Jtam-A4.dvi JOURNAL OF THEORETICAL AND APPLIED MECHANICS 54, 2, pp. 659-670, Warsaw 2016 DOI: 10.15632/jtam-pl.54.2.659 WAVELETS AND PRINCIPAL COMPONENT ANALYSIS METHOD FOR VIBRATION MONITORING OF ROTATING MACHINERY Hocine Bendjama Research Center in Industrial Technologies (CRTI), Cheraga, Algiers; e-mail: hocine bendjama@daad-alumni.de Mohamad S. Boucherit Department of Automatic, Polytechnic National School (ENP), El-Harrach, Algiers; e-mail: ms boucherit@yahoo.fr Fault diagnosis is playing today a crucial role in industrial systems. To improve relia- bility, safety and efficiency advanced monitoring methods have become increasingly im- portant for many systems. The vibration analysis method is essential in improving con- dition monitoring and fault diagnosis of rotating machinery. Effective utilization of vi- bration signals depends upon effectiveness of applied signal processing techniques. In this paper, fault diagnosis is performed using a combination between Wavelet Transform (WT) and Principal Component Analysis (PCA). The WT is employed to decompose the vibration signal of measurements data in different frequency bands. The obtained decomposition levels are used as the input to the PCA method for fault identification using, respectively, the Q-statistic, also called Squared Prediction Error (SPE) and the Q-contribution. Clearly, useful information about the fault can be contained in some levels of wavelet decomposition. For this purpose, the Q-contribution is used as an evaluation criterion to select the optimal level, which contains the maximum information.Associated to spectral analysis and envelope analysis, it allows clear visualization of fault frequencies. The objective of this method is to obtain the information contained in the measured data. Themonitoring results using real sensormeasurements from a pilot scale are presented and discussed. Keywords:vibration, fault diagnosis,wavelet analysis, principal componentanalysis, squared prediction error 1. Introduction Growingdemand forhigherperformance, safety andreliability of industrial systemshas increased the need for condition monitoring and fault diagnosis. During the two past decades, various monitoring methods have been developed, including dynamics, vibration, tribology and non- -destructive techniques (Altmann, 1999; Yang, 2004). The vibration analysis is one of the most important methods used for condition monitoring and fault diagnosis, because it always carries the dynamic information of a system. Effecti- ve utilization of vibration signals depends upon the effectiveness of applied signal processing techniques. The analysis of stationary vibration signals has largely been based on well-known spectral techniques such as: Fourier Transform (FT) and ShortTimeFourier Transform (STFT) (Seker and Ayaz, 2002; Shibata et al., 2000). Unfortunately, these methods are not suitable for non-stationary signal analysis (Wu and Liu, 2008). In order to solve this problem, Wavelet Transform (WT) has been developed. WT is used to extract approximations and detail coef- ficients of measurements data with different frequency bands by using successive low-pass and high-pass filtering. This makes the application of WT for non-stationary signal processing an area of active research over the past decade. An overview of the WT used in vibration signal analysiswas providedbyLitak andSawicki (2009), Al-Badour et al. (2011) andYan et al. (2014). 660 H. Bendjama,M.S. Boucherit The original signal using WT can be decomposed into approximations and details versions with different resolutions. The decomposed levels will not change their information in the time domain (Gaing, 2004). However, useful information can be contained in some sub-bands. So, the fault can be detected from a given level of resolution. This is based on the choice of an indicator to determine the optimal level where failure can occur. The selection of themost reliable indica- tor has been studied by several authors. A large number of applications have been reviewed, e.g. Prabhakar et al. (2002) selected periodic impulses of bearing faults in the time domain based on the low and high frequency nature of decomposed levels. Similar analyses were carried out by Purushotham et al. (2005) in order to extract periodic impulses from time signals using discre- te wavelet transform at Mel-frequency scales. Chinmaya and Mohanty (2006) used sidebands of gear meshing frequencies as an evaluation criterion for gear faults diagnosis. Djebala et al. (2008) analyzed vibration of faults inducing periodical impulsive forces by selecting the kurtosis as indicator. In another study,Gavrovska et al. (2009) described the optimal selection of decom- position levels in thewavelet transformused for both high-frequency and low-frequency filtering of the ECG signal. Yaqub et al. (2011) estimated the bandwidth of the resonant frequency band of vibration data by adaptive selection of wavelet decomposition levels. The adaptive criterion was based on saturated dissemination of the signal energy over thewavelet decomposition nodes. In thiswork,wepropose touse a combinationbetweenWTandPCAfor improving thevibra- tionmonitoring. PCA is amultivariate analysis technique, also a dimension reduction technique. It finds the directions of significant variability in the data by forming linear combinations of va- riables. Its application for vibration analysis is suggested in several papers (De Moura et al., 2011; Shao et al., 2014). The aim of the proposed combined method is to provide a solution to the fault diagnosis problem.WT is used to extract approximations and details vectors in order to obtain multiple data series atdifferent resolutions.Clearly, useful information is contained in somedecomposition levels. The obtained levels are used as the input to the PCA method for identifying abnormal situations in different frequency bands using the Q-statistic or Squared Prediction Error (SPE) and Q-contributions. In order to extract useful information, the Q-contribution is used as the principal criterion to select the optimal level of resolution. Associated to spectral analysis and envelope analysis, it allows clear visualization of the frequencies faults. The proposedmethod is evaluated using the experimental measurements data in the cases of mass unbalance and gear fault. The remainder of this paper is structured as follows. Section 2 presents the fault diagnosis method using WT and PCA along with its formulations. The experimental setup is discussed in Section 3. Themonitoring results and discussion are presented in Section 4. Finally, Section 5 concludes our work and contributions. 2. Fault diagnosis method 2.1. Wavelet transform Wavelet Transform (WT) is a relatively new signal processing tool. Due to its strong capa- bility in the time and frequency domain analysis, it is applied by many researchers in diverse applications (see, for example, Litak et al., 2009; Sen et al., 2010). WT describes a signal by using the correlation with translation and dilatation of a function called mother wavelet; it inc- ludes ContinuousWavelet Transform (CWT) andDiscreteWavelet Transform (DWT). Let s(t) be the signal, then the CWT of s(t) is defined as CWT(a,b)= 1 √ |a| ∞ ∫ −∞ s(t)ψ∗ (t− b a ) dt (2.1) Wavelets and principal component analysis method... 661 where ψ∗(t) is the conjugate function of the mother wavelet ψ(t), a and b are the dilation (scaling) and translation (shift) parameters, respectively. DWT is derived from discretization of CWT. The most common discretization is dyadic. DWT is found to yield fast computation of theWT. It is given by DWT(j,k) = 1√ 2j ∞ ∫ −∞ s(t)ψ∗ (t−2jk 2j ) dt (2.2) where a and b are replaced by 2j and 2jk, j is an integer. A very useful implementation of DWT, called multiresolution analysis (Mallat, 1989), is demonstrated inFig. 1. TheDWTanalyzes the signal at different frequencybandswith different resolutions by decomposing the signal into several levels; approximations (A1,A2,A3,...) and details (D1,D2,D3,...). The signal is decomposed at the expected level. DWT employs two sets of functions, called scaling function and wavelet function (Mallat, 1989), which are associated withLow-pass (L) andHigh-pass (H) filters, respectively. Theapproximations are thehigh-scale, low-frequency components and the details are the low-scale, high-frequency components of the signal. Fig. 1. Principle of DWT decomposition Selection of an appropriate wavelet is very important in signal analysis. There are many functions available that can be used, such as Haar, Daubechies, Meyer, and Morlet functions (Chui, 1992; Daubechies, 1988). In this application, we use the Daubechies wavelet for fault diagnosis of the mass unbalance and gear fault. 2.2. Principal Component Analysis Principal Component Analysis (PCA) (Chiang et al., 2001; Jolliffe, 2002) is a projection statistical method used for dimensionality reduction. It produces a lower-dimensional represen- tation in a way that preserves the correlation structure between the variables. Given a set of n observations or samples and m variables stacked into a matrix X, whose variance-covariance matrix has eigenvalue-eigenvector pairs (λ1,p1),(λ2,p2), . . . ,(λm,pm) (2.3) where λ1 ­ λ2 ­ λ3 ­ ··· ­ λm ­ 0. The principal decomposition component ofX can be represented as X=TPT+E= l ∑ i=1 tip T i +E l < m (2.4) where T = [t1, t2, . . . , tl] is defined to be the matrix of principal component scores, P = [p1,p2, . . . ,pl] is the matrix of principal component loadings, E is the residual matrix in the sense of the minimum Euclidean norm and l is the index of the Principal Components (PCs). 662 H. Bendjama,M.S. Boucherit The identification of the PCA model thus consists in estimating its parameters by an eigenvalue-eigenvector decomposition and determining the number of PCs l to retain. Many procedures have been proposed for selecting the number of PCs to be retained (Kano andHase- be, 2001). In this study, we use the experiential method (Nomikos andMacGregor, 1995) which judges that the cumulative sum contribution of the anterior l PCs is higher than 0.85, as follows 100 · ∑l i=1λi ∑m i=1λi > 85% (2.5) The basic idea of the process of fault detection using PCA is to collect normal observation data to establish the PCA model. When a new observation data is subject to fault, these new data can be compared to the PCA model and its threshold. In order to detect the abnormal changes of the new data, the Q-statistic or SPE is used as follows Q-statistic =SPE= eTe= ‖x(I−PPT)‖2 (2.6) The process is considered normal if Q-statistic¬ δ2Q (2.7) where δ2Q denotes the confidence limit or threshold for theQ-statistic. It can be calculated from its approximate distribution (Jackson andMudholkar, 1979) δ2Q = θ1 [ Cα √ 2θ2h 2 0 θ1 +1+ θ2h0(h0−1) θ21 ] (2.8) where θi = ∑m j=k+1λ i j, i =1,2,3 and h0 =1−2θ1θ3/(3θ22), and λj is the eigenvalue, Cα is the critical value of the normal distribution and m is the number of all PCs. The threshold is used todeterminewhether thedata iswithin rangeof themodel.Tocompare the new data to the PCA model, a confidence limit of α = 95% is used. Any point below the confidence limit line is considered to have a normal variance from the selected number of PCs, and any point above this line is considered to have an abnormally high level of variance. After the fault is detected, i.e. the newobservation data exceed the threshold line of thePCA model but can not be assure in what place the fault appears in the process. An assignable cause is determined by the contribution plot. The contribution plots are bar graphs of the Q-residual contribution of each variable calculated as in equation (2.9) (Xuet al., 2008). Thevariable having the largest residual produces the worst compliance to the PCAmodel, and indicates the source of the fault Q-contribution= conti = ‖ei‖2 Q-statistic (2.9) where ei presents the i-th element of the residual vector e and conti is the contribution of the i-th variable to the total sum of variations in the residual space. By using the WT, the time domain information will not be lost when the signal is decom- posed. In order to extract useful information in different decomposition levels, the contribution plots are used as an evaluation criterion for selecting the optimal level which contains the ma- ximum information about the fault. A flowchart of the fault detection and diagnosis method based onWT and PCA is illustrated in Fig. 2. Wavelets and principal component analysis method... 663 Fig. 2. Flowchart of the fault diagnosis method 3. Experimental setup Theproblemof diagnosing the degradation of working conditions of rotatingmachinery is extre- mely important in industries to reduce the productivity loss. The measurement of vibration applied to the condition monitoring and fault diagnosis requires different types and levels of equipment and techniques. In the present study, an experimental system is used and the vibra- tion response for mass unbalance and gear fault are obtained. 3.1. System description The experimental system consists of a test rig built by S’tell Diagnostic (France), a data acquisition system (OROS OR25, 4 channels), piezoelectric accelerometers (PCB Piezotronics 353B34) and a PC (see Fig. 3). The system is driven by a 0.18kW induction motor giving an output of 0-1500rpm, controlled by a variable speed drive. To confirm the feasibility of the proposedmethod, we collect real vibration signals using the experimental test rig illustrated in Fig. 3, where 1 and 2 are gears with 60 and 48 teeth, respectively, H1, H2, H3 and H4 are the bearing housings. Thevibration signals are taken onbearinghousingH1bymeans of twopiezoelectric accelero- meters measuring radial vibration, i.e. in the Vertical Direction (VD) and Horizontal Direction (HD). These measurements are repeated for different states at different rotation speeds. The data acquisition is performed using the OR25 software. The vibration signals measured have 664 H. Bendjama,M.S. Boucherit Fig. 3. Illustration of the experimental system length of acquisition of 400milliseconds. The sampling frequency is 5120Hz and each signal has 2048 samples. 3.2. Faults description The experiment described in this paper, performs the condition monitoring of rotating ma- chinery to predict some anomalies thatmay occur under differentmeasurement conditions such as: mass unbalance and gear fault. 3.2.1. Mass unbalance Mass unbalance is one of the most common causes of vibration. It is a condition when the center ofmass does not coincide with the center of rotation, due to unequal distribution ofmass about the center of rotation. It is simulated in this study by an additional weight on the disk. The mass unbalance creates a vibration frequency exactly equal to the rotational speed, with an amplitude proportional to the amount of unbalance (Tandon and Parey, 2006). Figure 4 represents themeasured signals of themass unbalance at a speed of 900 rpm inVD and HD. Fig. 4. Vibration signals of the mass unbalance collected at 900rpm: (a) VD and (b) HD 3.2.2. Gear fault Vibrations of a gear aremainly producedby the shock between teeth of the twowheels.Gear fault is simulated with a gap between teeth. The vibrationmonitored on a faulty gear generally exhibits a significant level of vibration at the toothmesh frequency (i.e. the number of teeth on the gear multiplied by its rotational speed) and its harmonics (Tandon and Parey, 2006). Figure 5 represents the measured signals of the gear fault at a speed of 900rpm in VD and HD. Wavelets and principal component analysis method... 665 Fig. 5. Vibration signals of the gear fault collected at 900rpm: (a) VD and (b) HD 4. Results and discussion The structure of proposed fault diagnosis technique involves two parts: the first one is develop- ment and training of thePCAmodel; the second is testing the process fault based on the trained model (see Fig. 2). The measurements used in training phase represent the normal operating conditions of the system, i.e. system without fault. The vibration signals used in thiswork have been gained through the practicalmeasurement, including a normal state, mass unbalance and gear fault. The data collection has been carried out according to the following routine: vibration signals used in the training are taken in VD andHD at seven different rotating speeds between 300 and 1425rpm. In the testing phase, each fault of the process has been measured in the radial direction at five different rotating speeds: 675, 900, 1125, 1275 and 1425rpmwhich, respectively, correspond to 11.25, 15, 18.75, 21.25 and 23.75Hz. Themultiresolution analysis is applied by using theDaubechies wavelet of the order 4 (db4) and level 4. It may be noted that the same wavelet with the same level of decomposition is applied to each signal for the training and testing phases. The results of db4 decomposition of the vibration signals of the mass unbalance and gear fault collected at 900rpm in the radial direction are given respectively in Figs. 6 and 7. Afterdecompositionwithdb4, theapproximationsanddetails vectors of each signalmeasured at the same rotation speed in the radial direction are collected in a matrix. Thus the input matrices of thePCAmethod are formed.During the training phase, sevenmatrices are collected for identifying the PCA model. Through PCA method, the anterior 8 PCs accumulation sum contribution rate is 92.46%, as shown in Table 1. These first 8 PCs are selected for the fault identification. The detection threshold of Q-statistic is calculated according to equation (2.8), which is 2.2414. To evaluate the fault detection method, the detection ratio is used. It is defined as the number of samples whoseQ-statistic values go beyond the threshold to the total samples. If the detection ratio is less than 20%, the fault is not successfully detected (Xuet al., 2008). In normal operating conditions, 2.83% of the Q-statistic samples are above the threshold value. It implies that the model has captured themajor correlation and variance among the system variables. In the testing phase, fivematrices of each fault are collected for validating the PCAmethod. The newdata are compared to the PCAmodel and its threshold. The evaluation results of fault detection rate are summarized in Tables 2 and 3.We show clearly that themajority samples of the Q-statistic at different rotation speeds are above the threshold value. By comparing these values to 2.83% obtained in the normal state, it is also clear that an abnormal situation has occurred in the process. The objective of the proposed method is to demonstrate the effectiveness of the Q- contribution as a principal criterion for selecting the optimal level of wavelet decomposition. 666 H. Bendjama,M.S. Boucherit Fig. 6. Decomposition results of the measured signals of the mass unbalance at 900rpm in VD (left) and HD (right) Fig. 7. Decomposition results of the measured signals of the gear fault at 900 rpm in VD (left) and HD (right) Wavelets and principal component analysis method... 667 Table 1.PCs values PCs Eigenvalues Variance [%] 1 1.6459 16.46 2 1.6159 16.16 3 1.2381 12.38 4 1.1741 11.74 5 1.0642 10.64 6 0.9358 9.36 7 0.8259 8.26 8 0.7458 7.46 9 0.3934 4.13 10 0.3611 3.54 Table 2.Evaluation results of the mass unbalance detection rate rpm 675 900 1125 1275 1425 SPE [%] 35.59 67.91 79.00 83.30 85.40 Table 3.Evaluation results of the gear fault detection rate rpm 675 900 1125 1275 1425 SPE [%] 70.94 81.59 76.36 80.17 87.50 Fig. 8. Contribution plots of the mass unbalance (left) and the gear fault (right) 668 H. Bendjama,M.S. Boucherit The contribution plots are bar graphs of the Q-residual contribution of each variable or decom- posed vector (Fig. 8). The level having the largest value produces the worst compliance to the PCAmodel and indicates the desired level of decomposition. As shown in Fig. 8, decomposition levels 1 to 4 and 6 to 9 present respectively the detail vectors for VD and HD, and levels 5 and 10 stand respectively for the approximation vectors. It shows an obvious difference between the levels. From Fig. 8 (left), it can be seen that levels 5 and 10 have the largest values. We show also in Fig. 8 (right), that the contribution plot using db4 occurs in the third level. So our choice is attached to approximation A4 for themass unbalance and detail D3 for the gear fault. In order to diagnose the mass unbalance and the gear fault from the selected level we use, respectively, spectral analysis and envelope analysis. Figure 9 (left) shows the spectra of fourth approximations (A4). Frequency peaks at 11.25, 15, 18.75, 21.25 and 23.75Hz are present, which could be related to a mass unbalance fault. Figure 9 (right) illustrates the FT of the envelopes of the selected details. It can be seen from this figure that the peaks at the rotation frequencies of the shaft (11.25, 15, 18.75, 21.25 and 23.75Hz) and their multiples (×2,×3, . . .) are present in the frequency spectra. This clearly indicates a gear fault. Fig. 9. Spectra of selected levels of the mass unbalance (left) and the gear fault (right) 5. Conclusion This paper presents a combined approach based on the WT and PCA method to improve the conditionmonitoring and fault diagnosis of rotatingmachinery. It is adapted to obtainmultiple data series at different levels using wavelet decomposition and reduce the number of variables needed tomonitor theprocess throughPCA.Thedetectionmethod is basedonaconfidence limit estimated from normal conditions. The aim is to select the optimal level of resolution using the Wavelets and principal component analysis method... 669 residual contribution of each decomposed vector, for a possible diagnosis. The proposedmethod has been tested on real measurement signals collected from a vibration system containing mass unbalance and gear fault. Better experimental results have been obtained by identifying the type of fault. Hence the WT combined with the PCA method is a successful approach to vibration monitoring. References 1. Al-Badour F., Sunar M., Cheded L., 2011, Vibration analysis of rotating machinery using time-frequency analysis and wavelet techniques, Mechanical Systems and Signal Processing, 25, 2083-2101 2. Altmann J., 1999, Application of discrete wavelet packet analysis for the detection and diagnosis of low speed rolling-element bearing faults, Ph.D. thesis,MonashUniversity,Melbourne, Australia 3. Chiang L.H., Russell E.L., Braatz R.D., 2001, Fault Detection and Diagnosis in Industrial Systems, Springer-Verlag, London 4. ChinmayaK.,MohantyA.R., 2006,Monitoringgearvibrations throughmotor current signature analysis and wavelet transform,Mechanical Systems and Signal Processing, 20, 1, 158-187 5. Chui Ch.K., 1992,An Introduction to Wavelets, Academic Press, NewYork 6. Daubechies I., 1988, Orthonormal bases of compactly supported wavelets, Communication on Pure and Applied Mathematics, 41, 909-996 7. De Moura E.P., Souto C.R., Silva A.A., Irmão M.A.S., 2011, Evaluation of principal com- ponent analysis and neural network performance for bearing fault diagnosis from vibration signal processed by RS andDF analyses,Mechanical Systems and Signal Processing, 25, 5, 1765-1772 8. DjebalaA.,OuelaaN.,HamzaouiN., 2008,Detection of rolling bearing defects using discrete wavelet analysis,Meccanica, 43, 339-348 9. Gaing Z.L., 2004,Wavelet-based neural network for power disturbance recognition and classifica- tion, IEEE Transactions on Power Delivery, 19, 1560-1568 10. GavrovskaA.M., JevticD.R.,ReljinB.D., 2009, Selection ofwaveletdecomposition levels in ECGfiltering,Proceedings of the International Conference onTelecommunication inModern Satel- lite, Cable and Broadcasting Services TELSIKS, Nǐs-Serbia, DOI: 10.1109/TELSKS.2009.5339423, 221-224 11. JacksonJ.E.,MudholkarG.S., 1979,Controlprocedures for residuals associatedwithprincipal components analysis,Technometrics, 21, 341-349 12. Jolliffe I.T., 2002,Principal Component Analysis, Springer-Verlag, NewYork 13. Kano M., Hasebe S., 2001, A new multivariate statistical process monitoring method using principal component analysis,Computers and Chemical Engineering, 25, 1103-1113 14. Litak G., Sawicki J.T., 2009, Intermittent behaviour of a cracked rotor in the resonance region, Chaos, Solitons and Fractals, 42, 1495-1501 15. LitakG., SenA.K., SytaA., 2009, Intermittent and chaotic vibrations in a regenerative cutting process,Chaos, Solitons and Fractals, 41, 2115-2122 16. MallatS.G., 1989,A theory formultiresolution signal decomposition: thewavelet representation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 11, 7, 674-693 17. Nomikos P., MacGregor J.F., 1995,Multivariate SPC charts for monitoring batch processes, Technometrics, 37, 1, 41-59 18. Prabhakar S., Mohanty A.R., Sekhar A.S., 2002, Application of discrete wavelet transform for detection of ball bearings race faults,Tribology International, 35, 793-800 670 H. Bendjama,M.S. Boucherit 19. Purushotham V., Narayanan S., Prasad S.A.N., 2005, Multi-fault diagnosis of rolling be- aring elements using wavelet analysis and hidden Markovmodel based fault recognition,NDT&E International, 38, 654-664 20. Seker S., Ayaz E., 2002, A study on conditionmonitoring for induction motors under the acce- lerated aging processes, IEEE Power Engineering, 22, 7, 35-37 21. SenA.K.,LitakG.,FinneyC.E.A.,DawcC.S.,WagnerR.M., 2010,Analysis of heat release dynamics in an internal combustion engine using multifractals and wavelets, Applied Energy, 87, 1736-1743 22. Shao R., Hu W., Wang Y., Qi X., 2014, The fault feature extraction and classification of gear using principal component analysis and kernel principal component analysis based on the wavelet packet transform,Measurement, 54, 118-132 23. Shibata K., Takahashi A., Shirai T., 2000, Fault diagnosis of rotating machinery through visualisation of sound signal,Mechanical Systems and Signal Processing, 14, 229-241 24. TandonN.,PareyA., 2006,Conditionmonitoringof rotarymachines, [In:]ConditionMonitoring and Control for Intelligent Manufacturing, Wang L., Gao R.X. (Edit.), Springer-Verlag, London, 109-136 25. WuJ.D., Liu C.H., 2008, Investigation of engine fault diagnosis using discrete wavelet transform and neural network,Expert Systems with Applications, 35, 1200-1213 26. XuX.,XiaoF.,WangS., 2008,Enhanced chiller sensor fault detection, diagnosis and estimation using wavelet analysis and principal component analysis methods, Applied Thermal Engineering, 28, 226-237 27. Yan R., Gao R.X., Chen X., 2014, Wavelets for fault diagnosis of rotary machines: A review with applications, Signal Processing, 96, 1-15 28. Yang H., 2004, Automatic fault diagnosis of rolling element bearings using wavelet based pursuit features, Ph.D. thesis, Queensland University of Technology, Australia 29. Yaqub M.F., Gondal I., Kamruzzaman J., 2011, Resonant frequency band estimation using adaptive wavelet decomposition level selection, Proceedings of the IEEE International Conferen- ce on Mechatronics and Automation ICMA, Beijing-China, DOI: 10.1109/ICMA.2011.5985687, 376-381 Manuscript received August 13, 2014; accepted for print December 10, 2015