Research ArticleFault Diagnosis of Reciprocating Compressor Based onConvolutional Neural Networks with Multisource RawVibration Signals

Hong-bai Yang ,1,2 Jiang-an Zhang,2 Lei-lei Chen,1 Hong-li Zhang,3 and Shu-lin Liu 3

1Shanghai Open University, Shanghai 200433, China2Shanghai University of Engineering Science, Shanghai 200437, China3Shanghai University, Shanghai 200072, China

Correspondence should be addressed to Hong-bai Yang; 948645047@qq.com

Received 4 March 2019; Revised 12 May 2019; Accepted 10 June 2019; Published 4 July 2019

Academic Editor: Juan P. Amezquita-Sanchez

Copyright © 2019 Hong-bai Yang et al.This is an open access article distributed under the Creative Commons Attribution License,which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Reciprocating compressors are widely used in petroleum industry. Due to containing complex nonlinear signal, it is difficult toextract the fault features from its vibration signals. This paper proposes a new method named Convolutional Neural Networkbased on Multisource Raw vibration signals (MSRCNN). The proposed method uses multisource raw vibration signals collectedby several sensors as input and uses the designed CNN to operate both the feature extraction and classification. The gas valvesignals of reciprocating compressor in different states are used as the experimental data. In order to test the effectiveness of theproposed method, it is compared with the traditional BP (Back-Propagation) neural network fault diagnosis method based onpower spectrum energy and wavelet packet energy. In order to further test the antinoise performance of the proposed method,some noisy signals with different signal-to-noise ratios were constructed by adding white noise into sampled signals for testing.The results show that the MSRCNN model has higher fault recognition rate than the traditional methods. This indicates that theMSRCNNmethod not only has good fault recognition effect, but also has certain antinoise performance.

1. Introduction

Commonly, fault diagnosis involves two steps: feature extrac-tion and fault decision. The fault feature is extracted fromthe raw signals, and then the fault diagnosis is performedby manual experience judgment or pattern recognitiontechnology [1–13]. Common feature extraction techniquesinclude time domain statistical analysis, Fourier spectrumanalysis, wavelet transform, empirical mode decomposition(EMD), and other time-frequency analysis methods [1, 2, 5,6]. Common pattern recognition classifiers could include k-nearest neighbor (KNN) [3], artificial neural network (ANN)[4–6], and support vector machine (SVM) [4, 8–11]. Inorder to improve the fault identification rate further, themultisensor fault signals are used to diagnose the devicetogether. However, this will lead to new problems such asincreased feature set dimensions and data redundancy. Afeature selection step will be implemented to get rid of

useless and insensitive features and reduce the dimensionsfor the sake of computational efficiency. Principal componentanalysis (PCA) is one of the common dimension reductionmethods [3, 11].

As general machinery, reciprocating compressor iswidely used in petroleum and petrochemical industries. Asmall fault in reciprocating compressor may cause seriousissues in operation. Its fault diagnosis also uses traditionalmethod—feature extraction adding pattern recognition[6–12]. Liu Shulin et al. [6] used wavelet packet energy ofdifficult frequency ranges and radial basis function (RBF)neural network to diagnose faults. Tang Youfu et al. [7]calculated the Lempel-Ziv complexity indexes and usedback-propagation (BP) neural network to identify the failureof the reciprocating compressor valve. A fault diagnosismethod for bearing fault of reciprocating compressor basedon LMD multiscale entropy and SVM is proposed [8].Wang Jindong et al. [9] applied compound fault diagnosis

HindawiMathematical Problems in EngineeringVolume 2019, Article ID 6921975, 7 pageshttps://doi.org/10.1155/2019/6921975

2 Mathematical Problems in Engineering

Conv.layer(24x24) Sub-samp. layer

(12x12) MLP layers

Input layer (28x28)R,G,B Conv.layer

(8x8)Sub-samp.

layer(4x4)

Output (class vectors)

Figure 1: Overview of a sample conventional CNN.

technique based on EMD-Information Entropy and SVMto fault diagnosis of reciprocating compressor bearings.Qi et al. [10], based on the pretrained dictionary by sparsecode method, supposed that faults are recognized andclassified by support vector machine (SVM). Liu et al.[11] proposed an integrated feature extraction methodbased on the variational mode decomposition (VMD) andmultifractal detrended fluctuation analysis (MFDFA) fora fault diagnosis of a reciprocating compressor valve, andthe faults were identified by pattern classifiers of BTSVMand CNN. The Teager–Kaiser energy operation (TKEO) isproposed to estimate the signal amplitude envelopes andstatistical measures are used to extract characteristics of thevalve conditions, while the deep belief networks (DBNs)are applied to classify the faults of compressor valves [12].The above research has achieved certain results. But dueto the complex structure of the reciprocating compressor,multisource nonlinear pulse signal [14], the existing featureextraction methods are difficult to capture the essential faultfeature, which make the diagnostic effects not ideal.

After the deep learning convolutional neural network(CNN) was proposed [13], the recognition accuracy was sig-nificantly improved in many traditional pattern recognitiontasks. CNN demonstrates its superior ability to deal withcomplex problems and attracts a large number of scholarsto study its theory and application [15]. The implicit featureextraction capabilities of CNNmake itself particularly attrac-tive in the field of pattern recognition where feature analysisis difficult. Some researchers have attempted to exploitCNN for both feature extraction and pattern classificationin mechanical equipment fault diagnosis [16–27]. In theseapplications, some input of theCNNmodel is a 1-dimensional(1D) raw signal [16–19], and some input of the CNN modelis 2-dimensional (2D) data [20–27]. In [16, 17] 1D vibrationsignal is used to diagnose the bearing. In [18] 1D motorcurrent signal pair is applied to diagnose the fault of motor.In [19] a 1D vibration signal is applied to health monitoringof themechanical structure. Papers of [20–22] simply processthe raw signals into a 2D raw signal to diagnose the bearing orthe pump. In [23–28] the time-frequency analysis of the rawsignal is performed, and the 2-dimensional time-frequencyimage data is taken as the input of the CNN model.

Although the CNN method has achieved good results inmany of the above-mentioned fault diagnosis fields, it is rarelyseen in the fault diagnosis of reciprocating compressors atpresent. In this paper, a CNN model based on multisourceraw vibration signals (MSRCNN) is proposed to diagnose

the valve failure of reciprocating compressor.Themultisourceraw vibration signals are directly input in MSRCNN model.MRCNN has the following advantages: (1)The multisourceraw signals are 2D data themselves which are very suitablefor CNNs; (2) the existing feature extraction methods aredifficult to capture the essential fault characteristics of recip-rocating compressors, andMRSCNNworks on raw temporalsignals, without feature extraction or manual selection; (3)more comprehensive fault information could be obtainedthrough multiple sensors.

The remainder of this paper is organized as follows:Section 2 briefly introduces the structure and principle ofCNN; Section 3 introduces the MSRCNN model based onmultisensor vibration signals; Section 4 evaluates the effectof the proposed method through experiments by comparingit with other traditional methods. Section 5 gives someconclusions and present to the future work.

2. Overview of CNNs

CNNs are a typical deep learning model and are biologicallyinspired feed-forward ANNs that present a simple model forthe mammalian visual cortex. They are now widely used andhave become the de facto standard in many image and videorecognition systems. Figure 1 illustrates a 2-D CNN modelwith an input layer accepting 28 × 28 pixel images [13, 18, 19].Each convolution layer after the input layer alternateswith thesubsampling (pooling) layerswhich decimate propagated 2-Dmaps from the neurons of previous layer. Unlike handcraftedand fixed parameters of the 2-D filter kernels, in CNNs theyare trained (optimized) by the BP algorithm. However, thekernel size and the subsampling factor that are set to 5 and2 for illustration in Figure 1 are the two major parametersof the CNN. The input layer is only a passive layer whichaccepts an input image and assigns its (R,G,B) color channelsas the feature maps of its three neurons. With forwardpropagation over sufficient number of subsampling layers,they are decimated to a scalar (1-D) at the output of the lastsubsampling layer. The following layers are identical to thehidden layers of a MLP, fully connected, and feed-forwardnetworks. These so-called fully connected layers end up withthe output layer that estimates the decision (classification)vector.

The convolutional layer of CNN is used to extract localfeatures of the input neuron data. Each convolution layer iscomposed of multiple feature matrices. Each feature matrix

Mathematical Problems in Engineering 3

1

3

2

4

1024x3 signalframe

Convolutionallayer @ 3x3

Convolutionallayer and pooling

layer @ 3x3Fully connected layer(196736, 4128, 132)

Softmaxclassifier

Poolinglayer @ 3x3

1024

Figure 2: Architectural hierarchy of proposed 2D CNN.

is viewed as a plane. The corresponding convolution kernelsare the same on each layer. They have the characteristicsof rotation, displacement invariance, and weight sharing.The features can be learned in parallel. The number of freeparameters is greatly reduced. The convolution kernels ofdifferent planes are different, and multiconvolution kernelsmake feature extraction more complete. The convolutionoperation is described as formula (1).

𝑥𝑙𝑗 = 𝑓( ∑𝑖∈𝑀𝑗

𝑥𝑙−1𝑖 × 𝑘𝑙𝑖𝑗 + 𝑏𝑙𝑗) (1)

𝑀𝑗 is the input feature map, we use 𝐾𝑙𝑗 and 𝑏𝑙𝑗 to denote theweights and bias of the 𝑗-th filter kernel in layer 𝑙, respectively,𝑥𝑙−1𝑗 is the input of the l-th layer, and 𝑥𝑙𝑗 is the output of the𝑙-th layer.

The pooling layer (P layer), also called the downsamplinglayer,merges similar features into one and reduces the param-eters of the network. Pooling process can extract featureswith scale invariance and prevent data from overfitting. Thepolling transformation is described as follows (in formula(2)).

𝑥𝑙𝑗 = 𝑓 (𝛼𝑙𝑗𝑑𝑜𝑤𝑛 (𝑥𝑙−1𝑖 ) + 𝑏𝑙𝑗) (2)

where 𝑑𝑜𝑤𝑛(⋅) is the subsampling function and 𝛼𝑙𝑗 is the 𝑗-thnetwork multiplicative bias of the 𝑙-th layer.

The last few layers are identical to the hidden layers ofa MLP, fully connected, and feed-forward. These so-calledfully connected layers end upwith the output layer. Generally,the softmax classifier is used to solve the multiclassificationproblem.

At present, most CNN is used for a supervised mode.Thetraining of CNN is similar to the traditional artificial neuralnetworks training method. The back-propagation algorithm(BP) is used to transmit the error layer by layer, and thegradient descent method is used to adjust the parametersbetween the layers.

Convolution has three important ideas: sparse connec-tivity, parameter sharing, and equivariant representations.The sparse connectivity establishes the non-fully connectedspatial relationship between the layers through the topologystructure to reduce the number of training parameters.Weight sharing can effectively avoid the algorithmoverfitting.Equivariant means that if the input changes, the outputchanges in the same way. Pooling helps to make the represen-tation become approximately invariant to small translationsof the input. CNN is very suitable for the processing andlearning of massive data. It can effectively realize fault diag-nosis and identification by extracting features layer by layer.

3. Proposed MSRCNN Model

In this study, fusing feature extraction and fault diagnosisof the reciprocating compressor, we propose a convolutionalneural network (MSRCNN) model as shown in Figure 2,which uses the multisource raw vibration signals withoutprior feature extraction as input.

In this model, the input data is 1024∗3 (3 sensors), andthe kernel size is set to 3∗3.Themodel parameters, includingthe number of layers and whether to use BatchNomalizarion,etc., were adjusted experimentally. Model is constructedby Python (a popular development language), using theKeras framework (the back end using Tensorflow). Kerasframework can be used to build models quickly. After many

4 Mathematical Problems in Engineering

Table 1: The network structure of MSRCNN.

Layer (type) Output Shape Param #input 1 (InputLayer) (None, 1024, 3, 1) 0conv1 (Conv2D) (None, 1024, 3, 8) 80bn1 (BatchNormalization) (None, 1024, 3, 8) 32activation 1 (Activation relu) (None, 1024, 3, 8) 0max pool1 (MaxPooling2D) (None, 512, 3, 8) 0conv2 (Conv2D) (None, 512, 3, 16) 3216bn2 (BatchNormalization) (None, 512, 3, 16) 64activation 2 (Activation relu) (None, 512, 3, 16) 0max pool2 (MaxPooling2D) (None, 256, 3, 16) 0conv3 (Conv2D) (None, 256, 3, 32) 4640bn3 (BatchNormalization) (None, 256, 3, 32) 128activation 3 (Activation relu) (None, 256, 3, 32) 0max pool3 (MaxPooling2D) (None, 128, 3, 32) 0conv4 (Conv2D) (None, 128, 3, 64) 2112bn4 (BatchNormalization) (None, 128, 3, 64) 256activation 4 (Activation relu) (None, 128, 3, 64) 0conv5 (Conv2D) (None, 128, 3, 32) 18464bn5 (BatchNormalization) (None, 128, 3, 32) 128activation 5 (Activation relu) (None, 128, 3, 32) 0max pool5 (MaxPooling2D) (None, 64, 3, 32) 0conv6 (Conv2D) (None, 64, 3, 16) 4624bn6 (BatchNormalization relu) (None, 64, 3, 16) 64activation 6 (Activation relu) (None, 64, 3, 16) 0max pool6 (MaxPooling2D) (None, 32, 3, 16) 0flatten 1 (Flatten) (None, 1536) 0fc1 (Dense) (None, 128) 196736fc2 (Dense) (None, 32) 4128fc3 (Dense) (None, 4) 132

trial adjustments, the MSRCNN model for fault diagnosisof reciprocating compressor adopts the network structureshown in Table 1, including 6 2-dimensional convolution(Conv2D) layers and 6 downsamplings (Max pooling) layers,6 BatchNormalization layers, 1 flat (Flatten) layer, and 3fully connected (Dense) layers whose length is 196736, 4128,and 132, respectively. The BatchNormalization [28] layer isdesigned to reduce the shift of internal covariance and accel-erate the training process of the deep network.The activationfunction uses the “ReLU” function, which overcomes thegradient disappearance problem and greatly speeds up thetraining.The full-connected layers operate data classificationusing “softmax” function.

4. Test and Verification

In order to get the effect of the convolutional neural networkmodel on the fault diagnosis of reciprocating compressors,we design two experiments. (1)The constructed MSRCNNmodel is applied to feature extraction and fault diagnosis ofthe multisource raw vibration signal and is compared withthe commonly used power spectrum (+BP) method and thewavelet packet energy (+BP) method. (2) The robustness

of the MSRCNN model to noise is studied. The raw dataunaddednoise is used to train theMSRCNNmodel. Diagnos-tic tests were performed with different SNR data and the testresults were compared with the wavelet packet energy (+BP)method.

Using the Daubechies-db16 wavelet packet and the com-pressor vibration signal is decomposed into three levels and8 frequency bands. The BP network adopts a three-layerneural network structure. The first layer is the input layer.The second layer is the hidden layer, in which the sigmoidactivation function is selected, and the number of hiddenlayer neurons is 10. The third layer is linear output layer, andthe corresponding expected outputs are normal (0, 0, 0, 1),less springs (0, 0, 1, 0), gaps (0, 1, 0, 0), and breaks (1, 0, 0, 0).4.1. Test Data Collection. The experimental data of the recip-rocating compressor vibration was collected from the No.1 working unit in the southern section of Daqing NaturalGas Company. In order to collect the vibration data underdifferent conditions, the valves in 4 states including normalstate, valve plate breakage, spring damage, and valve notchare installed in the secondary cylinder of No.1 working unitof Daqing Natural Gas Company. The gas valve is subjected

Mathematical Problems in Engineering 5

Motor

Crankcas

The first-stagecylinder

CrosshCrossh

AccelerationTransducer

X/Y/Z

Cover

Gas valve

The Second-stagecylinder

Figure 3: The reciprocating compressor structure.

shift

first frame

second frame

Figure 4: Data segmentation with overlap.

to a fault simulation test. The secondary gas inlet pressure is310 kPa, the outlet pressure is 1,040 kPa, the inlet temperatureis 32∘C, the outlet temperature is 104∘C, the gas flow is3,611 m3/h, and three acceleration sensors (X/Y/Z) are placedon the valve cover of the second-stage cylinder, as shownin Figure 3. The data collection system uses the INV-1021program-controlled multifunction signal conditioner fromBeijing Oriental Institute and the INV306U-6660 intelligentdata collection and processing analyzer. The sampling fre-quency is 20 kHz. In the normal state 120,000 points arecollected. In the other three valve states 80,000 points arecollected for each state.

4.2. Training Sample and Test Sample Library Building. Inorder to obtain a large number of training and test samples,the data acquisition uses a partial overlap cutting method, asshown in Figure 4.The simulation experiments show that thequantity of training datawill affect the correct rate of training.

The training frames are more, and the diagnosis accuracy ishigher. In this paper the step between sample frames is 10.Each sample is 1024 long, and a total of 5382 samples arecollected. 70% of them were randomly selected as trainingsamples and 30% as test samples.

For the power spectrum energy (+BP) method and thewavelet packet energy (+BP) method, there are four faultstates: the normal state of the valve, the broken valve plate,the spring damage, and the gap of the valve plate. Eachsample length is 1024 data items, and 78 samples in each stateare, respectively, taken. The 55 samples (70%) in each stateconstitute the fault training samples (55∗4=220 total), andthe remaining 30% samples constitute the test samples of themodel (23∗4=92 total).

In order to get the fault recognition rates under differentnoise conditions, different signal-to-noise ratio Gaussiannoise is added to the test samples to get the diagnosisaccuracy.

6 Mathematical Problems in Engineering

Table 2: Comparison of fault diagnosis results of different methods.

power spectrum energy (+BP) wavelet packet energy (+BP) MSRCNNRecognition rate 60% 87.27% 100%

Table 3: Fault diagnosis results under different SNR conditions.

SNR wavelet packet energy (+BP) MSRCNN10 dB 40.91% 90.2%20 dB 75.45% 100%50 dB 87.27% 100%no noise 87.27% 100%

4.3. Comparative Analysis of Test Results. Table 2 shows thecomparison of the recognition rates of three fault diagnosismethods for the reciprocating compressor valve failure. FromTable 2, the MSRCNN method based on the raw vibrationsignal has the highest recognition rate, which can reach100% accuracy, which effectively identifies and diagnoses thereciprocating compressor valve faults.

Table 3 compares fault diagnosis results between MSR-CNNmodel and the three-layer wavelet packet energy (+BP)model under different noise intensities. Table 3 shows that thediagnostic rate of the MSRCNNmethod has a relatively highfault recognition rate under 4 kinds of test conditions. Underno-noise, the SNR of 50 dB and the SNR of 20 dB conditions,MSRCNN can all get 100% fault recognition rate. While SNRdecreases to 10 dB, this model still reaches 90.2% recognitionrate. In the contrast, the wavelet packet energy (+BP)methodcan achieve a test recognition rate of 87.27% under no-noisesignals and 50 dB (less noise), but when the signal-to-noiseratio reaches 10 dB, the recognition rate is greatly reduced(40.91%) and the faults cannot be effectively identified.

5. Conclusions

Theraw 2Dvibration signal being directly applied as an input,MSRCNN model is constructed to perform reciprocatingcompressor fault diagnosis. Experiments were carried outusing the compressor valve fault data samples obtained fromthe test bench, and MSRCNN method obtains higher faultrecognition rate than power spectrum energy (+BP) methodand the wavelet packet energy (+BP) method. The noiseadded fault signals with different SNRs were tested. Theresults showed that, for the tested compressor, MSRCNNmodel method has higher recognition rate than waveletpacket energy (+BP) method under the different signal-to-noise ratio conditions. Experiments indicate that MSRCNNnot only has higher recognition effect, but also has certainantinoise performance.

At present, the focus of the article is the application of 2Ddata on CNN. The number of constructing model layers andeach parameter are adjusted manually based on experiments.In the future, it is expected that further research can find outthe inherent principle for model optimization.

Data Availability

The raw vibration data used in this paper are got fromthe cooperation enterprises project. There is confidentialityagreement. It can only be used by our research group andcannot be open to others.

Conflicts of Interest

The authors declare that they have no conflicts of interest.

Acknowledgments

This work was supported by the 2018 Shanghai Open Uni-versity Research Project (KX1805), Shanghai Key Laboratoryof Intelligent Manufacturing and Robotics (ZK1801), andNational Natural Science Foundation of China (51575331).

References

[1] Z. P. Feng, M. Liang, and F. L. Chu, “Recent advances intime-frequency analysismethods formachinery fault diagnosis:a review with application examples,” Mechanical Systems andSignal Processing, vol. 38, no. 1, pp. 165–205, 2013.

[2] H. Y. Zhao, J. D. Wang, J. Lee, and L. Ying, “A compound inter-polation envelope localmean decomposition and its applicationfor fault diagnosis of reciprocating compressors,” MechanicalSystems and Signal Processing, vol. 110, pp. 273–295, 2018.

[3] D. H. Pandya, S. H. Upadhyay, and S. P. Harsha, “Fault diagnosisof rolling element bearing with intrinsic mode function ofacoustic emission data using APF-KNN,” Expert Systems withApplications, vol. 40, no. 10, pp. 4137–4145, 2013.

[4] A. Hajnayeb, A. Ghasemloonia, S. Khadem, and M. Moradi,“Application and comparison of an ANN-based feature selec-tion method and the genetic algorithm in gearbox fault diagno-sis,” Expert Systems with Applications, vol. 38, no. 8, pp. 10205–10209, 2011.

[5] P. Potocnik and E. Govekar, “Semi-supervised vibration-based classification and condition monitoring of compressors,”Mechanical Systems and Signal Processing, vol. 93, pp. 51–65,2017.

[6] S. L. Liu, J. Z. Zhang,M. Q. Xu et al., “Wavelet packet and neuralnetwork based approach for fault diagnosis of reciprocatingcompressors,” Oil Field Equipment, vol. 31, no. 5, pp. 1–3, 2002.

[7] Y. F. Tang, S. L. Liu, Y. H. Liu et al., “Fault diagnosis basedon nonlinear complexitymeasurefor reciprocating compressor,”Journal of Mechanical Engineering, vol. 48, no. 03, pp. 102–107,2012.

[8] S. S. Yang, H. Zhou, H. Y. Zhao et al., “Fault diagnosismethod for the bearing of reciprocating compressor basedon LMD multiscale entropy and SVM,” Journal of MechanicalTransmission, vol. 39, no. 2, pp. 119–123, 2015.

[9] J. D. Wang, M. Dai, F. F. Xia et al., “Fault diagnosis forreciprocating compressor bearings based on EMD-information

Mathematical Problems in Engineering 7

entropy and SVM,” Fluid Machinery, vol. 42, no. 7, pp. 43–46,2014.

[10] G. Qi, Z. Zhu, K. Erqinhu, Y. Chen, Y. Chai, and J. Sun, “Fault-diagnosis for reciprocating compressors using big data andmachine learning,” Simulation Modelling Practice and Theory,vol. 80, pp. 104–127, 2018.

[11] Y. Liu, J. Wang, Y. Li, H. Zhao, and S. Chen, “Feature extractionmethod based on VMD and MFDFA for fault diagnosis ofreciprocating compressor valve,” Journal of Vibroengineering,vol. 19, no. 8, pp. 6007–6020, 2017.

[12] V. T. Tran, F. Althobiani, and A. Ball, “An approach to fault diag-nosis of reciprocating compressor valves using Teager–Kaiserenergy operator and deep belief networks,” Expert Systems withApplications, vol. 41, no. 9, pp. 4113–4122, 2014.

[13] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNetclassification with deep convolutional neural networks,” inProceedings of the Conference on Neural Information ProcessingSystems, vol. 60, pp. 1097–1105, Lake Tahoe, Nev, USA, Decem-ber 2012.

[14] S. Xiao, S. Liu, F. Jiang, M. Song, and S. Cheng, “Nonlineardynamic response of reciprocating compressor system withrub-impact fault caused by subsidence,” Journal of Vibration andControl, vol. 25, no. 11, pp. 1737–1751, 2019.

[15] H. Ren, J. F. Qu, Y. Chai, Q. Tang, and X. Ye, “Deep learning forfault diagnosis: the state of the art and challenge,” Control andDecision, vol. 32, no. 8, pp. 1345–1358, 2017.

[16] L. Eren, “Bearing fault detection by one-dimensional convolu-tional neural networks,”Mathematical Problems in Engineering,vol. 2017, Article ID 8617315, 9 pages, 2017.

[17] W. Zhang, G. Peng, C. Li, Y. Chen, and Z. Zhang, “A new deeplearning model for fault diagnosis with good anti-noise anddomain adaptation ability on raw vibration signals,” Sensors, vol.17, no. 2, 2017.

[18] T. Ince, S. Kiranyaz, L. Eren, M. Askar, and M. Gabbouj,“Real-time motor fault detection by 1-D convolutional neuralnetworks,” IEEE Transactions on Industrial Electronics, vol. 63,no. 11, pp. 7067–7075, 2016.

[19] O. Abdeljaber, O. Avci, S. Kiranyaz, M. Gabbouj, and D. J.Inman, “Real-time vibration-based structural damage detectionusing one-dimensional convolutional neural networks,” Journalof Sound and Vibration, vol. 388, pp. 154–170, 2017.

[20] W. Zhang, G. Peng, and C. Li, “Bearings fault diagnosis basedon convolutional neural networks with 2-D representation ofvibration signals as input,” in Proceedings of the 3rd Interna-tional Conference on Mechatronics and Mechanical Engineering(ICMME ’16), vol. 95, pp. 1–5, China, October 2016.

[21] X. Guo, L. Chen, and C. Shen, “Hierarchical adaptive deepconvolution neural network and its application to bearing faultdiagnosis,”Measurement, vol. 93, pp. 490–502, 2016.

[22] L. Wen, X. Li, L. Gao, and Y. Zhang, “A new convolutionalneural network-based data-driven fault diagnosis method,”IEEE Transactions on Industrial Electronics, vol. 65, no. 7, pp.5990–5998, 2018.

[23] O. Janssens, V. Slavkovikj, B. Vervisch et al., “Convolutionalneural network based fault detection for rotating machinery,”Journal of Sound and Vibration, vol. 377, pp. 331–345, 2016.

[24] S. Li, G. Liu, X. Tang, J. Lu, and J. Hu, “An ensemble deep con-volutional neural network model with improved D-S evidencefusion for bearing fault diagnosis,” Sensors, vol. 17, no. 8, ArticleID 1729, 2017.

[25] D. Verstraete, A. Ferrada, E. L. Droguett, V. Meruane, and M.Modarres, “Deep learning enabled fault diagnosis using time-frequency image analysis of rolling element bearings,” Shockand Vibration, vol. 2017, Article ID 5067651, 17 pages, 2017.

[26] M. Bach-Andersen, B. Rømer-Odgaard, and O.Winther, “Deeplearning for automated drivetrain fault detection,”Wind Energy,vol. 21, no. 1, pp. 29–41, 2018.

[27] Z. Chen, C. Li, and R.-V. Sanchez, “Gearbox fault identificationand classification with convolutional neural networks,” Shockand Vibration, vol. 2015, Article ID 390134, 10 pages, 2015.

[28] S. Ioffe and C. Szegedy, “Batch normalization: Acceleratingdeep network training by reducing internal covariate shift,” inProceedings of the 32nd International Conference on MachineLearning (ICML ’15), pp. 448–456, July 2015.

Hindawiwww.hindawi.com Volume 2018

MathematicsJournal of

Hindawiwww.hindawi.com Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwww.hindawi.com Volume 2018

Probability and StatisticsHindawiwww.hindawi.com Volume 2018

Journal of

Hindawiwww.hindawi.com Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwww.hindawi.com Volume 2018

OptimizationJournal of

Hindawiwww.hindawi.com Volume 2018

Hindawiwww.hindawi.com Volume 2018

Engineering Mathematics

International Journal of

Hindawiwww.hindawi.com Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwww.hindawi.com Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwww.hindawi.com Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwww.hindawi.com Volume 2018

Hindawi Publishing Corporation http://www.hindawi.com Volume 2013Hindawiwww.hindawi.com

The Scientific World Journal

Volume 2018

Hindawiwww.hindawi.com Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwww.hindawi.com Volume 2018

Hindawiwww.hindawi.com

Di�erential EquationsInternational Journal of

Volume 2018

Hindawiwww.hindawi.com Volume 2018

Decision SciencesAdvances in

Hindawiwww.hindawi.com Volume 2018

AnalysisInternational Journal of

Hindawiwww.hindawi.com Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwww.hindawi.com