Microsoft Word - 42landucci.docx CHEMICAL ENGINEERING TRANSACTIONS VOL. 82, 2020 A publication of The Italian Association of Chemical Engineering Online at www.cetjournal.it Guest Editors: Bruno Fabiano, Valerio Cozzani, Genserik Reniers Copyright © 2020, AIDIC Servizi S.r.l. ISBN 978-88-95608-80-8; ISSN 2283-9216 Subsea Oil Spill Risk Management Based on Sensor Networks Gianluca Tabellaa,*, Nicola Paltrinierib, Valerio Cozzanic, Pierluigi Salvo Rossia aDepartment of Electronic Systems, NTNU Norwegian University of Science and Technology, Norway bDepartment of Mechanical and Industrial Engineering, NTNU Norwegian University of Science and Technology, Norway cDepartment of Civil, Chemical, Environmental, and Materials Engineering, University of Bologna, Italy gianluca.tabella@ntnu.no The use of Wireless Sensor Networks (WSNs) in support of Dynamic Risk Assessment regarding oil spills still lacks a proper integration. WSNs enable prompt responses to such emergencies through an appropriate inspection, thus avoiding possible larger disasters. This work proposes a methodology for the setup of a WSN as a Leak Detection System in which a Fusion Center collects sensors’ binary decisions and provides a more reliable decision about the presence/absence of a leak. The detection rules are based on statistical signal processing techniques, and the choice of the optimal thresholds is made through the optimization of three objective functions tailored to the Oil&Gas industry. Detection performances are assessed in terms of the Receiver Operating Characteristic (ROC) curve. The case study is the Goliat FPSO, a production platform located in the Barents Sea, and related requirements dictated by Norwegian authorities to prevent oil spills. The considered WSN monitors the subsea manifolds through passive acoustic sensors. 1. Introduction Oil spills are known to cause a highly negative impact on the safety of offshore workers, the environment, and productivity. The early detection of a spill is crucial to limit its potential consequences. A Leak Detection System (LDS) is reliable if it can provide a high rate of correct detections ensuring a limited rate of false alarms, thus avoiding unnecessary production shutdowns and costly Remotely Operated Vehicles (ROV) inspections. Different technologies, among which the use of passive acoustic sensors, are nowadays available and are used to monitor the external underwater environment and the process conditions (Adegboye et al., 2019; Baroudi et al., 2019). Passive acoustic sensors have shown a high level of accuracy enabling the possibility to localize the spill source. This can be done without the need to install the sensors near the leaking component (which is a limitation of many other LDSs). Also, this technology can detect all hydrocarbon fluids. Acoustic sensors are easy to install and are appropriate for retrofitting. These properties make this LDS among the most used. The importance of a reliable LDS creates the need for a framework that integrates it into the Dynamic Risk Assessment (DRA). This is possible as the use of a distributed Wireless Sensor Network (WSN) can provide real-time monitoring of the subsea environment increasing the level of knowledge on the system allowing a more accurate DRA (Paltrinieri et al., 2014, 2019a). So far, the application of WSNs in the Oil&Gas industry has only been introduced (Paltrinieri et al., 2019b). This work gives a methodology for the setup of passive acoustic sensors in a WSN used for monitoring subsea templates and discusses its performances. 2. Signal Model The WSN aims at detecting possible oil spills, so the problem is formalized as a binary hypothesis testing with the null hypothesis corresponding to a non-spill scenario, and the alternative hypothesis corresponding to a spill scenario. For the generic th sensor, the two following different signal models are assumed for each hypothesis: DOI: 10.3303/CET2082034 Paper Received: 9 February 2020; Revised: 8 May 2020; Accepted: 26 July 2020 Please cite this article as: Tabella G., Paltrinieri N., Cozzani V., Salvo Rossi P., 2020, Subsea Oil Spill Risk Management Based on Sensor Networks, Chemical Engineering Transactions, 82, 199-204 DOI:10.3303/CET2082034 199 : = ∙ , +: = ⟹ : ~ 0, , ⋅ +: ~ 0, (1) where: • is the signal (sound pressure) received at the th sensor where = 1,2,…, ; • ~ 0, is a Gaussian random variable representing the emitted signal caused by the spill; • ~ 0, is Additive White Gaussian Noise having the same power for any sensor; • , is the Amplitude Attenuation Function (AAF) which only depends on the distance between the position ( th sensor position) and (leak position). The AAF is treated deterministically and represents the loss of the acoustic intensity level and accounts for seawater absorption and geometrical spreading (Stojanovic, 2006): 10log , = − ∙ 10 ‖ − ‖ − ℓ − sc ⋅ 10log ‖ − ‖ℓ (2) From which, the AAF can be obtained: , = ℓ‖ − ‖ sc 10 ℓ ‖ ‖ (3) where is the seawater absorption coefficient in dB km⁄ , ‖ − ‖ and ℓ (reference length) are in meters, and sc is the spreading coefficient. The absorption coefficient is obtained using the Francois & Garrison equation (Francois and Garrison, 1982a, 1982b). The speed of sound (required by Francois & Garrison) is obtained using the updated Chen & Millero equation (Wong and Zhu, 1995). 3. Wireless Sensor Network Model The modeled WSN is made of passive acoustic sensors monitoring the external environment (as shown in Figure 1). The th sensor, with a given sampling frequency, senses the received signal amplitude and sends to a Fusion Center (FC) its binary local decision on whether the sensed amplitude is caused by a spill. The choice of local binary decision is due to the energy constraints imposed by the use of a WSN (Shoari et al., 2016), such constraint will also reduce operating costs as only one bit is transmitted when a spill is detected. Finally, the FC takes a global decision on the occurrence of the spill based on the received ’s. Figure 1: Wireless Sensor Network Model 4. Detection Rules Each sensor performs an Energy Test, which is Uniformly Most Powerful for this application, where the signal energy is compared to a threshold to assess its decision (Ciuonzo and Salvo Rossi, 2017): = 1, ≥0, < (4) As the statistics of the received signal is known, the Local Probability of Detection ( , ) and Local Probability of False Alarm ( , ) for the th sensor can be defined as follows: H0 / H1 Fusion Center Sensor 1 Sensor K Sensor 2 y1 y2 yK d1 d2 dK Ĥ 200 , = Pr ≥ = 2 , ⋅ + (5) , = Pr ≥ = 2 (6) where is the complementary cumulative distribution function of the standard normal random variable: = 12 exp − 2 (7) The proposed method assumes the Signal-to-Noise Ratio = ⁄ at ℓ from the source to be known. The Counting Rule is used as Fusion Rule by the FC because of its simplicity which suits the constraint of low processing costs. This rule uses the local decisions as an input and has the following form: = , ≥ Λ, < Λ (8) This indicates that the FC counts the number of sensors detecting the spill and compares it to a threshold Λ. In case the sum is equal or higher than the threshold, the FC sends an alarm. 5. Threshold Selection Three different optimality criteria based on the Receiver Operating Characteristic (ROC) curve will be analyzed (Liu, 2012): • Youden Index ( ): ∗ = arg max = arg max − (9) • Closest-to-(0,1) ( ): ∗ = arg min = arg min 1 − + (10) • Concordance Probability ( ): ∗ = arg max = arg max ⋅ 1 − (11) These definitions are applicable both for the sensors and the FC with the appropriate substitutions ( is and Λ; is , and ; is , and ; ∗ is ∗ and Λ∗). The selection of the optimal threshold ∗ for the th sensor is carried out through a grid search where one optimal value is found for each one of the criteria. More specifically, the metrics in the optimality criteria are computed referring to average performances with respect to the hotspot positions , where = 1,2,…, . This is necessary as the Probabilities of Detection (both local and global) depend on the leak position. The hotspots are those components of the subsea production system that, in case of failure, would be the source of a spill. Also, it is assumed that the selected hotspots have the same failure rate and their spills cause signals having the same power . Therefore: , = 1 , , , = , optimize objective function ∗ (12) where, for the th sensor, ∗ is the chosen local threshold (using one of the criteria), , and , are its average performances and , , is , when the leak source is the th hotspot by using , . The choice of the optimal threshold at the FC follows the local threshold choice and uses the same procedure: = 1 ,= optimize objective function Λ∗ (13) where Λ∗ is the chosen global threshold. = Pr = and = Pr = are the values of Global Probability of Detection and Global Probability of False Alarm, where means that the th hotspot is 201 modeled as the leak source, and the bar denotes the average probability. , and are obtained via Monte Carlo Simulation requiring the simulation of the local decisions using the previously chosen local thresholds. 6. Case Study – Goliat FPSO The Goliat FPSO is an offshore platform located in the Norwegian Barents Sea equipped with a multi-template Subsea Production System. Each template can host up to four wellheads and the manifold. The latter is monitored by three passive acoustic sensors to detect the presence of an oil spill (Bjørnbom, 2011; Røsby, 2011). For an overview of the subsea equipment, the reader could refer to the specific literature (Bai and Bai, 2012). Figure 2: Scheme of Goliat’s subsea template: the grey elements are the structure and the Christmas Trees, the blue lines are the main streamlines, the green dots are the sensors, and the red dots are the hotspots 20 hotspots (connections and valves) were recognized in the manifold. Hotspots and sensors are assumed to be at the same height. The following parameters are used for this case study: Table 1: Parameters used to simulate the spill’s sound emission and its Amplitude Attenuation Function Parameter Value Note = ⁄ 13 dB ℓ = 1 m Noise Variance 1 Normalized Reference Frequency 2.5 kHz (Eckert et al., 1993), used for AAF Temperature 3.8 ℃ (Institute of Marine Research, 2020), used for AAF Salinity 3.5 % (Institute of Marine Research, 2020), used for AAF Depth 350 m (Bjørnbom, 2011), used for AAF pH 8 (Vetrov and Romankevich, 2004), used for AAF Spreading Coefficient sc 1.5 (Stojanovic, 2006) 7. Results The values of = , ∙ ⁄ averaged among all hotspots show a mean attenuation of 90.28 % and are the following: Table 2: Averaged Signal-to-Noise Ratio at the sensors Sensor 1 Sensor 2 Sensor 3 2.4 dB 3.4 dB 1.4 dB At sensor-level (Table 3 and Figure 3), the optimization of results in local thresholds with values distant from those obtained optimizing or which tend to be similar. When is used, in fact, the thresholds are oriented towards smaller values of , and , . The values of Area Under the Curve (AUC) of the averaged ROC curves among the three sensors have a standard deviation equal to 4 ∙ 10 , this justifies the similar average performances among the sensors when tuned using the same objective function. 1 2 3 5 m 202 Table 3: Local threshold selection’s results Sensor Value of Optimized Function Threshold , , 1 = 0.1896 1.6380 0.3902 0.2006 = 0.5886 0.7919 0.5451 0.3735 = 0.3417 0.8427 0.5328 0.3586 2 = 0.2064 1.7268 0.3952 0.1888 = 0.5785 0.8271 0.5496 0.3631 = 0.3504 0.8896 0.5354 0.3456 3 = 0.1923 1.5169 0.4104 0.2181 = 0.5845 0.7947 0.5497 0.3727 = 0.3450 0.8435 0.5378 0.3584 Figure 3: Averaged local ROC curves displaying the optimal points according to the different applied criteria At FC-level, results were obtained with 10 Monte Carlo runs (Table 4 and Figure 4). The performance at a given threshold varies according to the objective function used for the sensors. When sensors are tuned using , the optimal global threshold is divided between the value 1 if and are optimized, and 2 if is optimized. When sensors are tuned using or , the optimal global threshold is always 2 using any optimization criterion. The highest value of AUC at the FC is obtained when sensors are tuned using . Table 4: Global threshold selection’s results Function used for Sensors Value of Optimized Function Threshold Youden Index ( ) = 0.2546 2 0.3608 0.1062 = 0.5553; = 0.3774 1 0.7442 0.4928 Closest-to-(0,1) ( ) = 0.2562; = 0.5334; = 0.3905 2 0.5652 0.3090 Concordance Probability ( ) = 0.2598; = 0.5363; = 0.3899 2 0.5472 0.2875 Figure 4: Averaged global ROC curves displaying the optimal points according to the different applied criteria 203 8. Conclusions It is clear how the choice of the objective function at sensor-level is fundamental to determine the performances at FC-level for a given global threshold. The case study showed how tuning the sensors using the Youden Index increases the global AUC and orients the performances towards lower values of false alarm rate of the LDS, which may be preferable to avoid shutdowns. The optimal global thresholds show a similar behavior on the ROC space if compared to the results obtained when computing the optimal local thresholds. However, the tendency of the Youden Index to generate thresholds having a lower probability of detection and false alarm with respect to those generated by the other two indexes is less evident at FC-level since only + 2 points can be placed on the ROC space. The three objective functions can also be adapted and corrected using coefficients to fit specific applications and requirements. The proposed methodology shows how important the number of sensors and their positioning can be and how the network performances heavily rely on the signal model. For this reason, more information regarding the statistical properties of the signal and other contributions that influence the AAF should be integrated if available. These factors can be signal perturbations, interferences, ambient noise, and oceanic phenomena (currents, tides, internal waves, etc.). This work is a step towards the integration of subsea monitoring using WSNs with Risk Assessment techniques necessary to localize the hotspots and to select the most appropriate objective function. References Adegboye, M.A., Fung, W.K., Karnik, A., 2019. Recent Advances in Pipeline Monitoring and Oil Leakage Detection Technologies: Principles and Approaches. Sensors 19. Bai, Y., Bai, Q., 2012. Subsea Engineering Handbook, 1st ed. Gulf Professional Publishing, Houston, TX. Baroudi, U., Al-Roubaiey, A.A., Devendiran, A., 2019. Pipeline Leak Detection Systems and Data Fusion: A Survey. IEEE Access 7, 97426–97439. Bjørnbom, E., 2011. Goliat – Leak detection and monitoring from template to satellite. URL (accessed 6.6.20). Ciuonzo, D., Salvo Rossi, P., 2017. Distributed detection of a non-cooperative target via generalized locally- optimum approaches. Inf. Fusion 36, 261–274. Eckert, E.G., Maresca, J.W., Hillger, R.W., Yezzi, J.J., 1993. Location of Leaks in Pressurized Petroleum Pipelines by Means of Passive-Acoustic Sensing Methods, in: Durgin, P., Young, T. (Eds.), Leak Detection for Underground Storage Tanks. ASTM International, West Conshohocken, PA, pp. 53–69. Francois, R.E., Garrison, G.R., 1982a. Sound absorption based on ocean measurements: Part I: Pure water and magnesium sulfate contributions. J. Acoust. Soc. Am. 72, 896–907. Francois, R.E., Garrison, G.R., 1982b. Sound absorption based on ocean measurements. Part II: Boric acid contribution and equation for total absorption. J. Acoust. Soc. Am. 72, 1879–1890. Institute of Marine Research, 2020. Mareano. URL (accessed 6.6.20). Liu, X., 2012. Classification accuracy and cut point selection. Stat. Med. 31, 2676–2686. Paltrinieri, N., Comfort, L., Reniers, G., 2019a. Learning about risk: Machine learning for risk assessment. Saf. Sci. 118, 475–486. Paltrinieri, N., Landucci, G., Salvo Rossi, P., 2019b. An Integrated Approach to Support the Dynamic Risk Assessment of Complex Industrial Accidents. Chem. Eng. Trans. 77, 265–270. Paltrinieri, N., Scarponi, G.E., Khan, F., Hauge, S., 2014. Addressing dynamic risk in the petroleum industry by means of innovative analysis solutions. Chem. Eng. Trans. 36, 451–456. Røsby, E., 2011. Goliat development project - Subsea leak detection design. URL (accessed 6.6.20). Shoari, A., Mateos, G., Seyedi, A., 2016. Analysis of Target Localization With Ideal Binary Detectors via Likelihood Function Smoothing. IEEE Signal Process. Lett. 23, 737–741. Stojanovic, M., 2006. On the relationship between capacity and distance in an underwater acoustic communication channel, in: Proceedings of the 1st ACM International Workshop on Underwater Networks - WUWNet ’06. ACM Press, New York, NY, pp. 41–47. Vetrov, A., Romankevich, E., 2004. Carbon Cycle in the Russian Arctic Seas, 1st ed. Springer, Berlin, Germany. Wong, G.S.K., Zhu, S., 1995. Speed of sound in seawater as a function of salinity, temperature, and pressure. J. Acoust. Soc. Am. 97, 1732–1736. 204