1���������a��a����2��������������� ��������������� �����#����� �� ��� ������ �� ��� �������� ���� � 6� ���� �� 1��������� ��� �������������2� ���� ���� �� ��������� ��� ��������� ��� ������������'�������������� �� ������������������������������������������#� ���� ����� �����������1�����0 �� ��2�bedeutsamkeit'��������������������������������� ����� ������������������������ ������� �����������*��������� �����1������2������������������� ���������������������� �������dasein���������� 6� ������� ����'��������������������������� ������������8� ���������������������)����������� ������� ���' *��������+�� � ����� ��� �� ����'� *�� ����� ���� �#������ ��� ������ ���� �������� ������������ ����'�(������� ������������� ����������������c�����1������������������������2���� ���������������������� ���������������� ���������������� ��� ��������� '�-� ������� ����� ��� ���������������������0���� �������������������������������������������� ���� �����'� *��������� ��������� ��0���� ���� ����� ��� ����� ��������� ��� ������� �������� ��� �� ��� ��� ���������� �� ����� �����>umsicht?'������������ ������������ �� ������������0 �� ����� ������������������������������������������������������������������������� ���� ������� �� ��� ������������������������������������ ������ ��������1�����'2� andrew feenberg | function and meaning: the double aspects of technology i-etc, volume 1, nº 1, special inauguration issue, (2012) paper id-2 )��� ��� ��� ��� �� �� ���� ���� ����� ���� ��� �0 � ��������� ��������� ��� ��� ������� � ������������� ���� ��������� �������������c�������������������� ��������������������� �� ��� ����� ��� ���� ���������������� ��� ��� ����� ���� �������� ��� ������� �� ������'� &�� ����� ����� �����*������������������� ������������������������������������� �������������������� �������� ��� �������'����������� ��� *��������+�� ����� ����� ��������� ��� ���������� ������ ��� ����������� ����� ��������'� ���� ���� ��� �� ���� ���������������� � ������� ������� ���� ������ ���������� �����������������������������������'������������������ ��� ��������������� ������ ���� ������� ������ ��������� ���������������3���������� ������*������������������ ��������������a����������������������� ����������������������������'���������������� ���������������� �������� ������������������������������������������� ������ �� ���������� ������ �������� ����������� ��������'� *��������+�������������� ����������$� % ����� ����3��������������� �������������� �������������������� ��������������� ��������� ������������������������������������� ���������� �������'��������������� ����������������������������������������� ��������� ���� ������ ��� �������� ��� ��������� ��� ����� ������� ���������� � �#����� �� ��� �� ���� ����� ����������'������������������� ������������������ ����������������� �� ����������� ������� ���� ������ ���������� ��� ��������� ��� �#������ �'������ ��� ���� �� ���� ����#��� *��������+�� ������������� ����������������� ��� ���� ������� �' &����������������� ��������� ����� ��������� ������*��������+��������������c� the lecture course entitled fundamental concepts of metaphysics ���������������� ����� ����� ������ ������ ��� ����������� ����� �������' *��������� �#������� ����� ���� ����� �������� ��� dasein ��� �� ������ ������������ ����� ��� ����� ���� 1��2� ���������� ����� ���� ���� � ������ ��� �������������������������'�&��������� ��������������������������������������������������� ������ ��� ����� ��8�#���� ���������� ��� ����� ����� �������'� *��������� ������ ���� �� ����� ��� �������� ���� ������ ��� �#������ ���� ���� ���� ���6� ����� ��� ���� ��� ���� ������ ����'� (���������� ��� ������� �� �������������� ����'��������������� ������������������������� *���������������������������������dasein��������������������������� ����' )�������*������������������������������������������������� ����������������������� dasein������������������������ ����c�*�����������#��� ����������������� ����������� ������ ���� ��������������������� '�������1�� �������2������������������������� �������������a ��a������#�����������being and time.�(���������������������� ���� �� ������� ������������� ������� ������ �� ������������ �� �����'������� �������� �������3������� ������������������ ������������ ������������������� ���� �����������3��������������������'�*����� ������ � ���������� �������������� ��������������� ����������'�&����dasein���������������� �������������� � ����� �������������������������������������������� �� �������������'�(�� ����� ������������������*��������+���������������� ������������������������������������� ���� ��� ������������������ ���������1������2�����������dasein ��������� ����������������������� �� ���� ������������������� ����'� ���� �������� ������� ������ ���� ������ ��� � ����p����������� � ����� ���� ���������p��� ��� ���� ��� ��� ���� ��� ��� �� ��������� ���� ����� ���� ��� ���� �� ����� ��� ���� �����'��-������������ �� �������������������������������������������� ����������������� isel academic journal of electronics, telecommunications and computers http://editorial.isel.pt/journals ��� ��������������������'�(���� ����������� � ����������������������������������� ������ ������������������������������������'������� ����������������������������������������������� ���������������� �� �������������������������� ��������� ���� � ���������������������� ������� �#���������������������������������������� ��������������������'�)���*��������������� ��� ����������8������������������� �����������������������������������'�(�������� � ���� being and time ���� �� ����� ������� � �� ���� ��� �������� ����� ��� ���� �������� ����� ����� �� ��� ������ �� ���������������� � �� ����������������� ����������� �'������ �� ����������� ���������� �� ����� ���������������6����0���9�3��������*������������� ����������������� ��������"�� ���'� 7. technical creativity and democracy $��� ��� ���� �� ���������� �� '� (�� being and time, ���� ������� ��� �������� � � ����� ���������1��� ������������� ����������6� ������������������ �� ���������� ��������������2<:g='� *����������������������������������������������������� �� ��������������������������������� ���������� ����������'�)������ ���������������dasein ��������� ��� �������������������������'� ����� �������� ���� ��������� �� ��� ��� ���������� ������� ����� ����������'� (������� ���� ����� ��������������������������� ���������������������������������������������� ����������� � ��� � ����� ��� �� ��� ��� being and time. ������ � ������ ��������� �� ����������� ������ ��� ������ ���������1 �������������������������2��������� �����������'���������� �� ���� ��������������� �������� �� ���������������#������ �������������������������� ������' (� ����� ���������� ��� ��� � ���� ���� ����� ������� ��� ����� ������ �� ��� ������ ��� �� ������� ��� �� �������� ��� ���� ��� ������� ���� ����������'� ��� ��� ������� �� �������� ��������� ��� �� ������ ���� � ���� ��� �#�������� ��� �� ��� ��� ��� �������� ����� ���� ����������� �����������������������������'������������������ �������� ���������������������������� �� �� ���������������������������0 �� ������ ���������������������� ����� ��������������� ��������� ���������������� ���� ��������������� ����������������������� ���������������� ������ ��� ��� �� ��� ��� ��������'� ���� ���������� ��� ������� �� �������� �� ��� ��� �� ��� ����� ������������������ ����� �����������������������'�"�������� ����������������������������� ����������������� ���� �#������� ��������������� ������������ ����� ��������� ��������������� ������������������� �������������� ������������ ������ �����'�)��������� �� �������� ��� �������� ����� ������ ������� ���� ��������������� ����������'� (�� �� �� ����� ������ �� ���� ��8� �� ���� ����� ���������������� ��� ������� ���� ������������ ��� ����� �������� ��� �������� ��� ������������������������� ������� ��������������������� �����������' ����������� �������������#���������� ������ ��������� �������������� �'����� �� �����������"������������� �������������������+���� �������������#���������������������� ��������� �������'������������������������������������� ������ ���� ������������"������� ��� ���� ���������� ������ '� (�� ������� ���� �� ���� ������ ������ ��� ����������� ����� ����������������� ������������������� ���������������������'�(��������������������������� �������������������������������� ����'�9������������0 ��� ����������������������������������� andrew feenberg | function and meaning: the double aspects of technology i-etc, volume 1, nº 1, special inauguration issue, (2012) paper id-2 ���������������'�9����������������������������������������������� ��� �������� ���������.����a ��0����� ��� ���� ����'� 7������� ��� �������� ����� ����������� ���� ��� ����������� ���� ���� ����� ��� ���������������������������� �����#����������������� ������������������� ������������'� ���� ��������������������� ������� ������������������������ �������������� ���� ���������� ��������������� �0 ��� ��� �������������'� -#��������� �������������������������������������������� �� ����������������� ���'� &�� ���� ���� �������� ����� ����� ������� ���� ��� ������ ��� ����� ���� ������������ ��� ���� �� ��� ��� �������� � ��� ��������� ����'� ���� ����� ����������� ���� ������� ��� �� �������� ����� �� ������������� ��� �������� �� ��������� ��� ���� ���������� ��� ������� �� ������� ���� ���� ����������� ���� ���� ������ ��� � ����� ����� ������ ��� ����'� ��� ����� ��� ���� �� ������� ����������� ��� ����������������#������ ������� ��������������������������������������� �����a�����������������������������'�)������� ���������������� �������������a����������� � ����������������������������������������� ��������� ��� ������ ������������������ ��������� ����������� ��������������� ���� ���'��� ���������������������3��"�� ���+���������������� ��������������������������� �������������������� ���������������������� �� �������������� ����� ��� � �����0 a�� ��� ��� �����������'� 7������� ��� ������ ������� ��� �������� ���� ������ ��� �������������������������#������������� �������������� ������������������������������������ ��������������������� ����'� 8. references <:=�i�����$� % �� history and class consciousness�������'�7'�$�����������>��� �������"���"(�� ,������:n;:?��:e:' <4=�*�� ����"�� �����one-dimensional man�>)�������)�� ���,������:njf?��:gjagn' <e=�*�� ����"�� �����towards a critical theory of society����'�l�������q�������>9���o�� �� 7����������455:?��:ekaen' <f=��� ����)���������technology and the character of contemporary life�>��� ������������������� ��� ����,������:nkf?' <g=$����3���������� technology, time, and the conversations of modernity >9���o�� �� 7���������:nng?' <j=����������technology, time, and the conversations of modernity��f;' <;=�( ��'��k' <k=�( ��'��:;fa;g��:k4' <n=�)���������technology and the character of contemporary life��f;' <:5=�( ��'��:kn' <::=�( ��'��f' <:4=�( ��'��455' <:e=�( ��'��445' <:f=�������� ��� ��������l�����)����������'��community in the digital age�>$�������"l�� 7����������$�����0����455f?' <:g=�"������*����������being and time�������'�r'�"� .�����������-'�7� ������>9���o�� ��*������ s�7����:nj4?��efg'