Received July 1, 2020, accepted July 19, 2020, date of publication July 27, 2020, date of current version

August 6, 2020.
Digital Object Identifier 10.1109/ACCESS.2020.3012053

Rolling Bearing Fault Diagnosis Based

on Convolutional Neural Network
and Support Vector Machine
LAOHU YUAN 1 , DONGSHAN LIAN 1, XUE KANG 1, YUANQIANG CHEN 1,

AND KEJIA ZHAI 1

College of Aerospace Engineering, Shenyang Aerospace University, Shenyang 110136, China
Corresponding author: Laohu Yuan ([email protected])
This work was supported in part by the National Natural Science Foundation.

ABSTRACT Rolling bearings are one of the essential components in rotating machinery. Efficient bearing
fault diagnosis is necessary to ensure the regular operation of the mechanical system. Traditional fault
diagnosis methods usually rely on a complex artificial feature extraction process, which requires a lot of
human expertise. Emerging deep learning methods can reduce the dependence of the feature extraction
process on manual intervention effectively. However, its training requires a large number of fault signals,
which is difficult to obtain in actual engineering. In this paper, a rolling bearing fault diagnosis method based
on Convolutional Neural Network and Support Vector Machine is proposed to solve the above problems.
Firstly, the Continuous Wavelet Transform is used to convert one-dimensional original vibration signals
into two-dimensional time-frequency images. Secondly, the obtained time-frequency images are input for
training the constructed model. Finally, the diagnosis of the fault location and severity is completed. The
method is verified on the CWRU data set and the MFPT data set. The results demonstrate that the proposed
method achieves higher diagnostic accuracy and stability than other advanced techniques.

INDEX TERMS Convolutional neural network, continuous wavelet transform, fault diagnosis, rolling
bearing, support vector machine.

I. INTRODUCTION processing, which can analyze the time domain and fre-
Since some industrial machines need to work continuously in quency domain as a whole. Commonly used time-frequency
harsh environments, failures of critical components such as analysis methods include Empirical Mode Decomposition
bearings often occur. As one of the basic elements of many (EMD) [4], Short-Time Fourier Transform (STFT) [5], and
industrial machinery, the working state of rolling bearings has Wavelet Transform (WT) [6]. The signal can be adaptively
a great influence on the operation of the entire equipment [1]. decomposed by EMD into intrinsic modal function compo-
Therefore, the research on the fault diagnosis technology of nents with different scales. However, there is a problem of
rolling bearings is significant for the safety of the production modal confusion in this process. Although STFT has the
process and the reduction of economic losses. With the devel- ability to realize the time-frequency analysis of the signal,
opment of machine learning, many typical intelligent meth- it cannot adequately reflect the sudden change of the vibra-
ods have been successfully used in fault diagnosis research, tion signal because its time resolution is fixed. WT is a
mainly including of two stages: signal feature extraction and time-frequency analysis technology, of which time window
fault classification [2]. can shrink as the frequency of the signal increases, and vice
The vibration signals of bearings usually contain sufficient versa. WT expands STFT and effectively makes up for its
fault information, but most of them are nonlinear and non- shortcomings, so it is widely used. Yan et al. have sum-
stationary. Therefore, signal feature extraction is a crucial marized the application of Continuous Wavelet Transform
step [3]. Time-frequency analysis is a powerful tool in signal (CWT), Discrete Wavelet Transform (DWT), Wavelet Packet
Transform (WPT), and Second-Generation Wavelet Trans-
The associate editor coordinating the review of this manuscript and form (SGWT) in the field of fault diagnosis [7]. On the
approving it for publication was Yan-Jun Liu. other hand, fault identification also plays a vital role in

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 8, 2020 137395
 L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

fault diagnosis because it cannot meet the requirements of This paper is organized as follows: Section 2 is dedicated
data batch processing only by fault extraction. Traditional to the theory of CWT, CNN, and SVM. In Section 3, the
fault recognition tools include Bayesian classifier [8], Arti- proposed CNN-SVM model is presented. In Section 4, the
ficial Neural Network (ANN) [9]–[10], and Support Vector complete experimental procedure and analysis of the results
Machine (SVM). Under the condition that there are enough are introduced. Finally, the conclusion is given in Section 5.
training samples, the first two can distinguish the fault types
effectively. However, a large number of available failure sam- II. THEORETICAL BACKGROUND
ples are hard to obtain in actual work. SVM can achieve effec- In this paper, an intelligent diagnosis method of rolling bear-
tive classification through a small amount of samples due ing faults is proposed. Firstly, we convert the original vibra-
to its strong network generalization ability, good generality, tion signals into time-frequency images using CWT. Then,
and high classification accuracy. Therefore, SVM has been CNN is applied to extract the in-depth features of the time-
widely used in mechanical fault diagnosis research [11]–[14]. frequency images. Finally, the classifier SVM is trained using
However, SVM performs poorly on redundant data because the extracted features. The fundamental theories of CWT,
its shallow structure has some difficulties in learning in-depth CNN, and SVM are introduced as follows.
features [15].
Recently, deep learning has become a vital research direc- A. CWT FOR TIME-FREQUENCY ANALYSIS
tion and gradually applied to various fields [16]–[19]. The The CWT time-frequency analysis method performs multi-
deep learning model is composed of multi-layer neural net- scale refinement on the signal through scaling and transla-
works, which can extract and learn the in-depth features tion operations. Therefore, CWT can automatically adapt to
of the input signal. Deep learning method can handle the the requirements of time-frequency signal analysis, clearly
complex and high-dimensional problems in the massive data describing the change of signal frequency with time. [7].
that cannot be solved by shallow learning [20]. Due to its high Here, CWT is used for preliminary feature extraction,
efficiency, plasticity and universality, scholars have applied converting the original 1-D time-domain signals into 2-D
many deep learning models to the research of fault diagnosis, time-frequency images. The conversion process is shown in
such as Long Short-Term Memory (LSTM) [21], Deep Belief Fig. 1 [31].
Network (DBN) [22], Deep Auto-encoder (DAE) [23], Gated
Recurrent Unit Network (GRUN) [24], and Convolutional 1) CONTINUOUS WAVELET TRANSFORM
Neural Network (CNN). Among them, CNN has more sophis- CWT is a method to obtain characteristic signal information,
ticated applications in image processing, including image which can be used for the processing and analysis of nonlin-
classification [25], target positioning [26], and face recog- ear signals. Its algorithm is relatively mature, and the basic
nition [27]. It has received more attention in the study of definition can be expressed as [32]:
rolling bearing fault diagnosis. In [28], a deep CNN struc- 1
Z
t −b
tural model that can automatically classify rolling bearing Wϕ (a, b) = √ x(t)ϕ ∗ ( )dt, a > 0 (1)
a a
faults is established. In [29], a method for rolling bearing
fault diagnosis based on Cyclic Spectral Coherence (CSCoh) where a represents the scale parameter, b represents the time
and CNN is proposed, which improves the fault recognition or translation parameter, x(t) represents the original one-
performance. To simplify the network architecture, Hierar- dimensional data signal, ϕ represents the wavelet function
chical Symbolic Analysis (HSA) and CNN are combined for with scale a and position offset b, and ϕ ∗ is the complex
bearing fault diagnosis in [30]. However, the above method conjugate of ϕ.
relies on large quantities of samples that can be used for train-
2) THE OPTIMAL WAVELET BASIS
ing, which is difficult to obtain in actual engineering. More-
over, the construction time of the deep network is relatively An optimal wavelet basis function (WBF) is essential in any
long. signal processing using CWT. For the same signal, we will
Considering the advantages and disadvantages of all the get different results by using different WBF analysis. WBF
above work, CNN and SVM are combined to build a deep should be selected based on two aspects, including the general
neural network framework CNN-SVM to diagnose bearing properties of wavelet and the characteristics of the analyzed
faults in this paper. In the first place, CWT is used for prelimi- object. Its general principles include orthogonality, tight sup-
nary feature extraction. Then, the CNN-SVM network model port, symmetry, and smoothness. When we use CWT for
is constructed using the method of transfer learning. In the pulse signal processing, the closer the shape of WBF is to
end, the preliminary extracted features are used to train the the signal waveform, the more features can be obtained. The
network model to achieve bearing fault classification. This formula of the similarity between the two is as follows [33]:
method makes full use of the excellent feature extraction k
X m2i
capability of CNN and the exceptional classification perfor- δ= αi (2)
si
mance of SVM, which solves the problem of in-depth signal i=1
feature extraction and the difficulty of obtaining massive where δ represents the similarity coefficient, mi represents
samples in practical work. the maximum value of each peak after WBF is taken as the

137396 VOLUME 8, 2020

L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

FIGURE 1. The conversion process from time-domain signal to time-frequency image.

FIGURE 2. The architecture of the CNN model.

absolute value, si represents the area covered by each peak where xjl is the output of layer l, xil−1 represents the output of
after WBF is taken as the absolute value, αi represents the layer l − 1, that is, the input of layer l, Mj is the feature set of
weighted coefficient of each peak after WBF is taken as the layer l − 1, kijl represents the weight matrix, blj represents the
absolute value, αi = max(m
mi
i)
, and k is the number of peaks network bias, and f (·) represents the activation function.
after WBF is taken as the absolute value.
2) POOLING LAYER
B. CONVOLUTIONAL NEURAL NETWORK In the pooling layer, the data is down-sampled by calculat-
As a common method for extracting data features in deep ing the local average or maximum value, which reduces the
learning models, CNN has made remarkable achievements network calculation complexity and retains the most critical
in image recognition research [34]. The overall structure of features, thereby improving the efficiency of feature extrac-
CNN is demonstrated as Fig. 2. Its internal hidden layer tion. The calculation method can be expressed as:
structure is mainly composed of convolution layer, pooling    
layer, and fully-connected layer. xjl = f βjl down xjl−1 + blj (4)

1) CONVOLUTION LAYER where down (·) is the down-sampling function, β represents

In the convolution layer, a more advanced feature representa- the weight of the network.
tion is obtained through convolution operations. The number
of network parameters is decreased and the complexity of 3) FULLY-CONNECTED LAYER
the model is reduced by performing local feature extraction The image features are input into the fully connected layer
on input data. The convolution formula can be defined as after being alternately transferred through convolutional lay-
follows: ers and pooling layers. In the fully-connected layer, the deep
 
X feature information with category distinction is integrated,
xjl = f  xil−1 kijl + blj  (3) and the mapping relationship between the extracted features
i∈Mj and sample types is constructed. The mathematical formula

VOLUME 8, 2020 137397

 L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

TABLE 1. Architecture and parameters of ResNet-18.

FIGURE 3. Schematic diagram of the SVM classification principle.

To solve the above problem, the Lagrange function is

can be expressed as: introduced:
n n
  1 X X
yk = f ωk x k−1 + bk (5) L (ω, b, α) = kωk2 − αi yi (ω · xi + b) + αi (9)
2
i=1 i=1
where k represents the network of layer k, yk
represents where αi represents the Lagrange multiplier, αi ≥ 0. The total
the output of the fully connected layer, x k−1 is the input of decision function can be expressed as:
the fully connected layer, ωk is the weight coefficient, bk n
!
represents the network offset. f (x) = sgn
X
αi yi (xi · x) + b (10)
i=1
C. SUPPORT VECTOR MACHINE
However, the samples in the original space are often non-
SVM is a commonly used classifier that shows excellent
linear and inseparable. Therefore, the introduction of an
advantages in resolving small sample, nonlinear, and high-
appropriate kernel function is necessary. Finally, the basic
dimensional pattern recognition problems. As indicated in
model of SVM can be obtained:
Fig. 3, the basic model of SVM is to find an optimal hyper-
n
!
plane on the feature space to divide the dataset. In Fig. 3, H is X
the classification surface, H1 and H2 are straight lines parallel f (x) = sgn αi yi K (xi , x) + b (11)
i=1
to H and passing the data points closest to it in the two types of
data. The data point is called support vector, and the distance where K represents the kernel function. The commonly used
between H1 or H2 and H is called the geometric separation. kernel functions include linear, polynomial, S-shaped, and
Therefore, maximizing the geometric interval becomes the Gaussian radial basis kernel function. In this paper, the Gaus-
training goal of SVM [35]. sian radial basis function is selected.
Suppose a linearly categorizable training sample dataset,
D = {(xi , yi ) , i = 1, . . . , n}, where xi represents III. THE PROPOSED CNN-SVM MODEL
K-dimensional column vector, that is, x ∈ RK×1 , yi is the In this paper, a network architecture based on CNN and
category of xi , and yi ∈ {−1, 1}. The total decision function SVM is established for rolling bearing fault diagnosis, which
is as follows: we call CNN-SVM model. This chapter mainly includes the
construction of CNN-SVM model and bearing fault diagnosis
yi = sgn ((ω · x) + b) (6) based on the proposed model.

where sgn represents a symbolic function, ω represents the A. DESIGN OF CNN-SVM

weight vector, and b is the deviation. In this section, the CNN-SVM model is constructed, of which
Therefore, the problem of maximizing the geometric inter- structure is indicated as Fig. 4. The model uses transfer
val can be transformed into the solving of the quadratic learning of the pre-trained ResNet-18 network to extract
programming problem: 2-D image features, and uses the extracted features to train

1
 the classifier SVM, so as to achieve the purpose of image
2
min = kωk (7) classification.
2
In this system, the method of transfer learning in deep
the constraint is: learning is used, and the ResNet-18 network is selected for
feature extraction. Its structure and parameters are shown
yi (ω · xi + b) ≥ 1, i = 1, . . . , n. (8) in Table 1 [36]. Applying the relevant knowledge that has

137398 VOLUME 8, 2020

L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

FIGURE 4. The architecture of the proposed CNN-SVM model.

proposed method is demonstrated as Fig. 6, and the steps are

as follows:
Step 1: The vibration signals of faulty bearings are col-
lected on two different experimental platforms.
Step 2: First, the vibration signals are split into segments.
Then, we use CWT to convert the collected vibration signals
into time-frequency images with a size of 224 × 224 ×3.
Finally, the time-frequency images are partitioned into train-
ing samples and test samples, and a label is set for each
sample.
Step 3: First, load the pretrained ResNet-18 network. Then,
FIGURE 5. The residual structure.
input the samples processed in step 2 to the model and
get the high-level feature representations of the training and
test images from the fully connected layer. Finally, use the
features extracted from the training samples as predictor
been learned in the pre-training network directly to the tar-
variables to train the SVM.
get field can effectively solve the problem of complicated
Step 4: Test samples are input to the trained model to verify
and time-consuming manual construction of CNN and the
the validity of the diagnosis.
problem of insufficient data samples obtained in actual engi-
neering. The ResNet-18 network adopts the idea of residual IV. EXPERIMENTS
learning to guard against the problem of gradient disappear- To verify the flexibility and utility of the constructed model
ance to a certain extent. A typical residual structure is shown in bearing fault diagnosis, two open-source datasets are used
in Fig. 5. for research in this paper, including Case Western Reserve
SVM classifies different states by finding the best hyper- University (CWRU) Bearing Data Center dataset [37] and
plane that separates different data elements, and possesses the Society for Mechanical Failure Prevention Technology
exceptional generalization ability and high classification (MFPT) dataset [38].
accuracy. In this study, the Softmax layer in the CNN model
is replaced by an SVM classifier. After the CNN is finished A. CWRU DATA
trained, the extracted image features that output from the
1) DATA DESCRIPTION
fully-connected layer are used to train the SVM classifier.
The CWRU data set is one of the most widely used datasets
Finally, the trained SVM-CNN model can be used for image
in bearing fault diagnosis. As illustrated in Fig. 7, the experi-
type recognition.
ment platform consists of four parts: a driven motor, a torque
sensor, a dynamometer and an electronic controller. SKF
B. THE PROPOSED METHOD FOR BEARING FAULT 6205-2RS JEM deep groove ball bearings are used in the
DIAGNOSIS experiment. The sampling frequency of the drive end bearing
Since the amount of available bearing fault data in practical is 12 kHz and 48 kHz, and the sampling frequency of the
engineering is small and the feature extraction of traditional fan end bearing is 12 kHz. The single-point damage method
intelligent diagnosis methods is more complicated, a method of EDM is used to set the damage diameters of 0.1778,
for fault detection of rolling bearings based on CNN-SVM is 0.3556, 0.5334, and 0.7112 mm on the inner ring and rolling
proposes in this study. This method can take advantage of the element respectively, and the damage diameters of the outer
superiority of CNN in data feature obtaining and the classi- ring at 6 o’clock are set to 0.1778, 0.3556, and 0.5334 mm.
fication and generalization capabilities of SVM to improve The vibration data of bearings in normal status and 11 types
diagnostic performance. The corresponding flow chart of the of fault status are collected.

VOLUME 8, 2020 137399

 L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

FIGURE 6. The flowchart of the experiment.

In this study, the signals collected at the frequency of is selected. In this way, different fault locations and damage
12kHz under 12 health conditions of the drive end bearing severities of bearings can be better simulated.

137400 VOLUME 8, 2020

L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

TABLE 2. Sample description of the CWRU dataset.

FIGURE 7. CWRU bearing test rig.

TABLE 3. Similarity coefficients of five commonly used wavelet basis

functions.

FIGURE 8. Signal segmentation.

2) DATA PROCESSING 3) RESULTS OF THE EXPERIMENT

In this experiment, 120000 data points are taken from the First, the proposed CNN-SVM model is trained by the con-
data set under four load conditions of 12 health states of structed training samples of the 12 bearing fault states. After
the bearing, respectively. In order to make each sample the training is completed, the remaining test samples are used
contain enough information, 12000 data points are taken to verify the trained model. To prove the stability of the
as the sample length. As shown in Fig. 8, 6000 data proposed model, the experiment is repeated ten times. The
points are repeated between adjacent parts for a division diagnosis accuracy and training time of 10 trials are revealed
to obtain more experimental samples. In this way, each in Table 4. The average classification accuracy is 98.75%, and
group of fault data can be divided into 19 segments. There- the highest diagnosis accuracy is 99.29%. A fault diagnosis
fore, 19 time-frequency images in 48 states can be obtained result is shown in Fig. 10.
through CWT. By merging the load types under 12 health con-
ditions, 76 time-frequency images for each health state are B. MFPT DATA
obtained. Finally, a sample set containing 912 time-frequency 1) DATA DESCRIPTION
images is obtained, as shown in Table 2. Among them, 70% The MFPT dataset provided by the Society for Machinery
of the images under each fault type are randomly chosen as Failure Prevention Technology is also applied for the analysis
training samples, and the remaining 30% of the images are and research of rolling bearing faults. The MFPT dataset con-
used as test samples. In Fig. 9 (a)-(l), the time-frequency sists of three sets of experimental bearing vibration data and
images of the 12 health conditions in Table 2 are shown in three actual fault data. Among them, the three sets of experi-
order. mental bearing vibration data include baseline set, inner race
Db10, Meyer, Sym8, Coif5, Morlet are several commonly fault set, and outer race fault set. The baseline dataset contains
used wavelet basis functions, and their similarity coefficients three files, and the data in each file are obtained by sampling
to bearing fault characteristics are listed in Table 3 [33]. at a frequency of 97656 Hz for 6 seconds under a load of
According to the Table 3, the similarity coefficient of db10 270 pounds. The inner race fault dataset contains seven files,
wavelet is the largest. However, Db10 wavelet has an asym- which are respectively obtained by sampling at 48828 Hz for
metric shape, poor smoothness, and complicated theoretical 3 seconds under seven load conditions, including 0, 50, 100,
analysis. Morlet wavelet is a single-frequency sine function 150, 200, 250, and 300 pounds. The outer race fault dataset
under Gaussian envelope, of which the similarity coefficient contains seven files, which are respectively obtained by sam-
is second only to Db10 and shape is symmetrical and smooth. pling at 48828 Hz for 3 seconds under seven load conditions,
Therefore, Morlet wavelet is chosen here as the WBF of CWT including 25, 50, 100, 150, 200, 250, and 300 pounds. The
for the generation of time-frequency images of vibration data points are from a single channel radial accelerometer.
signals. The experiment bearings of the MFPT data set is also deep

VOLUME 8, 2020 137401

 L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

FIGURE 9. Time-frequency images of twelve health conditions on CWRU dataset: (a) Class 1; (b) Class 2; (c) Class 3; (d) Class 4; (e) Class 5; (f) Class 6;
(g) Class 7; (h) Class 8; (i) Class 9; (j) Class 10; (k) Class 11; (l) Class 12.

TABLE 4. Diagnostic results of 10 trials on CWRU dataset.

groove ball bearing with a 31.62 mm pitch diameter, a 5.97

mm ball diameter, a 0◦ contact angle, and an element number
of 8.
In this study, the three sets of bearing vibration data col-
lected at the test rig are selected.

2) DATA PROCESSING
In this experiment, all the data points in the three fault data
sets are used, and 48828 data points are taken as the sample
length. In order to obtain more experimental samples, 24414
data points are repeated between adjacent parts. The baseline
set is down-sampled to 48828 Hz to match other fault sets.
The three files in the baseline set are divided into 22 segments
respectively, the seven inner fault files are divided into five FIGURE 10. Classification results on CWRU dataset.

segments respectively, and the seven outer fault files are

also divided into five segments respectively. Through the under the three fault types, 66 time-frequency images for
CWT, a corresponding number of time-frequency images in the baseline set, 35 time-frequency images for the inner race
each state can be obtained. By merging the load conditions and 35 time-frequency images for the outer race are obtained.

137402 VOLUME 8, 2020

L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

FIGURE 11. Time-frequency images of three health conditions on MFPT dataset: (a) Class 1; (b) Class 2; (c) Class 3.

TABLE 5. Sample description of the MFPT dataset.

Finally, a sample set containing 136 time-frequency images

is obtained, as shown in Table 5. Among them, 70% of the
images in each state are randomly chosen as training samples,
and the remaining 30% of the images are used as test samples. FIGURE 12. Classification results on MFPT dataset.

In Fig. 11 (a)-(c), the time-frequency images of the three

health conditions in Table 5 are shown in sequence.

3) RESULTS OF THE EXPERIMENT

First, the CNN-SVM model is trained by using the con-
structed training samples of the three bearing fault types.
After the training is completed, the trained model is verified
by the test sample. To prove the stability of the proposed
model, the experiment is repeated ten times. The diagnosis
accuracy and training time of 10 trials are revealed in Table 6.
The average classification accuracy is 98.89%, and the high-
est diagnosis accuracy reaches 100%. A fault diagnosis result
is indicated in Fig. 12.

C. COMPARISON WITH OTHER METHODS

FIGURE 13. Classification accuracy of three different methods.
To prove the superiority of the proposed model, we compare it
with standard CNN and standard SVM models. The inputs of
standard CNN and SVM are also the time-frequency images, using the CWRU dataset similarly. The diagnosis accuracy of
and the classifier of the standard CNN is Softmax. In order to each method is shown in Fig. 13. The diagnosis accuracy and
ensure the fairness of the experiment and the reliability of the average training time of the three methods are summarized in
results, the test is repeated five times on these two models by Table 7.

TABLE 6. Diagnostic results of 10 trials on MFPT dataset.

VOLUME 8, 2020 137403

 L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

TABLE 7. Accuracy and average training time of the three models. under a motor load of 0hp, and then complete classification
for the 10 health conditions under the working situation with
a motor load of 1, 2, and 3hp.
The architecture and parameters of ResNet-18 which
selected by the proposed method for feature extraction are
given in Table 1. The selection of other hyperparameters is as
follows. The mini batch size is 256. The learning rate starts
from 0.1 and is divided by 10 when the error reaches a plateau,
TABLE 8. Average accuracy of the proposed method and several other and the models are trained for up to 60 × 104 iterations, the
methods.
weight decay is 0.0001, and the momentum equals to 0.9
[36]. The accuracies of the 5 contrastive methods are 97.82%,
77.86%, 98.125%, 94.73%, and 97.81%, respectively.
By comparing the results with other methods, it can be
easily seen that the proposed method achieves a higher diag-
nosis accuracy, which further shows the effectiveness of the
proposed method.

V. CONCLUSION
In this study, a new deep neural network model CNN-SVM
As can be seen, the diagnosis accuracy of the proposed is built for fault diagnosis of rolling bearings. First, we use
method in 5 trials is higher than that of the other two standard the CWT to construct the time-frequency images of vibration
models, and the training time is considerably shortened com- signals. Then, the obtained images are input to the proposed
pared to the standard CNN model. The results show that SVM model for training. Finally, the diagnosis of fault location and
is better than the default classifier, i.e., Softmax. It is proved severity of the rolling bearing is completed. The experiments
that the proposed model can effectively solve the difficulty of indicate that the diagnostic accuracy of this method can reach
extracting in-depth features of miscellaneous data by SVM 98.75% for the CWRU dataset and 98.89% for the MFPT
and the challenge of meeting the needs of massive samples dataset which verify the flexibility and practicability of the
for the training of CNN, and has excellent stability. constructed model. By comparing with the standard CNN
In recent years, there are many works about deep learning- and standard SVM, it is showed that the proposed model is
based fault diagnosis methods. Thus, to further illustrate the able to resolve the difficulty of deep feature extraction in the
innovation of the proposed method in the field of bearing fault traditional method and the small sample problem in actual
diagnosis, we compare it with existing advanced techniques engineering and has excellent stability. By comparing with
such as PNN-SFAM [9], BPNN [39], CNN-HMM [39], advanced methods such as PNN-SFAM, BPNN, CNN-HMM,
DAFD [40], and DGNN [41]. The average diagnosis accuracy DAFD, and DGNN, the effectiveness of the proposed method
of each method is listed in Table 8. is further verified.
The details of the different methods are listed below. In [9], However, for relatively noisy data sources, the accuracy of
the PNN is introduced to classify rolling bearing states in the proposed method still needs to be enhanced. Therefore,
healthy and not healthy. If the PNN decision shows that the the structure of the proposed model needs further improve-
processed state is not healthy, then the SFAM is used to clas- ment in the future. Moreover, in this paper, only single-fault
sify 7 types of faults. In [39], for the BPNN method, 50 units bearing vibration signals are used for model training, and no
in hidden layer, learning rate is 0.2, momentum is 0.05, and compound fault samples are created to simulate the actual
iteration number is 800. 12 types of health conditions are clas- situation, which is a certain challenge for the application
sified by using the learned features of the experimental data of the proposed model in practical engineering. It has also
as inputs. For the CNN-HMM method, a 50 × 50 matrix of become a research direction for us in the future.
each sample is constructed, learning rate is 1, and maximum
REFERENCES
iteration is 100. 12 health conditions are classified by using
[1] D.-T. Hoang and H.-J. Kang, ‘‘A survey on deep learning based bearing
raw data without any signal preprocessing as inputs. In [40], fault diagnosis,’’ Neurocomputing, vol. 335, pp. 327–335, Mar. 2019.
a deep neural network model with domain adaptability, called [2] M. Kang, M. R. Islam, J. Kim, J.-M. Kim, and M. Pecht, ‘‘A hybrid feature
DAFD, has been proposed for fault diagnosis. The classifica- selection scheme for reducing diagnostic performance deterioration caused
by outliers in data-driven diagnostics,’’ IEEE Trans. Ind. Electron., vol. 63,
tion model is directly trained by using the labeled samples of
no. 5, pp. 3299–3310, May 2016.
4 health conditions collected without a motor load, and then [3] Q. Hu, X.-S. Si, A.-S. Qin, Y.-R. Lv, and Q.-H. Zhang, ‘‘Machinery fault
the data collected under a motor load of 3hp is classified. diagnosis scheme using redefined dimensionless indicators and mRMR
In [41], a fault diagnosis method based on deep generative feature selection,’’ IEEE Access, vol. 8, pp. 40313–40326, 2020.
[4] X. Yu, F. Dong, E. Ding, S. Wu, and C. Fan, ‘‘Rolling bearing fault diag-
neural networks (DGNN) has been proposed. The model is nosis using modified LFDA and EMD with sensitive feature selection,’’
trained by using the samples of 10 health conditions collected IEEE Access, vol. 6, pp. 3715–3730, 2018.

137404 VOLUME 8, 2020

L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

[5] D. Zhong, W. Guo, and D. He, ‘‘An intelligent fault diagnosis method based [27] S. Banerjee and S. Das, ‘‘Mutual variation of information on transfer-CNN
on STFT and convolutional neural network for bearings under variable for face recognition with degraded probe samples,’’ Neurocomputing,
working conditions,’’ in Proc. Prognostics Syst. Health Manage. Conf. vol. 310, pp. 229–315, Oct. 2018.
(PHM-Qingdao), Qingdao, China, Oct. 2019, pp. 1–6. [28] D.-T. Hoang and H.-J. Kang, ‘‘Rolling element bearing fault diagnosis
[6] D. Verstraete, A. Ferrada, E. L. Droguett, V. Meruane, and M. Modar- using convolutional neural network and vibration image,’’ Cognit. Syst.
res, ‘‘Deep learning enabled fault diagnosis using time-frequency image Res., vol. 53, pp. 42–50, Jan. 2019.
analysis of rolling element bearings,’’ Shock Vib., vol. 2017, Oct. 2017, [29] Z. Chen, A. Mauricio, W. Li, and K. Gryllias, ‘‘A deep learning method
Art. no. 5067651. for bearing fault diagnosis based on cyclic spectral coherence and convo-
[7] R. Yan, R. X. Gao, and X. Chen, ‘‘Wavelets for fault diagnosis of rotary lutional neural networks,’’ Mech. Syst. Signal Process., vol. 140, Jun. 2020,
machines: A review with applications,’’ Signal Process., vol. 96, pp. 1–15, Art. no. 106683.
Mar. 2014. [30] Y. Yang, H. Zheng, Y. Li, M. Xu, and Y. Chen, ‘‘A fault diagnosis scheme
[8] V. Muralidharan and V. Sugumaran, ‘‘A comparative study of Naïve Bayes for rotating machinery using hierarchical symbolic analysis and convolu-
classifier and Bayes net classifier for fault diagnosis of monoblock cen- tional neural network,’’ ISA Trans., vol. 91, pp. 235–252, Aug. 2019.
trifugal pump using wavelet analysis,’’ Appl. Soft Comput., vol. 12, no. 8, [31] R. Chen, X. Huang, L. Yang, X. Xu, X. Zhang, and Y. Zhang, ‘‘Intelli-
pp. 2023–2029, Aug. 2012. gent fault diagnosis method of planetary gearboxes based on convolution
[9] J. Ben Ali, L. Saidi, A. Mouelhi, B. Chebel-Morello, and F. Fnaiech, neural network and discrete wavelet transform,’’ Comput. Ind., vol. 106,
‘‘Linear feature selection and classification using PNN and SFAM neural pp. 48–59, Apr. 2019.
networks for a nearly online diagnosis of bearing naturally progressing [32] L. Huang and J. Wang, ‘‘Forecasting energy fluctuation model by wavelet
degradations,’’ Eng. Appl. Artif. Intell., vol. 42, pp. 67–81, Jun. 2015. decomposition and stochastic recurrent wavelet neural network,’’ Neuro-
[10] J. Li, X. Yao, X. Wang, Q. Yu, and Y. Zhang, ‘‘Multiscale local features computing, vol. 309, pp. 70–82, Oct. 2018.
learning based on BP neural network for rolling bearing intelligent fault [33] P. Liang, C. Deng, J. Wu, and Z. Yang, ‘‘Intelligent fault diagnosis
diagnosis,’’ Measurement, vol. 153, Mar. 2020, Art. no. 107419. of rotating machinery via wavelet transform, generative adversarial nets
[11] Z. Huo, Y. Zhang, L. Shu, and M. Gallimore, ‘‘A new bearing fault and convolutional neural network,’’ Measurement, vol. 159, Jul. 2020,
diagnosis method based on Fine-to-Coarse multiscale permutation entropy, Art. no. 107768.
Laplacian score and SVM,’’ IEEE Access, vol. 7, pp. 17050–17066, 2019. [34] P. Liang, C. Deng, J. Wu, Z. Yang, J. Zhu, and Z. Zhang, ‘‘Compound fault
diagnosis of gearboxes via multi-label convolutional neural network and
[12] X. Yan and M. Jia, ‘‘A novel optimized SVM classification algorithm
wavelet transform,’’ Comput. Ind., vol. 113, Dec. 2019, Art. no. 103132.
with multi-domain feature and its application to fault diagnosis of rolling
[35] Z. Lv, B. Tang, Y. Zhou, and C. Zhou, ‘‘A novel method for mechanical
bearing,’’ Neurocomputing, vol. 313, pp. 47–64, Nov. 2018.
fault diagnosis based on variational mode decomposition and multikernel
[13] Y. Li, Y. Yang, X. Wang, B. Liu, and X. Liang, ‘‘Early fault diagnosis
support vector machine,’’ Shock Vib., vol. 2016, pp. 1–11, Jan. 2016.
of rolling bearings based on hierarchical symbol dynamic entropy and
[36] K. He, X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for image
binary tree support vector machine,’’ J. Sound Vib., vol. 428, pp. 72–86,
recognition,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR),
Aug. 2018.
Las Vegas, NV, USA, Jun. 2016, pp. 770–778.
[14] X. Zhang, Y. Liang, J. Zhou, and Y. Zang, ‘‘A novel bearing fault diag-
[37] CWRU Dataset. Case Western Reserve University Bearing Data Center.
nosis model integrated permutation entropy, ensemble empirical mode
Accessed: Jul. 20, 2012. [Online]. Available: http://csegroups.case.edu/
decomposition and optimized SVM,’’ Measurement, vol. 69, pp. 164–179,
bearingdatacenter/pages/download-data-file
Jun. 2015.
[38] MFPT Dataset. Society for Machinery Failure Prevention Technology.
[15] S. Ahlawat and A. Choudhary, ‘‘Hybrid CNN-SVM classifier for handwrit- Accessed: 2012. [Online]. Available: https://mfpt.org/fault-data-sets/
ten digit recognition,’’ Procedia Comput. Sci., vol. 167, pp. 2554–2560, [39] S. Wang, J. Xiang, Y. Zhong, and Y. Zhou, ‘‘Convolutional neural
Jan. 2020. network-based hidden Markov models for rolling element bearing fault
[16] H. Rathore, A. K. Al-Ali, A. Mohamed, X. Du, and M. Guizani, ‘‘A novel identification,’’ Knowl.-Based Syst., vol. 144, pp. 65–76, Mar. 2018.
deep learning strategy for classifying different attack patterns for deep [40] W. Lu, B. Liang, Y. Cheng, D. Meng, J. Yang, and T. Zhang, ‘‘Deep model
brain implants,’’ IEEE Access, vol. 7, pp. 24154–24164, 2019. based domain adaptation for fault diagnosis,’’ IEEE Trans. Ind. Electron.,
[17] S. Zhang, X. Pan, Y. Cui, X. Zhao, and L. Liu, ‘‘Learning affective video vol. 64, no. 3, pp. 2296–2305, Mar. 2017.
features for facial expression recognition via hybrid deep learning,’’ IEEE [41] X. Li, W. Zhang, and Q. Ding, ‘‘Cross-domain fault diagnosis of rolling
Access, vol. 7, pp. 32297–32304, 2019. element bearings using deep generative neural networks,’’ IEEE Trans. Ind.
[18] Y. Li, T. Zhang, S. Sun, and X. Gao, ‘‘Accelerating flash calculation Electron., vol. 66, no. 7, pp. 5525–5534, Jul. 2019.
through deep learning methods,’’ J. Comput. Phys., vol. 394, pp. 153–165,
Oct. 2019.
[19] C. Zhang, H. Zhang, J. Qiao, D. Yuan, and M. Zhang, ‘‘Deep transfer
learning for intelligent cellular traffic prediction based on cross-domain LAOHU YUAN received the B.S. degree in
big data,’’ IEEE J. Sel. Areas Commun., vol. 37, no. 6, pp. 1389–1401,
automation from Air Force Engineering Univer-
Jun. 2019.
sity, Xi’an, China, in 2002, and the M.S. and
[20] J. Schmidhuber, ‘‘Deep learning in neural networks: An overview,’’ Neural
Ph.D. degrees in aircraft design from the Harbin
Netw., vol. 61, pp. 85–117, Jan. 2015.
Institute of Technology, Harbin, China, in 2004
[21] M. Qiao, S. Yan, X. Tang, and C. Xu, ‘‘Deep convolutional and LSTM
recurrent neural networks for rolling bearing fault diagnosis under strong and 2010, respectively. He is currently a Profes-
noises and variable loads,’’ IEEE Access, vol. 8, pp. 66257–66269, 2020. sor Shenyang Aerospace University, Shenyang,
[22] H. Shao, H. Jiang, X. Li, and T. Liang, ‘‘Rolling bearing fault detection China. His research interests include intelligent
using continuous deep belief network with locally linear embedding,’’ systems and intelligent control.
Comput. Ind., vol. 96, pp. 27–39, Apr. 2018.
[23] S. Haidong, J. Hongkai, L. Xingqiu, and W. Shuaipeng, ‘‘Intelligent fault
diagnosis of rolling bearing using deep wavelet auto-encoder with extreme
learning machine,’’ Knowl.-Based Syst., vol. 140, pp. 1–14, Jan. 2018.
[24] X. Li, H. Jiang, X. Xiong, and H. Shao, ‘‘Rolling bearing health prognosis DONGSHAN LIAN received the B.S. degree in
using a modified health index based hierarchical gated recurrent unit automation from the Hunan University of Science
network,’’ Mechanism Mach. Theory, vol. 133, pp. 229–249, Mar. 2019. and Technology, Xiangtan, China, in 2018. She
[25] A. Abu Mallouh, Z. Qawaqneh, and B. D. Barkana, ‘‘Utilizing CNNs is currently pursuing the M.S. degree in aircraft
and transfer learning of pre-trained models for age range classification design and engineering with Shenyang Aerospace
from unconstrained face images,’’ Image Vis. Comput., vol. 88, pp. 41–51, University, Shenyang, China. Her research inter-
Aug. 2019. ests are in intelligent systems and the application
[26] G. Xu, X. Su, W. Liu, and C. Xiu, ‘‘Target detection method based on of machine learning to fault diagnosis.
improved particle search and convolution neural network,’’ IEEE Access,
vol. 7, pp. 25972–25979, 2019.

VOLUME 8, 2020 137405

 L. Yuan et al.: Rolling Bearing Fault Diagnosis Based on Convolutional Neural Network and Support Vector Machine

XUE KANG received the B.S. degree in aircraft KEJIA ZHAI received the B.S. degree in traffic and
design and engineering from Shenyang Aerospace transport from the Anyang Institute of Technology,
University, Shenyang, China, in 2018, where she Anyang, China, in 2019. He is currently pursuing
is currently pursuing the M.S. degree in aircraft the M.S. degree in aircraft design and engineering
design. Her research interests are in intelligent with Shenyang Aerospace University, Shenyang,
systems and the application of machine learning China. His research interests are computer vision
to fault diagnosis. and artificial intelligence.

YUANQIANG CHEN received the B.S. degree

in aircraft design and engineering from Shenyang
Aerospace University, Shenyang, China, in 2019,
where he is currently pursuing the M.S. degree
in aircraft design. His research interest includes
intelligent control.

137406 VOLUME 8, 2020