This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

A Novel Fault Diagnosis Method Based on

Integrating Empirical Wavelet Transform and
Fuzzy Entropy for Motor Bearing
WU DENG1,3,4,5,6, SHENGJIE ZHANG2, HUIMIN ZHAO1,3,6* and XINHUA YANG1,6
1
Software Institute, Dalian Jiaotong University, Dalian, 116028 China
2
School of Electronics and Information Engineering, Dalian Jiaotong University, Dalian 116028, China
3
Traction Power State Key Laboratory of Southwest Jiaotong University, Chengdu 610031 China
4
Guangxi Key Lab of Multi-source Information Mining & Security, Guangxi Normal University, Guilin 541004, China
5
Sichuan Provincial Key Lab of Process Equipment and Control (Sichuan University of Science and Engineering),Zigong 64300, China
6
Liaoning Key Laboratory of Welding and Reliability of Rail Transportation Equipment, Dalian Jiaotong University, Dalian 116028, China
Corresponding author: Huimin Zhao (e-mail: hm_zhao1977@126.com).
This research was supported by the National Natural Science Foundation of China under Grant 51605068, Grant 51475065,and Grant 61771087, the Open
Project Program of the Traction Power State Key Laboratory of Southwest Jiaotong University under Grant TPL1705 and Grant TPL1803, the Research
Fund of Guangxi Key Lab of Multi-source Information Mining & Security under Grant MIMS17-03, the Open Project Program of Sichuan Provincial Key
Lab of Process Equipment and Control under Grant GK201613.

ABSTRACT Motor bearing is subjected to the joint effects of much more loads, transmissions and shocks
that cause bearing fault and machinery breakdown. Vibration signal analysis method is the most popular
technique that is used to monitor and diagnose the fault of motor bearing. However, the application of
vibration signal analysis method for motor bearing is very limited in engineering practice. In this paper, on
the basis of comparing fault feature extraction by using empirical wavelet transform(EWT) and Hilbert
transform with the theoretical calculation, a new motor bearing fault diagnosis method based on integrating
EWT, fuzzy entropy and support vector machine(SVM), called EWTFSFD is proposed. In the proposed
method, a novel signal processing method called EWT is used to decompose vibration signal into multiple
components in order to extract a series of amplitude modulated-frequency modulated(AM-FM) components
with supporting Fourier spectrum under an orthogonal basis. Then fuzzy entropy is utilized to measure the
complexity of vibration signal, reflect complexity changes of intrinsic oscillation, and compute the fuzzy
entropy values of AM-FM components, which are regarded as the inputs of SVM model to train and
construct a SVM classifier for fulfilling fault pattern recognition. Finally, the effectiveness of the proposed
method is validated by using the simulated signal and real motor bearing vibration signals. The experiment
results show that the EWT outperforms empirical mode decomposition for decomposing the signal into
multiple components, and the proposed EWTFSFD method can accurately and effectively achieve the fault
diagnosis of motor bearing.

INDEX TERMS Motor bearing, Fault diagnosis, Empirical wavelet transform, Fuzzy entropy, Support
vector machine, Fourier spectrum segmentation, AM-FM components

I.INTRODUCTION
With the constantly increasing automation level in rotating
machinery, fault diagnosis is paying more and more attention
in order to increase the reliability and decrease the possible
loss due to the unforeseen damage and fault. Bearing is one
of the most important components in various rotating
machinery[1]. The motor bearing is subjected to the joint
effects of much more loads, transmissions and shocks, which
cause bearing fault and machinery breakdown, and lead to
the loss of production and income, and even occur human
casualties. It was reported that bearing faults accommodate
motor failures of 45~55%[2,3]. Therefore, the fault diagnosis
of motor bearing plays a key role in the reliable operation of
motors. It must be monitored and diagnosed to guarantee the
safe operation and further prevent the breakdowns of rotating
machines. In general, when the motor bearing has broken
down, the components of outer race, inner race and rolling
element induce impulses, which can be detected in the

VOLUME XX, 2017 1

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
acquired vibration signal. When one defective bearing rotates
at a constant speed, the generated impulses are periodic or
quasi-periodic. Because the mechanical vibration signals
consist of plentiful information with the dynamical features,
the vibration-based signal processing method is one of the
principal tools for diagnosing the malfunctions of motor
bearing. Therefore, it is great importance to deeply study the
fault diagnosis method and maintain the safety and stability
operation of motors[4-7].
In recent years, a variety of fault diagnosis methods have
been effectively exploited to detect motor bearing faults at
an early stage in order to keep machinery performing state,
avoid abnormal progression and reduce productivity loss.
For fault diagnosis, the signal analysis is one of the most
important steps, because it can effectively analyze the
dynamic information. There exist various kinds of the
vibration signal analysis methods, such as spectral analysis,
correlation analysis, time series analysis, autoregressive
moving average(ARMA), fast Fourier transform(FFT),
wavelet transform(WT), empirical mode decomposition
(EMD), independent component analysis(ICA), blind signal
separation(BSS), ensemble empirical mode decomposition
(EEMD), neural network, fault tree analysis, support vector
machine(SVM), gray theory and so on. Ming et al.[8]
proposed a novel fault diagnosis method based on the
orthogonal projection theory. van Wyk et al.[9] proposed a
powerful tool for bearing time series feature extraction and
classification methods for real-time applications. Mohsen et
al.[10] proposed a fault diagnosis method of analog circuits
based on dictionary approach. Liu et al.[11] proposed a
signal processing method-amplitude recovery method.
Rubini and Meneghetti[12] proposed the limits of the
mentioned methodologies by showing their application to
bearings affected by different pitting failures. Gao et al.[13]
proposed an empirical mode decomposition method for
decomposing complicated signal. Widodo et al.[14] applied
independent component analysis for feature extraction and
data reduction from original features. Guo et al.[15]
proposed a robust feature extraction scheme for the fault
diagnosis of rolling element. Dong et al.[16] proposed a
repeated blind source separation method based on
morphological filtering and singular value decomposition to
separate the mixed sources. Xue et al.[17] proposed an
adaptively fast EEMD method. Cheng et al.[18] proposed a
new gear fault diagnosis method in strong noise based on
multi-sensor information fusion. Jayaswal et al.[19]
proposed a new approach based on wavelet transform,
artificial neural network and fuzzy rules for detecting and
localizing defects in rolling element. Widodo et al.[20]
proposed two multi-class classification techniques for fault
diagnosis through RVM and support vector machine. Hu et
al. [21] proposed a central-tapped node linked modular fault
tolerance topology for SRM. Liu et al. [22] proposed a
speculative approach for spatial-temporal efficiency with
multi-objective optimization. Deng et al. [23] proposed a
2 VOLUME XX, 2017
2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Author Name: Preparation of Papers for IEEE Access (February 2017)
genetic and ant colony adaptive collaborative optimization
algorithm. Deng et al. [24] proposed an improved adaptive
PSO algorithm for solving complex problem. Gu et al. [25]
proposed a structural minimax probability machine for
constructing a margin classifier. Rong et al. [26] proposed a
novel K+-isomorphism method. Zhao et al. [27] proposed a
new feature extraction method based on EEMD and multi-
scale fuzzy entropy. Chen et al. [28] proposed an improved
quaternion principal component analysis method for
processing nonlinear quaternion signals. Wang et al. [29]
proposed a back propagation neural network model. Deng
et al. [30] proposed a novel intelligent diagnosis method
using optimal LS-SVM with improved PSO algorithm.
Xiong et al. [31] proposed a novel reversible data hiding
scheme using integer wavelet transform, histogram shifting
and orthogonal decomposition. Zhang et al. [32] proposed
an efficient algorithm to achieve k-barrier coverage. Qu et
al. [33] proposed a multilevel pattern mining architecture to
support automatic network management. These proposed
methods have better realized the signal analysis, feature
extraction, fault diagnosis, and some research results are
obtained in recent years. But these proposed methods also
have their own limitations in the actual engineering
applications. Some methods cannot effectively analyze the
fault signal, and extract the fault feature, which result in the
lower fault diagnosis accuracy, the poor generalization
ability of fault diagnosis model, and the higher calculation
complexity, and so on.
Wavelet transform(WT) is an effective tool to analyze
the non-stationary signal [34]. A wavelet has good local
time-frequency properties. It can provide time domain and
frequency domain information by inner production between
the analyzed signal and a predetermined wavelet basis. So
this method has already shown its tremendous ability in
mechanical equipment fault diagnosis due to multi-
resolution analysis ability [35]. Dyadic wavelet
transform(DWT) is proposed to extract the fault feature. It
can save some computation time because it is a fast method.
This method has widely applied in mechanical fault
diagnosis [36,37]. But the DWT only can decompose the
low frequency sub-band. Wavelet packet transform(WPT)
is based on improving its frequency resolution through a
parallel decomposition of the high-frequency and low-
frequency bands. However, the WPT would not allow a
data-driven frequency partition due to the dyadic partition
scheme. Although these wavelet methods have gained a lot
of interesting results in engineering applications, they still
are limited in the subdivision scheme. They are very
difficult to be altered, and they can lead to severe damages
on exacting transient vibration features in transition areas of
packets. Meanwhile, the EMD method is proposed to
overcome the prescribed dyadic subdivision scheme and
achieve adaptive representation [38]. The EMD is a
completely different method, which is used to identify the
principal modes with representing the signal. It decomposes
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
the multi-modulated vibration signal into a few intrinsic
mode functions(IMFs). Each IMF is considered as a
component. Because the EMD method can effectively
extract stationary and non-stationary components from a
signal, it has been paid attentions to signal processing and
actual engineering applications. However, there is lack of
mathematical theory for EMD method, a lot of improved
issues and combining other methods are proposed in recent
years[39]. Empirical wavelet transform(EWT) based on
combining the advantages of wavelet transform and EMD
method was proposed by Gilles [40]. It is used to extract a
series of amplitude modulated-frequency modulated(AM-
FM) signals from the given signal. The EWT method is
used in fault diagnosis for weak fault and compound fault
identification. The AM-FM components can be transformed
into a compact supporting Fourier spectrum, and the
different approach is developed to construct adaptive
wavelets capable of identifying AM-FM components for
vibration signal. The different modes are identified to
segment Fourier spectrum and apply some filtering for each
detected supporting. Therefore, the EWT method can
effectively analyze a signal and extract the internal features.
For the nonlinear dynamic features of bearing fault
signals, some nonlinear analysis methods are also used in
fault diagnosis, such as approximate entropy, sample
entropy and fuzzy entropy and so on [41-43]. These
methods have made some successful applications in the
field of fault diagnosis. But the approximate entropy exits
the shortcomings of dependency on record length and lower
estimation value. The sample entropy is based on the
Heaviside step function, it is discontinuous and mutational
at the boundary. The fuzzy entropy is defined by using the
concept of membership function. It uses least mean square
and membership function method to process the similarity
measures of sequence, so it is an effective tool to measure
the data fuzziness. The SVM is a powerful machine
learning method based on statistical learning theory and
structural risk minimization principle [44]. It has been
successfully applied to fault diagnosis in order to improve
the accuracy of fault detection and diagnosis.
In this study, in order to make full use of the advantages of
EWT, fuzzy entropy and SVM, a new fault diagnosis method
is proposed for motor bearing. Firstly, the EWT method is
used to adaptively decompose vibration signal into a number
of AM-FM components. Next, the values of fuzzy entropy of
AM-FM components are calculated in order to generate
feature vectors, which are fed into the SVM model to train
and construct a SVM classifier for fulfilling fault pattern
recognition. Finally, the effectiveness of the proposed
method is validated by using the simulated signal and real
motor bearing vibration signals. The fault diagnosis results
are compared with the other fault diagnosis methods in the
literatures.
II. METHODS AND TECHNIQUES
2 VOLUME XX, 2017
2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Author Name: Preparation of Papers for IEEE Access (February 2017)
A. EMPIRICAL WAVELET TRANSFORM
1) EMPIRICAL MODE DECOMPOSITION
The EMD method is a non-linear multi-resolution self-
adaptive decomposition technique [45], which provides a
different approach to signal processing since it is not
defined as integral transform. It can adaptively decompose
a complicated signal into a set of intrinsic mode
functions(IMFs) without preliminary knowledge from the
high frequency to low frequency by means of a process,
called sifting. The sifting process decomposes the original
signal( S (t ) ) into a number of IMFs according to the
following expression:
n
S (t )   ci (t )  rn (t ) (1)
i 1
where rn (t ) represents the residual of the signal S (t ) ,
ci (t ) indicates the i th IMF. The IMF components, such as
c1 , c2 , c3 , ... , cn represent the frequency bands from the
low frequency to high frequency.
An IMF is the function that satisfies two following
conditions:
a) The number of extrema and the number of zero-
crossings must be either equal or differ at most by one in one
data set.
b) The mean value of the moving envelope defined by
local maxima and the envelope defined by local minima are
zero at any point.
2) WAVELET TRANSFORM
Wavelet theory has emerged as a signal processing tool in
many fields, and has some distinct merits. It was firstly
proposed by Morlet in 1984 [46]. Wavelets are
mathematical functions that divide data into different
frequency components. It is different from short time
Fourier transform(STFT) in that each component is studied
with a resolution to its scale.
For the sake of simplicity, one-dimensional wavelets are
considered to illustrate the related concepts. Wavelet
transform(WT) is a method for obtaining harmonic
components. It employs a mother wavelet instead of
orthogonal sine and cosine functions in order to extract
frequency components. A dyadic frequency filter is used to
obtain the high frequency and low frequency components.
The low frequency components are decomposed into higher
frequency components. It is possible to obtain signal levels at
frequency bands instead of exact components. The advantage
of WT method is simple calculation and it can be completed
in less time. Generally, the WT method can be described as
follow:

F (a, b)    f ( x) (*a , b ) ( x)dx (2)
where the * is the complex conjugate symbol, and the
 is a function, which can arbitrarily be chosen that it
obeys certain rules.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

3) EMPIRICAL WAVELET TRANSFORM Parameter  can ensure that there is no overlap between
The EWT method was proposed by Gilles to obtain two consecutive transition areas. So the parameter  must
adaptive wavelet capable for identifying and extracting meet the following equation:
AM-FM components of a signal. The main idea is to extract   n
the signal modes by constructing an appropriate wavelet   min n ( n1 ) (7)
n1  n
filter bank. In the EWT method, the Fourier frequency
spectrum is normalized and separated into the number of The EWT method is defined by the similar way with
intervals, then the compactly supported orthonormal wavelet transform. The detailed coefficients W f (n, t ) are
wavelet basis is explicitly constructed in each interval. given by the inner products with empirical wavelets:
Hence, the segmentation operation on the Fourier spectrum W f (n, t )  f , n   f ( ) n (  t )d (8)
is the important step to make the wavelet adaptive to the And the approximation coefficients W f (0, t ) is given by 
signal. Firstly, the local maxima in the Fourier spectrum of the inner products with scaling function:
the analyzed signal is detected. Then the spectrum based on
the detected maxima is segmented. Lastly, a corresponding W f (0, t )  f , 1   f ( )1 (  t )d (9)
wavelet filter bank is constructed. Then the reconstruction signal and empirical mode are
Assuming that a signal consists of N components and given as follows:
each component is well separated in Fourier spectrum. Its N

Fourier spectrum is divided into N segments, and each f (t )  W f (0, t ) * 1 (t )   W f (n, t ) * n (t ) (10)
n 1
segment is corresponding to one mode.  n is the boundary
limits between each segments, 0 =0 and  N =  . f 0 (t )  W f (0, t ) * 1 (t ) (11)
Partitioning of Fourier axis is shown Figure 1. For each  n , 
f k (t )  W f (k , t ) * k (t ) (12)
the transition phase( Tn ) with the 2 n width is defined as the
center.  n is set as  n  n ( 0    1 ) in order to B. FUZZY ENTROPY
simplify the equation. Entropy is a general concept, and it is used to measure the
uncertainty of one system or a piece of information. Fuzzy
degree is a quantitative index to describe the degree of
fuzzy set. Fuzzy entropy is a method based on the concept
of approximate entropy and sample entropy to measure the
complexity of time series [47]. The fuzzy entropy is
described as follows:
(1) Give a sample time series with N {u (i ) : 1  i  N } .
FIGURE 1. Partitioning of Fourier axis For given m , n and r , a vector set
The empirical scaling function and wavelets are defined as { X im , i  1,2, ... , N  m  1} is formed. Each vector contains
follows:
m sequential elements from u (i ) .
 1 if |  |(1   )n (3)

 
ˆn ( )  cos[  (
1
(|  | (1   )n ))] if (1   )n |  | (1   )n X im  {u (i ), u (i  1),, u (i  m  1)}  u 0 (i ) (13)
 2 2n
 0 otherwise where u0 (i ) is the average of X im .
 1 if (1 )n |  |(1 )n1 1 m1
 
cos[ ( 1 (| | (1 )n1))] (4)
u 0 (i )   u (i  j ) (14)
if (1 )n1 |  |(1 )n1 m j 0
 2 2n1
ˆn()   (2) For certain vector X im , the distance d ijm between X im
sin[ (  )(| | (1 ) )] if (1 )n |  | (1 )n
 2 2n n
and X mj ( j  1,2,3,, N  m, i  j ) is defined as the

 0 otherwise
maximum absolute difference of corresponding scalar
The function  (x) is an arbitrary C k [0,1] function, components.
0 if x  0 and  (x)   (1 x)  1  x [0,1] (5) dijm  d[Xim, X mj]  max (u(i  k) u0 (i))(u( j  k) u0 ( j)) (15)
 (x)   k(0,m1)
1 if x  0
(3) Give n and r , the similarity degree Dijm of X mj to
Many polynomial functions meet these properties. The
following polynomial was firstly suggested by Daubechies X im is calculated by using a fuzzy function  (d ijm , n, r ) ,
and used by Gilles.  In 2 ( d ijm / r ) n
Dijm (n, r )   (d ijm , n, r )  e (16)
 ( x)  x 4 (35  84 x  70 x 2  20 x 3 ) (6)
(4) Define the function  . m

2 VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

1 N m 1 N m
 x(t )  x1 (t )  x2 (t )  x3 (t )  x4 ( x)  x5 ( x)
 m (n, r )   (  Dijm )
N  m i 1 N  m! j 1, j i
(17) 
 x1 (t )  36cos(30t)
(5) Similarly, { X im1 }is formed and the function  m 1 is  x2 (t )  10 sin(150t )
 (21)
obtained.  x3 (t )  60sin(340t )
1 N m 1 N m  x4 (t )  25 cos(520t )
 m1 (n, r )   (  Dijm1 )
N  m i 1 N  m! j 1, j i
(18) 
 x5 (t )  130 sin(600t )
(6) Fuzzy entropy of sequence {u (i ) : 1  i  N } for the The time domain waveforms of five components and the
negative natural logarithm of the deviation is defined as simulated signal are displayed in Figure 2. The sampling
follow. frequency is 1024Hz and data length is 512.
50

FuzzyEn(m, n, r )  lim [ln  m (n, r )  ln  m1 (n, r )] (19)

x1
0

N  -50

10
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

(7) If the length N is finite, FuzzyEn(m, n, r ) can be

x2
0

-10
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

changed as follow: 100

x3
FuzzyEn (m, n, r )  ln  m ( n, r )  ln  m1 (n, r ) (20) -100

50
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

x4
-50
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

C. SUPPORT VECTOR MACHINE 200

x5
0

The SVM is one of the most popular tools based on -200

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

structural risk minimization. The basic feature of SVM is to

500

0
x

map the original nonlinear data into a higher-dimensional -500

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

feature space where a hyperplane is constructed to bisect

Time(s)

FIGURE 2. Time domain waveforms of five components and simulated

two classes of data and maximize the margin of separation signal

between itself and those points in lying nearest to it. The For the comparison, the EWT and EMD methods are
hyperplane should be used as the basis for classifying employed to decompose the simulated signal. The
unknown data. Therefore, the SVM was widely applied in segmentation of the frequency spectrum and extracted
pattern recognition, nonlinear system identification, modes are obtained by EWT method in Figure .3. and
modeling, predication and control and so on [48-51]. Figure 4. The extracted modes are obtained by EMD
method in Figure 5.
The SVM is mainly used to solve the binary
classification problem. The basic theory was originally
derived from data classification. It is to find one division
plane with meeting the given requirement in order to keep
the point of the training set far away the plane. The SVM
originated from the optimal classification surface from the
linearly separable circumstance. It is used to solve the
linear constrained quadratic programming problem by
mapping the input space into the high dimensional inner
product space to obtain the global optimal solution and
guarantee the convergence speed and avoid the local FIGURE 3. Segmentation of frequency spectrum by EWT method
minimum value. 50
A M -F M 1

-50
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

III. SIMULATION EXPERIMENT OF EWT 50

In order to verify the effectiveness of EWT method, a

A M -F M 2

simulated signal is formulated in here. In general, when one -50

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

fault occurs on one of main parts of rotating machinery, the 100

A M -F M 3

corresponding vibration signal is often comprised of different 0

harmonic components and impulses, which always can be

-100
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
50

detected in the vibration signal. In order to simulate the

A M -F M 4

actual condition, the simulated signal x(t ) is designed as -50

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

follow.
200
A M -F M 5

-200
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time(s)

FIGURE 4. Extracted modes of simulated signal by EWT method

2 VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

200
Therefore, the EWT method is more robust to the noise, and
IMF1

0
-200

can obtain more accurate components by comparing with

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

100
IMF2

0
-100
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 EMD method. The EWT method has better performance in
dealing with multi-component signal, and can effectively
20
IMF3

0
-20
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

50 distinguish the different components.

IMF4

0
-50
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

1
IV. A NOVEL FAULT DIAGNOSIS METHOD BASED ON
IMF5

0
-1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

2
EWT, FUZZY ENTROPY AND SVM
The experiment vibration data comes from Bearing Data
IMF6

0
-2
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

2
Center of Case Western Reserve University [52]. The 6205-
IMF7

0
-2

2RS JEM SKF deep groove ball bearing is employed in the

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time(s)

FIGURE 5. Extracted modes of simulated signal by EMD method

experiment. The bearing information is described in Tab.1.
As can be seen from Figure 3.~Figure 5., it can be found
The motor is connected to a dynamometer and torque sensor
that the EWT method has performed well and the number
of component modes is 5, which is in accordance with the by a self-aligning coupling. The data were collected from
real component. The EMD method decomposes the accelerometers on the motor housing at the drive end of
simulated signal into 7 IMFs from the high frequency to the motor. The vibration signals were measured under no-load
low frequency in Figure 5. Due to the noise interference, (0hp) at rotating speed of 1797r/min. Faults were introduced
the EMD method generated false components without into the test bearings by using electro-discharge machining
physical significance. Therefore, it can observe that the method. The fault diameter was 0.007''. Four different
EMD method failed to determine the internal modes for the operating conditions are:(1) normal condition (2) inner race
simulated signal, and then separate some information which fault (3) outer race fault (4) rolling element fault. The
are original part of same component. For the obtained vibration signals of motor bearing were sampled at the
results by EWT method, it can see that the EWT method frequency of 12800Hz and the duration of each vibration
can better extract the presence of internal modes in the signal was 10 seconds. The original vibration signals were
spectrum, and provides different components which are divided into segmentations of samples, and each sample
close to the original ones. covered 2048 data points.
In order to further verify the effectiveness of EWT TABLE I.
method for processing the simulated signal, the The bearing information of 6205-2RS JEM SKF deep groove
decomposed signals are reconstructed, respectively. The ball bearing
reconstructed signals are in shown in Figure 6. and Figure 7. Inside Outside Ball Pitch Roller Rotating
Thickness
diameter diameter diameter diameter number speed
Original signal
300

200

0.9843 2.0472 0.5906 0.3126 1.537

9 1797r/min
100
Amplitude

0
inches inches inches inches inches
-100

-200
According to the theoretical calculation values of fault
feature frequency, the faults of outer race, inner race and
-300
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time(s)

rolling element of motor bearing are determined [53].

Reconstruction signal
300

200

100
Chandra et al. [54] gave the calculation method of fault
Amplitude

feature frequency of each element of the bearing. The

0

-100

-200

-300
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
calculated results are shown in TABLE II.
Time(s)

FIGURE 6. Reconstructed signal by EWT method

TABLE II
Fault feature frequencies
Inner race(Hz) Outer race(Hz) Rolling Element (Hz)

162.19 107.29 141.08

In here, the EWT and Hilbert transform are combined to

extract the fault feature frequencies of outer race, inner race
and rolling element of rolling bearing. When the EWT
method is used to divide the frequency band, a threshold
method by Gilles is used to determine the value of N . In
here, the value of N is set as N  4 . After the EWT
method is transformed, one mode with the most concentrated
FIGURE 7. Reconstructed signal by EMD method
energy is analyzed by using Hilbert transform. The Hilbert
As can be seen from Figure 6 and Figure 7, it can be found
envelope spectrums of fault signals are shown in Figure
that the reconstructed signal by EWT method is quite similar
8~Figure 10.
to original simulated signal. But the reconstructed signal by
EMD method is similar to partially original simulated signal.

2 VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
FIGURE 8. Envelope spectrum of frequency band with the most
concentrated energy for rolling element fault signal
FIGURE 9. Envelope spectrum of frequency band with the most
concentrated energy for inner race fault signal
FIGURE 10. Envelope spectrum of frequency band with the most
concentrated energy for outer race fault signal
As can be seen from Figure 8, the largest amplitude peak
of fault feature frequency of rolling element fault signal
appears at 105.4688Hz. The theoretical fault feature
frequency of rolling element is 141.08Hz. As can be seen
from Figure 9.,the largest amplitude peak of fault feature
frequency of inner race fault signal appears at 164.0625Hz.
The theoretical fault feature frequency of inner race is
162.19Hz. As can be seen from Figure 10.,the largest
amplitude peak of fault feature frequency of outer race fault
signal appears at 105.4688Hz. The theoretical fault feature
frequency of outer race is 107.29Hz. For the inner race fault
signal and outer race fault signal, it can be found that the
feature frequency value in the envelope spectrum is close to
the calculated fault feature frequency. But for rolling element
fault signal, the deviation between the theoretical fault
feature frequency and the largest amplitude peak of fault
feature frequency is 25.6112Hz. It is difficult to obviously
transmit the fault of rolling element by using vibration signal
and extract fault features from vibration signal. Therefore,
the fault diagnosis method based on the feature frequency has
2 VOLUME XX, 2017
2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Author Name: Preparation of Papers for IEEE Access (February 2017)
its limitation, and it is easy to be affected by the signal
acquisition condition.
According to the experimental analysis results, the EWT is
effective method for extracting the fault features of inner race
and outer race, but it is poor effect for extracting the fault
feature of rolling element. Therefore, the EWT method is
combined with other methods in order to propose a novel and
effective fault diagnosis method.
A. THE IDEA OF FAULT DIAGNOSIS METHOD
In the past decades, a lot of methods or algorithms are used
to diagnose the faults in order to propose various fault
diagnosis methods. Although each fault diagnosis method
takes on better diagnosing ability, they exist their own
shortcomings. The EWT method is a new adaptive signal
decomposition method. It can effectively avoid end effects
and mode mixing phenomenon, extract the AM-FM
components with supporting Fourier spectrum from the
original signal, and improve the reliability and the calculation
speed of noise ratio. Fuzzy entropy is an effective tool to
measure the fuzziness for data set. It uses least mean square
and membership function method to process the similar
measures of sequence, and it can effectively keep the features
of approximate entropy and sample entropy, and overcome
the effect of entropy stability due to the changed parameters.
The SVM is a classification method, which can improve the
generalization ability of machine learning by seeking the
structure risk minimization. It takes on better performance in
solving pattern recognition problems with small samples,
nonlinear and high dimension. However, the bottleneck
problem of fault diagnosis is the lack of fault samples.
Clearly, the hybrid fault diagnosis method has better features
than single fault diagnosis method. Therefore, the EWT,
Fuzzy entropy and SVM are introduced into fault diagnosis
in order to propose a novel fault diagnosis method (called
EWTFSFD) for motor bearing in this paper. Firstly, the EWT
method is used to decompose the vibration signal of motor
bearing into multiple components in order to extract the
frequency domain local maximum points, which are used to
adaptively split Fourier spectrum and obtain different modes.
The wavelet filter banks of band pass are adaptively
constructed in the frequency domain in order to extract the
AM-FM components with supporting Fourier spectrum. Next,
the fuzzy entropy is used to calculate the entropy values of
AM-FM components, which are regarded as the feature
vectors for training and testing. Finally, the SVM model is
used to construct a classifier in order to realize the
classification of fault features and obtain the diagnosis results.
B. THE MODEL OF FAULT DIAGNOSIS METHOD
In this paper, the proposed EWTFSFD method based on
integrating EWT, fuzzy entropy and SVM is proposed. The
flow of the proposed EWTFSFD method is shown in Figure
11.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
FIGURE 11. Flow of the proposed EWTFSFD method
V. APPLICATION OF THE PROPOSED EWTFSFD FOR
MOTOR BEARING
A. EXPERIMENT DATA
Data set consists of 160 data samples of four conditions
(normal condition, outer race fault, inner race fault and
rolling element fault) under no-load. Each of four
conditions includes 40 data samples. There are 80 samples
for training and 80 samples for testing. The experiment
results are from one experiment time.
B. SELECT PARAMETERS
1) SELECT PARAMETERS FOR FUZZY ENTROPY
According to the definition of fuzzy entropy, there are three
parameters of embedded dimension m , similar tolerance r
and gradient of fuzzy function n that must be fixed for
each calculation.
a) Embedded dimension m
Fuzzy entropy is the same as approximate entropy and
sample entropy. If the dimension m selects larger value, the
more detailed information is obtained in dynamically
reconfiguring the sequence of joint probability. But the
value of dimension m is larger, the required data length is
longer in calculating fuzzy entropy. In general, the data
length is N  10 m ~ 30 m . According to the experience and
actual data in the experiment, the embedded dimension
is m  2 .
b) Similar tolerance r
This parameter is used to represent the boundary width of
fuzzy function. If the value of the similar tolerance r is
larger, it will cause the information to be lost. However, if
the value of the similar tolerance r is smaller, it will cause
the unfavorable statistics features, and increase the
2 VOLUME XX, 2017
2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Author Name: Preparation of Papers for IEEE Access (February 2017)
sensitivity to noise. Typically, the similar tolerance r is
0.1~0.25SD(SD is the standard deviation of original data).
In this paper, the similarity tolerance is r =0.20SD.
c) Gradient of fuzzy function n
This parameter is used to determine the boundary
gradient of similar tolerance. If the fuzzy function n is
larger value, the boundary gradient is larger. The boundary
gradient of fuzzy function n plays a important role in the
process of calculating similarity between fuzzy entropy
vectors. If the boundary gradient of fuzzy function n is
greater than one, the more similarity contribution of the
near vectors and the less similar contribution of the farther
vectors are included. The larger value of fuzzy function n
will lead to lose the details of information. If the boundary
gradient of fuzzy function n is less than one, it is opposite.
In fact, if the boundary gradient of fuzzy function n tends
to the infinity, the fuzzy function is an unit step function. In
order to obtain the detailed information, the smaller integer
values are selected in general, such as 2, 3, 4, and so on.
Due to these reasons, the boundary gradient of fuzzy
function is n  2 .
d) Data length N
In general, the entropy values are not high for data length.
If there is m  2 , then the data length is N  100 ~ 900.
2) SELECT KERNEL FUNCTION AND PARAMETERS
OF SVM
The kernel function of SVM is introduced to map the
input space into a high-dimensional feature space, then the
optimal superplane is studied in feature space. The kernel
function K meets the following expression:
K ( xi , x j )   ( xi )  ( x j ) (22)
The process can be described as follow: the input vector
x is mapped into high-dimensional space. The commonly
used kernel functions have linear kernel function,
polynomial kernel function, radial basis kernel function,
Sigmoid kernel function and Fourier kernel function. For
different systems in the process of classification, there has
optimal kernel function with best corresponding effect. The
radial basis function has widely applied due to its simple
expression form, good smoothness and analyticity.
Therefore, the radial basis function is selected as the kernel
function of classification model, the specific expression is
described as follow.
2
K ( x, xi )  exp( x  xi /  2 ) (23)
where x is a m -dimensional input vector, xi is the
centre of the i th radial basis function and has the same
dimension with x .
In the SVM, the penalty coefficient C is used to adjust
the proportion of confidence range and experience risk in
the determined subspace in order to make best classification
ability of learning machine. For the determined
classification data, when the value of penalty coefficient C
is smaller, it means that the penalty of experience error is
smaller. Otherwise, the complexity of machine learning is
smaller and the experience risk of machine learning is
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

2 140

larger. However, the larger or smaller value of penalty

coefficient C will make poor generalization ability of
1.5 120

system. The kernel function parameter  2 is used to

1 100

determine the corresponding relation between mapping

Amplitude(m /s 2 )
Amplitude(m/s 2 )
0.5 80

function and feature space. The appropriate kernel 0 60

parameter  2 will be able to map the data into appropriate -0.5 40

feature space. However, the values of appropriate kernel -1 20

parameter  2 and penalty coefficient C are very difficult to -1.5

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
0
0 1000 2000 3000 4000 5000 6000 7000

be pre-determined. It is critical how to effectively select the Time(s)

FIGURE 14. The time-domain signal and its corresponding frequency

Frequency(Hz)

values of SVM parameters to obtain the experience risk and 0.5

spectrum of fault vibration signal of inner race
45

classification ability. In our experiments, the alternative 0.4

40

values of the kernel parameter  2 and penalty coefficient 0.3

35

C were tested and modified for some cases to obtain the

0.2
30

0.1

Amplitude(m/s 2)

Amplitude(m/s 2)
most reasonable initial values. These selected values of the
25
0
20

parameters take on the best classification ability and the -0.1

-0.2
15

most reasonable running time to efficiently complete the -0.3

10

classification. Therefore, the finally obtained values of the -0.4 5

kernel parameter  2 and penalty coefficient C are 1.0 and -0.5

0 0.02 0.04 0.06 0.08
Time(s)
0.1 0.12 0.14 0.16
0
0 1000 2000 3000 4000
Frequency(Hz)
5000 6000 7000

1.5, respectively. FIGURE 15. Time-domain signal and its corresponding frequency
spectrum of fault vibration signal of rolling element
2) SPLIT FOURIER SPECTRUM AND SELECT
C. THE REALIZATION STEPS OF EWTFSFD
FREQUENCY BAND
1) ORIGINAL VIBRATION SIGNAL AND FREQUENCY The EWT method provides the adaptability with respect to
SPECTRUM the analyzed signal. At same time, in the EWT method,
The experimental vibration signals are firstly analyzed to how to effectively segment Fourier spectrum is very
validate the effectiveness of the EWTFSFD method for important study work. The goal of segmentation is to
bearing fault diagnosis. Data set consists of 160 data separate different portions of the spectrum, which
samples of normal condition, outer race fault, inner race correspond to the different centered modes with specific
fault and rolling element fault under no-load. Each frequency of compact support. So it is a key technical
condition includes 40 data samples. 20 data samples are difficulty to segment Fourier spectrum in engineering
used as training data, and 20 data samples are used as applications. Assuming that a signal consists of N mono-
testing data. The first data in each data set is used to components and each component is well separated in
analyze in detail. The time-domain signals and Fourier spectrum. The Fourier spectrum is divided
corresponding frequency spectrums are shown in Figure into N segments, each corresponding to one mode.
12.~Figure 15. 0.3 70 i (i  1,2,3,..., N ) is denoted to be the boundary between
0.2 60 each segments, the started boundary is 0 and the end
0.1 50
boundary is  n . This implies that a total of N  1 extra
boundaries are required to find. In order to find N  1
Amplitude(m/s 2)

Amplitude(m/s 2)

0 40

-0.1 30
boundaries, the N will be determined. Gilles[31] proposed
-0.2 20
a threshold method to determine the value of N . {M i }kM1
is the magnitude of the detected local maximum points in
the frequency domain. Then these values are sorted
-0.3 10

according to the law of diminishing, and normalized to

-0.4 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0 1000 2000 3000 4000 5000 6000 7000
Time(s) Frequency(Hz)

FIGURE 12. Time-domain signal and its corresponding frequency

spectrum of normal vibration signal
[0,1]. The threshold value is M M   ( M 1  M M ) ,
1.5 100

90
 (   (0,1) ) is relative amplitude ratio. For certain  , the
1
80
number of maximum points greater than threshold value is
0.5
70
defined N , and the maximum points of the former N is
used to solve the boundary.
60
Amplitude(m/s 2)
Amplitude(m/s 2 )

0 50

40 In this paper, the features of vibration signal spectrum in

Figure 8.~ Figure11. are studied and analyzed, it may be
-0.5
30

known that the local maximum number in the spectrum is

20
-1

10

-1.5
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
0
0 1000 2000 3000 4000
Frequency(Hz)
5000 6000 7000 more, and the frequencies of some local maximum values
are very close. If all local maximum points are selected to
Time(s)

FIGURE 13. Time-domain signal and its corresponding frequency

spectrum of fault vibration signal of outer race
compute frequency bands, the segmentation intervals are
too much, it will lead to low computational efficiency and
larger storage space for extracting features. Therefore, we

2 VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

45

comprehensively considered the following factors for the 40

used segmentation method. 35

a) Study the features of stronger nonlinear and non- 30

stationary of motor vibration signal in detail. 25

Magnitude
b) Propose a compromise method between the 20

computational complexity and the classification accuracy. 15

c) Use the early experience and repeated experiments for 10

EWT method in signal processing. 5

Based on the comprehensive analysis and study, the 0

0 1000 2000 3000 4000 5000 6000 7000

selection of the boundary(local maximum value) is

Frequency (Hz)

FIGURE 19. Fourier spectrum and selected frequency band of fault

considered as follow. vibration signal of rolling element

a) The magnitude threshold is set as 30%. That is to As can be seen from Figure 16., it could be found that the
remove the local maximum points of amplitude below low frequency components occupy larger proportion. The
magnitude threshold(30%). whole spectrum is divided into three regions and three
b) The frequency threshold is set as 8 Hz. That is to different frequency bands are obtained in total. As can be
combine two adjacent maxima with interval width less than seen from Figure17., it could be found that the intermediate
8 Hz. frequency components occupy larger proportion. The whole
When the selection of the local maximum points is spectrum is divided into four regions and three different
completed, the boundary is the center of two adjacent local frequency bands are obtained in total. As can be seen from
maximum points. Fourier spectrum and selected frequency Figure 18., it could be found that the low and intermediate
band are shown in Figure 16.~Figure 19. The dotted line is frequency components occupy larger proportion. The whole
the obtained boundary in these figures. spectrum is divided into thirteen regions and thirteen
70 different frequency bands are obtained in total. As can be
60
seen from Figure 19., it could be found that the low and
intermediate frequency components occupy larger
50
proportion. The whole spectrum is divided into ten regions
40
and ten different frequency bands are obtained.
Magnitude

30 3) CONSTRUCT FILTER BANKS AND EXTRACT THE

20
AM-FM COMPONENTS
On the basis of the obtained boundary, the filter banks are
10
constructed according to the equations(3) and equations(4).
0
0 1000 2000 3000 4000 5000 6000 7000
The extracted modes for normal signal, fault vibration
signal of outer race, fault vibration signal of inner race and
Frequency (Hz)

FIGURE 16. Fourier spectrum and selected frequency band of normal

signal fault vibration signal of rolling element are shown in Figure
20.~Figure 23.
100

90
0.3

0.2
AM-FM1

80
0.1

0
70
-0.1
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16

-3
60 x 10
5
Magnitude

AM-FM2

50
0

40 -5
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16

30 0.15

0.1
AM-FM3

20
0.05

0
10
-0.05
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Time(s)
0
0 1000 2000 3000
Frequency (Hz)
4000 5000 6000 7000
FIGURE 20. Extracted modes for normal signal
0.5
FIGURE 17. Fourier spectrum and selected frequency band of fault
AM-FM1

0
vibration signal of outer race
-0.5
120 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16

0.1
100
AM-FM2

80 -0.1
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Magnitude

0.05
60
AM-FM3

40 -0.05
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16

0.1
20
AM-FM4

0 -0.1
0 1000 2000 3000 4000 5000 6000 7000 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Frequency (Hz) Time(s)

FIGURE 18. Fourier spectrum and selected frequency band of fault FIGURE 21. Extracted modes for fault vibration signal of outer race
vibration signal of inner race

2 VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

4) CALCULATE FUZZY ENTROPY VALUES

Because the fault vibration signal of motor bearing is
stronger nonlinear and non-stationary signal. It takes on the
features of complex amplitude modulation, frequency
modulation and phase modulation and so on. The
complexity of different fault vibration signal is different,
and its entropy value is also different. Some faults usually
occur in a particular frequency band, the vibration signal in
fault frequency band will be greatly changed, and its
complexity will be changed. Therefore, fuzzy entropy is
selected to describe the complexity of vibration signal of
motor bearing. It is used to calculate the extracted modes
for the different vibration signals in order to obtain the
values of fuzzy entropy. This method can reduce the
interference or coupling of feature information among these
vibration signals, and can reflect the essence of fault
information. The values of fuzzy entropy for different
vibration signals are shown in TABLE III ~TABLE VI. The
partial values of fuzzy entropy are given due to the limited
space of paper.
FIGURE 22. Extracted modes for fault vibration signal of inner race TABLE III
Fuzzy entropy values for normal signal
NO. AM-FM1 AM-FM2 AM-FM3
1 0.1888 0.0416 2.5874
2 0.1966 0.0491 3.1627
... ... ... ...
19 0.2132 0.046 3.3413
... ... ... ...
39 0.1947 0.0497 3.289
40 9.3085 3.2823 0

FIGURE 23. Extracted modes for fault vibration signal of rolling element
TABLE IV
Fuzzy entropy values for fault vibration signal of outer race
NO. AM-FM1 AM-FM2 AM-FM3 AM-FM4 AM-FM5 AM-FM6 AM-FM7 AM-FM8
1 18.2683 3.367 4.1124 5.43 0 0 0 0
2 27.3231 0.6371 3.3588 3.8643 2.1886 9.3481 0 0
... ... ... ... ... ... ... ... ...
19 46.4064 0.8942 4.3507 1.8425 3.9928 0.6988 7.1575 0
... ... ... ... ... ... ... ... ...
39 38.632 0.754 4.2778 4.7607 0.7095 5.2605 0 0
40 17.2208 3.4741 5.0154 0 0 0 0 0

2 VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

TABLE V
Fuzzy entropy values for fault vibration signal of inner race
NO AM- AM- AM- AM- AM- AM- AM- AM- AM- AM- AM- AM- AM- AM-
. FM1 FM2 FM3 FM4 FM5 FM6 FM7 FM8 FM9 FM10 FM11 FM12 FM13 FM14
1 2.5039 0.1843 0.1568 0.1299 0.4462 1.4816 0.6898 1.5806 1.7288 3.5671 8.3556 8.8272 20.9354 0
2 2.5641 0.2157 0.1707 0.1263 0.4283 0.3574 0.7371 0.8766 1.9536 2.9969 6.8966 8.6242 19.694 0
... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
19 1.7208 0.2188 0.1672 0.0917 0.1442 0.2595 0.7993 0.5418 0.825 1.657 2.8614 6.9673 7.7769 20.2835
... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
39 2.0915 0.2132 0.3247 2.3553 0.7527 3.9039 1.7234 3.3647 9.8304 8.9222 20.1835 0 0 0
40 2.7898 0.2221 0.2856 1.4043 0.8141 1.0309 1.6839 3.6083 8.2804 9.5573 0 0 0 0

TABLE VI
Fuzzy entropy values for fault vibration signal of rolling element
NO. AM-FM1 AM-FM2 AM-FM3 AM-FM4 AM-FM5 AM-FM6 AM-FM7 AM-FM8 AM-FM9 AM-FM10 AM-FM11 AM-FM12 AM-FM13
1 0.5594 0.069 0.0511 0.1302 0.724 0.6123 0.7549 0.9736 0.8992 1.1813 1.168 1.4437 0
... ... ... ... ... ... ... ... ... ... ... ... ... ...
8 0.2142 0.0269 0.0691 0.1257 0.5177 0.7578 0.9242 0.9089 1.864 0.9947 0.916 1.2841 2.0743
... ... ... ... ... ... ... ... ... ... ... ... ... ...
19 9.3939 1.7529 1.8966 1.5913 2.0483 0 0 0 0 0 0 0 0
... ... ... ... ... ... ... ... ... ... ... ... ... ...
39 8.8117 0.8651 0.4381 0.8285 1.2855 1.6908 1.7177 1.5223 1.3898 0 0 0 0
40 9.276 1.9436 1.1639 1.7804 2.2951 1.4294 0 0 0 0 0 0 0

5) TRAIN SVM MODEL obtained expectation classification result of SVM. The RBF
For actual fault diagnosis of motor bearing, it is a small is selected as the kernel function of SVM.
sample and high dimension problem. In order to illustrate 6) TEST AND OUTPUT THE RESULTS
the effectiveness of the EWTFSFD method, the SVM with In the classification experiment, the feature vectors T of
small sample and short training time is selected as a testing samples are gradually input into the trained SVM1.
classifier in here. The obtained values of fuzzy entropy are If the output of SVM1 is zero, then the normal bearing is
selected as feature vectors. That's to say, T=[AM-FM1,AM- determined and the testing experiment ends. Otherwise the
FM2, AM-FM3, ...,AM-FMn-1,AM-FMn]. Then the feature testing samples are input into the trained SVM2. If the
vectors T are input into SVM model to train the SVM output of SVM2 is zero, then the fault bearing of outer race
model and obtain a SVM classifier. According to the above is determined and the testing experiment ends. Otherwise,
method, 160 feature vectors are obtained. For each the testing samples are input into the trained SVM3. If the
vibration signal(40 samples), 20 samples are randomly output of SVM3 is zero, then the fault bearing of inner race
selected as training samples and the rest 20 samples are is determined, otherwise the fault bearing of rolling element
selected as testing samples. Because there are four different is determined. The detailed diagnosis processes do not list
vibration signals, it is necessary to establish three two- one by one. The results of testing samples are show in
classifiers. The established multi-fault classifier is TABLE VII.
described in Figure 24. TABLE VII
Diagnosis results of testing samples
Test Correct Wrong Correctness
f ( x)  0 ? f ( x)  0 ? f ( x)  0 ? States
samples number number rate (%)
Normal signal 20 20 0 100
Fault vibration signal of
20 18 2 90
outer race
Fault vibration signal of
20 20 0 100
FIGURE 24. Multi-fault classifier inner race
Fault vibration signal of
In Figure 24., SVM1 is a classifier to determine the 20 19 1 95
rolling element
normal bearing or fault bearing. SVM2 is a classifier to
determine the fault bearing of outer race or other fault D. EXPERIMENT RESULTS
bearing. SVM3 is a classifier to determine the fault bearing In order to verify the effectiveness of proposed EWTFSFD
of inner race or rolling element. f ( x)  0 represents the method for motor bearing, the fault diagnosis method based

2 VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

on EMD, fuzzy entropy and SVM(called EMDFSFD) is TABLE X.

selected to study and compare with the proposed Effects of magnitude threshold for vibration signals
Correctness rate of Correctness rate of
EWTFSFD method. The flow of EMFZSM method is States Test magnitude magnitude
samples threshold (30%) threshold (40%)
similar to the proposed EWTFSFD method. Therefore, the
Fault vibration signal
EMFZSM method is not described in detail due to the of outer race 20 95% 95%
limited space. The comparison results of fault diagnosis Fault vibration signal 20 100% 90%
methods are shown in TABLE VIII. of inner race
TABLE VIII Fault vibration signal 20 95% 90%
of rolling element
Comparison results of fault diagnosis methods
EMDFSFD EWTFSFD
States
Test
Correct rate
As can be seen from TABLE X, for the selected
samples Correct rate(%) magnitude threshold(30%), the correctness rate of fault
(%)
Normal signal 20 100 100 diagnosis for vibration signals of outer race, inner race and
Fault vibration signal rolling element are 95%, 100% and 95%, respectively. For
20 70 90
of outer race the selected magnitude threshold(40%), the correctness rate
Fault vibration signal
of inner race 20 45 100 of fault diagnosis for vibration signals of outer race, inner
Fault vibration signal 20 40 95 race and rolling element are 95%, 90% and 90%,
of rolling element
respectively. Therefore, the selected magnitude threshold
At the same time, for fault vibration signals of outer race for vibration signals has the effect on the results of fault
and inner race, two-class method is executed in order to diagnosis. By the analysis of experimental results, we can
further verify the effectiveness of proposed EWTFSFD see that the selected magnitude threshold(30%) can obtain
method for solving fault diagnosis problem. The results of more accurate diagnosis results than the selected magnitude
two-class method are show in TABLE IX. threshold(40%). For the selected magnitude threshold(20%)
TABLE IX in the experiment, the correctness rate of fault diagnosis is
Comparison results of fault diagnosis methods by using two- similar to the correctness rate of fault diagnosis of
class method magnitude threshold(30%), but the time complexity of
EMDFSFD EWTFSFD magnitude threshold(20%) is far greater than the time
States Test samples
Correct rate (%) Correct rate (%) complexity of magnitude threshold(30%). Therefore, it can
Fault vibration signal of
outer race
20 90 95 know that the selected magnitude threshold(30%) is more
Fault vibration signal of
20 70 100 reasonable and effective from this experiment.
inner race

As can be seen from TABLE VI and TABLE IX, for
normal vibration signal, it can be seen that the proposed
EWTFSFD method and EMDFSFD method have the same
testing accuracy of 100%. The testing accuracy of proposed
EWTFSFD method is 100% for fault vibration signal of
inner race. For all fault vibration signals of motor bearing,
it can be seen that the testing accuracy of proposed
EWTFSFD method is higher than the EMDFSFD method
in the experiments. At same time, it is difficult to determine
the faults between inner race and rolling element, and
between inner race and outer race for motor bearing. But
for the proposed EWTFSFD method, it can obtain
outstanding diagnosis results. Thus, we believe that the
performance and overall diagnostic capability of the
proposed EWTFSFD method outperforms the EMDFSFD
method.
E. ANALYSIS OF SELECTED
MAGNITUDE THRESHOLD
The section mainly analyzes the effects of selected method
for magnitude threshold. The magnitude thresholds are 30%
and 40%, respectively, which are selected to study and
analyze the fault diagnosis results in detail in the
experiment. The correctness rates of fault diagnosis under
two conditions are obtained and shown in TABLE X.
VI. CONCLUSION AND FUTURE RESEARCH
In the past decades, various methods and techniques have
been proposed to diagnose the faults of motor bearing.
These methods and techniques have their own shortcomings
in fault diagnosis. A novel fault diagnosis method based on
integrating EWT, fuzzy entropy and SVM is proposed to
identify fault of motor bearing. The EWT method breaks
through the limits of pre-defined dyadic scheme of
traditional DWT and allows to segment Fourier spectrum. It
is used to decompose vibration signal into several empirical
modes with physical meanings and extract a series of AM-
FM components with supporting Fourier spectrum under an
orthogonal basis. Then fuzzy entropy is used to extract the
values of fuzzy entropy of AM-FM components in order to
construct a feature vectors, which are used to train SVM
model for obtaining SVM classifier. The effectiveness of
the EWT method is firstly validated by using simulated
METHOD FOR signal. The proposed fault diagnosis method had been
applied to identify the motor bearing fault(normal signal,
outer race fault signal, inner race fault signal, and rolling
element fault signal). Some attractive features of the EWT
method and proposed fault diagnosis method are
summarized as follows:
1) The EWT method breaks through the limitations of
traditional DWT method and allow to adaptively
segment Fourier spectrum. It can effectively extract the
AM-FM components with supporting Fourier spectrum

VOLUME XX, 2017 9

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
from vibration signals with strong noise.
2) Fuzzy entropy can easily and clearly calculate the
values of fuzzy entropy of extracted modes.
3) The analysis results from experimental signals with
simulated signal indicates that the EWT method
outperforms the EMD method, and the proposed
EWTFSFD method can accurately and effectively
diagnose the different faults of motor bearing.
4) The proposed EWTFSFD method outperforms the
EMDFSFD method and the method based on EWT and
Hilbert transform.
Moreover, there are also some problems of EWT method,
such as how to adaptively segment Fourier spectrum, select
the optimal filter banks, and so on. In future work, we will
continue to study how to better adaptively segment the
spectrum and further combine EWT method with other
diagnosing technologies, optimizing parameter methods and
so on.
REFERENCES
[1] R.B. Randall, J. Antoni, “Rolling element bearing diagnostics--
Atutorial,” Mech. Syst. Signal Pr., vol. 25, no. 2, pp. 485-520, 2011.
[2] Y.H. Hu, C. Gan, Q.G. Sun, P. Li, J.H. Wu, “Modular tri-port high-
power converter for SRM based plug-in hybrid electrical trucks.
IEEE T. Power Electr., vol. 33, no. 4, pp. 3247-3257, 2018.
[3] S.B. Wang, G.G. Cai, Z.K. Zhu, W.G. Huang, X.W. Zhang,
“Transient signal analysis based on Levenberg-Marquardt method for
fault feature extraction of rotating machines,” Mech. Syst. Signal Pr.,
vol. 54-55, pp. 485-520, 2015.
[4] S.D. Wu, P.H. Wu, C.W. Wu, J.J. Ding, C.C. Wang, “Bearing fault
diagnosis based on multiscale permutation entropy and support
vector machine,” Entropy, vol. 14, no.8, pp. 1343-1356, 2012.
[5] Y.H. Hu, J.F. Zhang, W.P. Cao, J.D. Wu, G.Y. Tian, S.J. Finney,
J.L .Kirtley, “Online two-section PV array fault diagnosis with
optimized voltage sensor locations,” IEEE T. Ind..Electron., vol.62,
no.11, pp. 7237-7246, 2015.
[6] S.D. Wu, C.W. Wu, K.Y. Lee, S.G. Lin, “Modified multiscale
entropy for short-term time series analysis,”Physica A, vol.392,
no.23, pp. 5865-5873, 2013.
[7] P.E. William, M.W. Hoffman, “Identification of bearing faults using
time domain zero-crossings,” Mech. Syst. Signal Pr., vol. 25, no. 8,
pp. 3078-3088, 2011.
[8] A.B. Ming, Z.Y. Qin, W. Zhang, F.L. Chu, “Spectrum auto-
correlation analysis and its application to fault diagnosis of rolling
element bearings,” Mech. Syst. Signal Pr., vol. 41, no. 1-2, pp. 141-
154, 2013.
[9] B.J. van Wyk, M.A. van Wyk, G. Qi, “Difference Histograms: A
new tool for time series analysis applied to bearing fault diagnosis,”
Pattern Recogn. Lett., vol. 30, no.6, pp. 595-599, 2009.
[10] A.A.K. Mohsen, M.F. Abu El-Yazeed, “Selection of input stimulus
for fault diagnosis of analog circuits using ARMA model,”AEU- Int.
J. Electron. C., vol.58, no.3, pp. 212-217, 2004.
[11] Y.K. Liu, L.W. Guo, Q.X. Wang, G.Q. An, M. Guo, H. Lian,
“Application to induction motor faults diagnosis of the amplitude
recovery method combined with FFT,” Mech. Syst. Signal Pr., vol.
24, no. 8, pp. 2961-2971, 2010.
[12] R. Rubini, U.Meneghetti , “Application of the envelope and wavelet
transform analyses for the diagnosis of incipient faults in ball
bearings,” Mech. Syst. Signal Pr., vol. 15, no.2, pp. 287-302, 2001.
[13] Q. Gao, C. Duan, H. Fan, Q. Meng, “Rotating machine fault
diagnosis using empirical mode decomposition,” Mech. Syst. Signal
Pr., vol. 22, no. 5, pp. 1072-1081, 2008.
[14] A. Widodo, B.S. Yang, T. Han, “Combination of independent
component analysis and support vector machines for intelligent faults
diagnosis of induction motors,” Expert Syst. Appl., vol. 32, no.2, pp.
299-312, 2007.
VOLUME XX, 2017
2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Author Name: Preparation of Papers for IEEE Access (February 2017)
[15] Y. Guo, J. Na , B. Li, R.F. Fung, “Envelope extraction based
dimension reduction for independent component analysis in fault
diagnosis of rolling element bearing,” J. Sound Vib., vol. 333, no.13,
pp. 2983-2994, 2014.
[16] S.J. Dong, B.P. Tang, Y. Zhang, “A repeated single-channel
mechanical signal blind separation method based on morphological
filtering and singular value decomposition,” Measurement, vol. 45,
no.8, pp. 2052-2063, 2012.
[17] X.M. Xue, J.Z. Zhou, Y.H. Xu, W.L. Zhu, C.S. Li, “An adaptively
fast ensemble empirical mode decomposition method and its
applications to rolling element bearing fault diagnosis,” Mech. Syst.
Signal Pr., vol. 62-63, pp. 444-459, 2015.
[18] G. Cheng, X.H. Chen, X.L. Shan, H.G. Liu , C.F. Zhou, “A new
method of gear fault diagnosis in strong noise based on multi-sensor
information fusion,” J. Vib. Control, vol.22,no.6, pp. 1504-1515,
2016.
[19] P. Jayaswal, S.N. Verma, A.K. Wadhwani , “Development of EBP-
Artificial neural network expert system for rolling element bearing
fault diagnosis,” J. Vib. Control, vol.17,no.8, pp. 1131-1148, 2011.
[20] A. Widodo, E.Y. Kim, J.D. Son, B.S. Yang, A.C.C. Tan, D.S. Gu,
B.K. Choi, J. Mathew , “Fault diagnosis of low speed bearing based
on relevance vector machine and support vector machine,” Expert
Syst. Appl., vol. 36, no.3, pp. 7252-7261, 2009.
[21] Y.H. Hu, C. Gan, W.P. Cao, W.H. Li, S. Finney, “Central-tapped
node linked modular fault tolerance topology for SRM based
EV/HEV applications,” IEEE T. Power Electr., vol. 31, no.2, pp.
1541 - 1554, 2016.
[22] Q. Liu, W.D. Cai, J. Shen, Z.J. Fu, X.D. Liu, N. Linge , “A
speculative approach to spatial-temporal efficiency with multi-
objective optimization in a heterogeneous cloud environment,”
Secur . Commun. Netw., vol.9, no.17, pp. 4002-4012, 2016.
[23] W. Deng, H.M. Zhao, L. Zou, G.Y. Li, X.H. Yang, D.Q. Wu, “A
novel collaborative optimization algorithm in solving complex
optimization problems”, Soft Comput., vol. 21, no. 15, pp. 4387-
4398,2017.
[24] W. Deng, H.M. Zhao, X.H. Yang, J.X. Xiong, M. Sun, B. Li, “Study
on an improved adaptive PSO algorithm for solving multi-objective
gate assignment”, Appl. Soft Comput., vol.59, pp. 288-302,2017.
[25] B. Gu, X.M. Sun, V.S. Sheng , “ Structural minimax probability
machine, IEEE T. Neural Networ., vol.28, no.7, pp. 1646-1656, 2017.
[26] H. Rong, T.H. Ma, M.L. Tang, J. Cao , “A novel subgraph K+ -
isomorphism method in social network based on graph similarity
detection,”Soft Comput., vol.22, no.8, pp. 2583–2601, 2018.
[27] H.M. Zhao, M. Sun, W. Deng, X.H. Yang, “A new feature extraction
method based on EEMD and multi-scale fuzzy entropy for motor
bearing,” Entropy, vol.19, no.1, pp. 1-12, 2017.
[28] B.J. Chen, J.H. Yang, B. Jeon, X.P. Zhang, “Kernel quaternion
principal component analysis and its application in RGB-D object
recognition,” Neurocomputing, vol.266, pp. 293-303, 2017.
[29] B.W. Wang, X.D. Gu, L. Ma, S.S. Yan. , “Temperature error
correction based on BP neural network in meteorological WSN,” Int.
J. Sens. Netw., vol.23, no.4, pp. 265-278, 2017.
[30] W. Deng, R. Yao, H.M. Zhao, X.H. Yang, G.Y. Li , “A novel
intelligent diagnosis method using optimal LS-SVM with improved
PSO algorithm,” Soft Comput.,2017,DOI: 10.1007/s00500-017-
2940-9
[31] L.Z. Xiong, Z.Q. Xu, Y.Q. Shi , “An integer wavelet transform based
scheme for reversible data hiding in encrypted images,” Multidim.
Syst. Sign. P., vol.29, no.3, pp. 1191-1202, 2018.
[32] Y.H. Zhang, X.M. Sun, B.W. Wang , “Efficient algorithm for K-
barrier coverage based on integer linear programming,” China
Commun., vol.13, no.7, pp. 16-23, 2016.
[33] Z.G. Qu, J. Keeney, S. Robitzsch, F. Zaman, X.J. Wang, “Multilevel
pattern mining architecture for automatic network monitoring in
heterogeneous wireless communication networks,” China Commun.,
vol.13, no.7, pp. 108-116, 2016.
[34] R.H. Jiang, S.L. Liu, Y.F. Tang, Y.H. Liu. , “A novel method of fault
diagnosis for rolling element bearings based on the accumulated
envelope spectrum of the wavelet packet,” J. Vib. Control, vol.21,
no.8, pp. 1580-1593, 2015.
[35] J.L. Chen, Z.P. Li, J. Pan, G.G. Chen, Y.Y. Zi, J. Yuan, B.Q. Chen,
Z.J. He , “Wavelet transform based on inner product in fault
9
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2834540, IEEE
Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

diagnosis of rotating machinery: A review,” Mech. Syst. Signal Pr., artificial intelligence, optimization method and fault diagnosis.
vol. 70-71, pp. 1-35, 2016.
[36] B.M. Ebrahimi ,J. Faiz, S. Lotfi-Fard, P. Pillay, “Novel indices for SHENGJIE ZHANG (M’92), received the B.S.
broken rotor bars fault diagnosis in induction motors using wavelet degree in Electrical Engineering and Its
transform,” Mech. Syst. Signal Pr., vol. 30,no.7, pp. 131-145, 2012. Automation from Hebei University of Engineering,
[37] R.Q. Yan, R.X. Gao, X.F. Chen , “Wavelets for fault diagnosis of Handan, China, in 2015. From 2016 to now, he was
rotary machines: a review with applications,” Signal Process., vol. 96, a graduate students at School of Electronics and
pp. 1-15,2014. Information Engineering, Dalian Jiaotong
[38] N.E. Huang, Z. Shen, S.R. Long, “The empirical mode University, Dalian, China. His research interest
decomposition and the Hilbert spectrum for nonlinear and non- includes artificial intelligence, fault diagnosis.
stationary time series analysis,” Proc. R. Soc. Math. Phys. Eng. Sci.,
vol. 454, pp. 903-995,1998.
[39] J.H. Ahn, D.H. Kwak, B.H. Koh, “Fault detection of a roller-bearing HUIMIN ZHAO (F’77), received the B.S. degree in
system through the EMD of a wavelet denoised signal,”Sensors, vol. Electrical Technology from Dalian Jiaotong
14, no.8, pp. 15022-15038,2014. University, Dalian, China, in 2000, the M.S. degree
[40] J. Gilles, “ Empirical wavelet transform,” IEEE T. Signal Proces., in Traffic information engineering and control from
vol. 61, no.1, pp. 3999-4010,2013. Dalian Jiaotong University, Dalian, China, in 2004
[41] S.F. Zhao, L. Liang, G.H. Xu, J. Wang, W.M. Zhang , “Quantitative and Ph.D. degree in mechanical engineering from
diagnosis of a spall-like fault of a rolling element bearing by Dalian Maritime University, Dalian, China, in 2013.
empirical mode decomposition and the approximate entropy From 2012 to now, she was a Professor with the
method,” Mech. Syst. Signal Pr., vol. 40,no.1, pp. 154-177, 2013. Software Institute, Dalian Jiaotong University,
[42] M.D. Han, J.L. Pan , “A fault diagnosis method combined with LMD, Dalian, China. Her research interest includes artificial intelligence, signal
sample entropy and energy ratio for roller bearings,” Measurement, processing and fault diagnosis.
vol.76, pp. 7-19, 2015.
[43] Zheng JD, Cheng JS, Yang Y. , “A rolling bearing fault diagnosis
XINHUA YANG (M’69), received the B.S. degree
approach based on LCD and fuzzy entropy,” Mech. Mach. Theory,
in Mechanical Engineering from Dalian Jiaotong
vol.70, pp. 441-453, 2013.
University, Dalian, China, in 1992, the M.S. degree
[44] V. Vapnik The Nature of Statistical Learning Theory. New York;
in mechanical engineering from Dalian Jiaotong
Springer Verlag,1995.
University, Dalian, China, in 1998 and Ph.D.
[45] Y.G. Lei, J. Lin, Z.J. He, J.M. Zuo , “A review on empirical mode
degree in mechanical engineering from Dalian
decomposition in fault diagnosis of rotating machinery,” Mech. Syst.
University of Technology, Dalian, China, in 2003.
Signal Pr., vol. 35, no.1-2, pp. 108-126, 2013.
From 2007 to now, he was a Professor with the
[46] A. Grossman, J. Morlet , “Decomposition of Hardy functions into
Liaoning Key Laboratory of Welding and
square integrable wavelets of constant shape,” SIAM J. Math. Anal.,
Reliability of Rail Transportation Equipment,
vol. 15, no.4, pp. 723-736, 1984.
Dalian Jiaotong University, Dalian 116028, China. His research interest
[47] H.Y. Zhang, S.Y. Yang , “Inclusion measure for typical hesitant
includes artificial intelligence.
fuzzy sets, the relative similarity measure and fuzzy entropy,” Soft
Comput., vol. 20, no.4, pp. 1277-1287, 2016.
[48] B. Gu, V.S. Sheng, K.Y. Tay, W. Romano, S. Li, “Incremental
support vector learning for ordinal regression,” IEEE T. Neural
Networ., vol.26, no.7, pp. 1403-1416, 2015.
[49] B. Gu, V.S. Sheng , “A Robust regularization path algorithm for ν-
support vector classification,” IEEE T. Neural Networ., vol.26, no.5,
pp. 1241-1248, 2015.
[50] X.Y. Zhang, J.Z. Zhou , “Multi-fault diagnosis for rolling element
bearings based on ensemble empirical mode decomposition and
optimized support vector machines,” Mech. Syst. Signal Pr., vol.
41,no.1-2, pp. 127-140, 2013.
[51] Y. Xue, J.M. Jiang, B.P. Zhao, T.H. Ma, “A self-adaptive artificial
bee colony algorithm based on global best for global optimization,”
Soft Comput., vol. 22,no.9, pp. 2935-2952, 2018.
[52] http://csegroups.case.edu/bearingdatacenter/home
[53] X.M. Li, W. Deng, H.M. Zhao, G.H. Zheng , “Study on a novel fault
diagnosis method of rolling bearing in motor,” Recent Patents on
Mechanical Engineering, vol. 9,no.2, pp. 144-152, 2016.
[54] N.H. Chandra, A.S. Sekhar , “Fault detection in rotor bearing
systems using time frequency techniques,” Mech. Syst. Signal Pr.,
vol. 72-73, pp. 105-133, 2016.

WU DENG (M’76), received the B.S. degree in

Electrical Technology from Dalian Jiaotong
University, Dalian, China, in 2001, the M.S. degree
in computer application technology from Dalian
Jiaotong University, Dalian, China, in 2006 and
Ph.D. degree in computer application technology
from Dalian Maritime University, Dalian, China, in
2012. From 2015 to now, he was a Professor with
the Software Institute, Dalian Jiaotong University,
Dalian, China. His research interest includes

VOLUME XX, 2017 9

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.