Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Measurement 75 (2015) 180–191

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Rolling element bearing fault detection using PPCA


and spectral kurtosis
Jiawei Xiang ⇑, Yongteng Zhong, Haifeng Gao
College of Mechanical and Electrical Engineering, Wenzhou University, Wenzhou 325035, PR China

a r t i c l e i n f o a b s t r a c t

Article history: A hybrid approach using probabilistic principal component analysis (PPCA) and spectral kur-
Received 17 February 2015 tosis (SK) is proposed to detect rolling element bearing faults. The approach includes three
Received in revised form 7 June 2015 main steps. In a first step, the signal-to-noise ratio (SNR) of PPCA denoising model is improved
Accepted 22 July 2015
through the selection of two key parameters. In the model, the primary information and fault
Available online 6 August 2015
signals will be preserved by allotted in the principal component subspace, while noises and
linear interrelated information will be discarded by projected to the residual subspace. In a
Keywords:
second step a band-pass filter for the denoising signal is designed using rapid spectral kurto-
Probabilistic principal component analysis
Denoising
sis procedure to determine optimal center frequency and bandwidth. The third step is to per-
Spectral kurtosis form a Hilbert envelope spectrum analysis of the filtered signal to extract the fault
Rolling element bearing frequencies of the rolling element bearings. The effectiveness of the proposed approach is
Fault diagnosis demonstrated by numerical simulation and experimental investigation of rolling element
bearing with different kind of faults. It indicates that employing the proposed scheme with
PPCA and SK results in the effective detection of faults in rolling element bearings.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction classification using intelligence techniques, which divide


the vibration signals into health type or fault types
Some modulation signals, e.g., amplitude modulation [16,17]. However, a great deal of fault samples should be
caused by inner race or out race faults, pulse modulation given to perform these methods. The second directly
caused by rolling element fault, will be appeared when employs signal processing method to extract feature fre-
one part of bearing has fault. The reason is that series of quencies, which have achieved a great progress [1–
periodic impact load will be generated when every rolling 15,18,19]. The most effective method is the envelope anal-
element passes through the damage position. Generally, ysis (EA), and the effectiveness of this method depends on
the bearing fault can be detected according to this modula- how well the vibration signal is preprocessed and what
tion frequency [1–7]. However, these modulation signals denoising method is adopted. Recently, wavelet transform
are commonly carried by a high frequency signal and are has been proved to be a good choice for impact signal
easily contaminated by environmental noise. Therefore, denoising [1–7,18–23]. Empirical mode decomposition
many novel signal processing methods are introduced to (EMD) and the further improvement methods were
deal with such impacting signals [8–15]. explored for reducing the white noise [23–28]. Local mean
Generally, there are two groups of methods to detect decomposition and its variant approach were developed to
faults in rolling element bearings. The first one is fault protrude the impact component whereas the stable com-
ponent is filtered effectively [29–32]. To further improve
the performance of detecting bearing faults, a generalized
⇑ Corresponding author. demodulation time–frequency analysis was proposed to
E-mail address: wxw8627@163.com (J. Xiang).

http://dx.doi.org/10.1016/j.measurement.2015.07.045
0263-2241/Ó 2015 Elsevier Ltd. All rights reserved.
J. Xiang et al. / Measurement 75 (2015) 180–191 181

demodulate the fault feature frequency [33,34]. However, the number of samples. The n-by-k matrix
the above methods are limited to apply on certain condi- P ¼ fp1 ; p2 ;    ; pk g 2 Rnk is the loading matrix (principal
tions constrained, e.g., the impact period cannot be too component matrix) with limited condition of k < n, where
short, the sampling data must be long enough, etc. k is the number of principal components. The k-by-m
Therefore, spectral kurtosis (SK) in associate with the fast matrix u ¼ fu1 ; u2 ;    ; um g 2 Rkm is the principal compo-
computation of the kurtogram technique was developed nent matrix and is set to meet Gaussian distribution with
to identify rolling element bearing faults [35–41]. zero mean and I (unit matrix) covariance, i.e., N(0, I). E is
The methods mentioned above have an obvious limita- an isotropic Gaussian noise matrix and is set to satisfy
tion for signal denoising because of the absence of proba-
Nð0; r2 IÞ; where r2 denotes the variance parameter of noise
bilistic model for raw data. As a result, high frequency
variable. Therefore, X meets Nð0; PPT þ r2 IÞ.
noises may not be separated correctly or some fault infor-
The probability distribution of u is given by
mation may be filtered as noises [42,43]. Therefore, it is
 
necessary to improve the signal-to-noise ratio (SNR) and 1 XT X

probabilistic principal component analysis (PPCA) [42,43] pðuÞ ¼ ð2pÞ k=2
e 2r2
ð2Þ
might be a perfect denoising method. It firstly established
The conditional probability distribution over x-space
an appropriate probabilistic model for every variable. Then
for a given u is
a new sample data can be afresh generated based on the
 
new model. Last, principal component can be finally  1 jjXPujj2

acquired through orthogonal projection onto the principal pðxjuÞ ¼ ð2pÞ n=2
e 2r 2
ð3Þ
component subspace using raw data. Generally, PPCA is
Therefore, the probability distribution of x is
utilized as a constrained Gaussian density model. The ben-
Z
efits of doing so is that the unknown parameters for the
pðxjuÞpðuÞdx ¼ ð2pÞn=2 jCj1=2 eð2X C XÞ
1 T 1
pðxÞ ¼ ð4Þ
covariance matrix can be iteratively solved using EM algo-
rithm [44–46].
In PPCA model, every variable is expressed with proba- where C ¼ PPT þ r2 I is n  n covariance matrix deter-
bilistic distribution and a new model is reconstructed mined by P and r2 .
based on the linear combination between the principal According to Bayes’ theorem [47,48], conditional proba-
component and noise. With this model, the fault informa- bility distribution of u for a given x is
tion and noise can be distinguished by comparing the dis-  T 1

k 1 12½uM1 PT XT  ððr2 Þ MÞ½uM1 PT X
pðujxÞ ¼ ð2prÞ 2 jðr2 Þ Mj2 e
1

tortion degree of the raw data. However, it is possible that


some portion of original data may not well match the new ð5Þ
model. Therefore, an improved PPCA is required and the
original variable number and the principal component where M ¼ PT P þ r2 I is a k  k matrix whose dimension
number are two important parameters, which determine has been reduced.
the raw data dimensionality and the reduced data dimen- As shown in Eqs. ((2)–(5)), the probability model can be
sionality, respectively [45,46]. obtained once the parameters P and r2 are determined.
The purpose of the present work is to establish a robust The expectation–maximization (EM) algorithm [49,50]
hybrid approach using PPCA and SK to overcome the weak- can be selected to estimate parameters as
ness that the SK is sensitive to non-Gaussian noises.
e ¼ SPðr2 I þ M1 PT SPÞ1
P ð6Þ
According to the denoising performance, the PPCA model
is improved through the selection of the principal compo- 1
nent variable number and the original variable number.
r~ 2 ¼ trðS  SPM1 Pe T Þ ð7Þ
n
The performance of the present approach has also been Pm
where S ¼ m i¼1 xi xTi is the covariance matrix of the origi-
1
examined using numerical simulation and experimental
investigation. nal variables, trðÞ represents the trace of matrix.
Repeatedly iterate between Eqs. (6) and (7) until con-
vergence, P and r2 are calculated and the PPCA model is
2. PPCA denoising and SK finally obtained.

2.1. PPCA model 2.2. PPCA denoising

In this section, a brief review of PPCA is summarized When PPCA model is constructed, the denoising model
[42–46]. PPCA model is a latent variable model with factor will be obtained using the following transform
analysis method and its probability distribution model is
given by ui ¼ pTi X ð8Þ
Eq. (8) reveals that each principal component data
X¼ PuþE ð1Þ
(denoising signal) is the projection from the raw data X
In Eq. (1), X ¼ fx1 ; x2 ;    ; xm g 2 Rnm is a n-by-m matrix to the corresponding principal component vector pi .
(generated by one-dimensional signal x ¼ xr  x  r , where xr Finally, the PPCA denoising model with reduction matrix
is the raw signal and x r is the mean of xr ), where n is the dimension is obtained and the denoising purpose for raw
number of original variables (embedding dimension), m is signal is realized.
182 J. Xiang et al. / Measurement 75 (2015) 180–191

0 4 1
2.3. The selection of two key parameters Xþ1 
  A
E@ xðiÞwði  tÞej2 pfi

i¼1 
Similar to principal component analysis (PCA) [51],
SKð f Þ ¼ 0 0 2 112  2 ð13Þ
PPCA denoising effect is determined by the number of X þ1 

@E@ 
principal components k and the number of original vari- xðiÞwði  tÞej2pfi  AA
ables n [44]. Generally, n is determined by personal expe- i¼1 
rience that enough fault information is contained in the
matrix X [52], which will also be evaluated using embed- From Eq. (13), we can see that SK is related to the center
ding dimensional estimation algorithms, such as near frequency and bandwidth of filter (window function). If the
neighbor method, correlation integral method, and singu- bandwidth is chosen infinitely small, SK tends to zero.
lar value decomposition [51]. In the present, we are using Alternatively, if the bandwidth is chosen too wide, over-
the numerical simulation to select an agreeable parameter much noise will be included and the narrowband transient
n when the parameter k is fixed. signal cannot be detected with SK computation. Therefore,
k is commonly determined by the cumulative variance for non-stationary signals, SK is a function of the center
contribution rate (CVCR). A single CVCR in term of every frequency and the bandwidth.
principal component is defined as [51] The optimal center frequency and the bandwidth can be
calculated by fast kurtosis diagram [39]. Then the optimal
" #
X
n parameters are employed to perform band-pass filter to
g% ¼ ki = kt  100% i ¼ 1; 2; 3; :::; k ð9Þ obtain the largest SK [39].
t¼1

where k is the eigenvalue of the covariance matrix of the


original variables S. Therefore, the sum CVCR of all princi-
pal components can be expressed by 3. The hybrid approach using PPCA denoising and SK
" , #
X
k Xk Xn In this section, we propose a hybrid approach using
g% ¼ kt kt  100% ð10Þ PPCA and SK to detect faults of the rolling element bearing.
t¼1 t¼1 t¼1
As shown in Fig. 1, three steps in the proposed approach
P
Generally, the threshold value of kt¼1 g% is suggested are
as 85% to determine the parameter k [51]. Once the calcu-
lation result using Eq. (10) is approximation equals to (or (1) Build PPCA model
lager than) 85%, the corresponding k is selected. From Eq.
(10), we can see that the values of the CVCR will exceed PPCA model is generated using different n and k. Then
P
100% for the cumulative effects. the relationship among k, n and the sum CVCR kt¼1 g% is
constructed and the relatively optimal parameters n and
k will be obtain using Eq. (10) and the corresponding
2.4. Spectral kurtosis and band-pass filter PPCA denoising model will be constructed. Finally, the
raw signal is denoised by PPCA denoising model and noises
SK [35–41] is sensitive to impulse signal and is com- are restrained. In the present, the Matlab Toolbox for
monly used for signal processing not only detects transient Dimensionality Reduction coded by Dr. Maaten [53] is
impact component but also confirms the kurtosis fre- applied to build up PPCA model.
quency. However, SK is easily obstructed by noise and it
is necessary to take preprocessing procedures (denoising (2) Perform SK analysis
and filter) for the raw signal. Generally, SK is expressed
by the ratio of fourth order cyclic cumulant and variance, According to the fast kurtosis diagram, the optimal cen-
is given by [35] ter frequency and the bandwidth are determined. Then
EðjYðt; f Þj4 Þ band-pass filter is designed to filter the PPCA denoising
SKðf Þ ¼ 2
2 ð11Þ signal. For the filtered signal, Hilbert envelope spectrum
ðEðjYðt; f Þj2 ÞÞ analysis is applied and the demodulation frequency is
where EðÞ and j  j stand for expectation and module, finally obtained. In this paper, the Matlab Toolbox for
respectively, Yðt; f Þ is the complex envelope of raw signal Fast Kurtogram coded by Dr. Antoni [54] is employed to
perform SK decomposition. The Toolbox is also applied to
at the position of frequency f. Yðt; f Þ is calculated with
propose an adaptive SK technique to detect single and mul-
short time Fourier transform (STFT) as [35,36]
tiple transient faults of bearings [55,56].
X
þ1
STFTðt; f Þ ¼ xðiÞwði  tÞej2pfi ð12Þ (3) Detect result
i¼1

where wðtÞ is the window function used for the raw signal Compared the demodulation frequency with the theo-
filter. retical fault feature frequency, the type of bearing fault
In Eq. (11), if Xðt; f Þ replaced by STFTðt; f Þ, we obtain can be determined
J. Xiang et al. / Measurement 75 (2015) 180–191 183

with standard deviation 5. The sampling frequency


Step1 Raw signal f s ¼ 40; 960 Hz, the sampling points are N = 4096 and the
raw signal is xr .
The simulation signal and its frequency spectrum are
PPCA model generation shown in Fig. 2. The raw data matrix X is commonly con-
structed using xr by the following Hankel matrix as [51]
No 2 3
xr ð1Þ xr ð2Þ . . . xr ðN  n þ 1Þ
Parameters 6 x ð2Þ
optimized? 6 r xr ð3Þ . . . xr ðN  n þ 2Þ 7
7
X¼6 7 ð17Þ
4 ... ... ...  5
Yes xr ðnÞ xr ðn þ 1Þ . . . xr ðNÞ
PPCA denoising From Eq. (17), we can see clearly that the raw data ser-
ies xr is conveyed into n  ðN  n þ 1Þ matrix X. It points
out that Hankel matrix X (state-space model) is formed
Step2 Fast kurtosis diagram using a data sequence xr , which is a matrix that is symmet-
ric and constant across the anti-diagonals, and has ele-
ments xij ¼ xr ði þ j  1Þ.
Determine the center
frequency and the bandwidth The principal component number k and the original
of band-pass filter variable number n are carefully chosen and the corre-
sponding calculation values of CVCR are shown in
Table 1, one conclusion can be obtained: the more the orig-
Band-pass filter
inal variable number n is given, the lower the CVCR value
should be.
Calculate Hilbert envelope Furthermore, to establish a PPCA model, we consider
spectrum the simulation case, i.e., k is fixed to 2, different n is consid-
ered to testify the PPCA denoising effect. SNR (SNR is calcu-
Demodulation frequency lated by SNR ¼ 10log10 ðrs =rn Þ, where rs and rn are the
average power spectrums of signal and noise, respectively)
of the raw signal and purified signal are shown in Table 2.
Step3 Detection result The effectiveness of PPCA denoising is obvious and the SNR
is the highest when n is set to 20. Moreover, in Table 1, the
CVCR value (85.61%) is lager than 85% when k is increased
Fig. 1. The flowchart of fault detection approach.
to 9. Therefore, to detect the feature frequency of a rolling
element bearing with faults, the relative optimal parame-
4. Numerical simulations ters of PPCA denoising might be n = 20 and k = 9. The wave-
form and its spectrum of purified signal using PPCA with
To validate the performance of PPCA denoising model, n = 20 and k = 9 are shown in Fig. 3(a) and (b), respectively.
we consider a simulation signal (includes bearing vibration The graph of local area G (see the dotted ellipse circle in
signal, low-frequency interference signal and white noise). Fig. 3(b)) indicates that a series of indistinctly visible mod-
The simulation signal is defined as ulation frequency is produced around the center frequency
(4000 Hz) of the bearing. However, the modulation fre-
xðtÞ ¼ sðt þ TÞ þ cðtÞ þ eðtÞ ð14Þ quency (100 Hz) cannot be directly detected.
To further verify the performance of the hybrid fault
where sðtÞ is the bearing impulse response signal, T is
detection approach, the optimal center frequency and
impulse period, cðtÞ is the interference signal relative with
bandwidth can be established from fast kurtosis diagram
rotating frequency and its harmonic components, and eðtÞ
as shown in Fig. 4(a). The center frequency fc correspond-
is white noise.
ing to the maximum SK is 4160 Hz. The optimal bandwidth
cðtÞ and sðtÞ are given by
Bw is 640 Hz and maximum kurtosis is located in the fourth
cðtÞ ¼ 5 sinð2pf 0 tÞ þ 1:5 sinð4pf 0 tÞ þ 0:5 sinð6pf 0 tÞ ð15Þ layer by the magnitude 0.3. Therefore, the purified signal is
band-pass filtered using the optimal center frequency and
and bandwidth. Finally Hilbert envelope spectrum is calculated
sðtÞ ¼ eBt cosð2pf n tÞ ð16Þ based on the filtered signal, as shown in Fig. 4(b). The
faults frequency 100 Hz and its harmonics components
in which f 0 is the shaft rotating frequency, f n is the natural (200 Hz and 300 Hz) are detected exactly. However, when
frequency of the bearing. B ¼ 2pf n n is the attenuation coef- the purified signal is replaced by the raw signal, the fast
ficient depended on damping ratio n and natural frequency kurtogram and the corresponding Hilbert envelope spec-
f n. trum are shown in Fig. 5(a) and (b), respectively.
Suppose the parameters are: T ¼ 0:01 s, f 0 ¼ 1480 rpm Obviously, the fault frequency (100 Hz) will be mixed with
=60 s ¼ 24:67 Hz, f n ¼ 4000 Hz, n ¼ 0:02387 and the corre- 30 Hz for its harmonics components are completely sub-
sponding B ¼ 600, eðtÞ is a standard normal distribution merged in noise and only the unknown frequency
184 J. Xiang et al. / Measurement 75 (2015) 180–191

15 0.7

10 0.6

5 0.5
Amplitude

Amplitude
0.4
0
0.3
-5
0.2
-10 0.1

-15 0
0 0.05 0.1 0.15 0.2 0 1000 2000 3000 4000 5000 6000
t /s f /Hz

Fig. 2. The waveform and frequency spectrum of the simulation signal with noises.

Table 1
The input parameters and the CVCR values (%).

k n k n
10 15 20 25 10 15 20 25
1 88.96 82.28 75.12 68.59 14 – 94.92 94.27 94.02
2 89.07 82.71 76.69 70.93 15 – 99.86 95.96 96.49
3 89.23 83.85 77.50 71.61 16 – – 97.43 97.81
4 89.61 84.45 78.56 73.48 17 – – 99.71 100.51
5 89.96 85.32 80.11 75.21 18 – – 101.37 102.95
6 90.30 86.04 81.33 77.40 19 – – 105.46 105.65
7 90.46 86.99 82.89 79.21 20 – – 108.92 108.29
8 90.97 87.69 84.39 81.36 21 – – – 111.21
9 91.54 88.63 85.61 83.67 22 – – – 113.99
10 91.84 89.62 87.41 85.56 23 – – – 118.11
11 – 90.13 89.01 88.15 24 – – – 121.18
12 – 91.23 90.50 90.17 25 – – – 135.03
13 – 92.64 92.54 92.32

Table 2
SNR of different n with k ¼ 2. of speed monitor, manual speed governor, acceleration
n 10 15 20 25 sensors, speed sensors, motors, spindles and computer
with VQ data acquisition software [57].
SNR* (db) 5.0 5.0 5.0 5.0
SNR** (db) 5.9 7.5 7.7 7.0 A deep groove bearing with bearing symbol ER-12 K
(3/4 in. bearing) is employed in our experiment. In order
SNR* stands for SNR of original signal.
to simulate the contact fatigue failure, the single pitting
SNR** stands for SNR of purified signal.
defection processed by the electro-discharge machining is
pre-installed on the inner race and the rolling element pro-
components, such as 160 Hz and 320 Hz can be found in vided by SpectraQuest Inc. [57], which are shown in Fig. 7.
the graph. The bearing parameters are: the number of rolling ele-
According to the above analysis, it is shown that the ments N b ¼ 8, ball diameter Bd ¼ 0:3125 in:, pitch diameter
PPCA denoising method with the parameters k = 9 and Pd ¼ 1:318 in:, the contact angle a ¼ 0 [57]. In the present,
n = 20 can effectively protrude the fault frequency around the shaft speed is f shaft ¼ 30 Hz (1800 rpm), therefore, the
the bearing center frequency. Moreover, the performance ball pass frequency on inner race (BPFI) is
of the present hybrid fault detection approach is verified.  
Nb Bd
BPFI ¼ f shaft 1 þ cos a ¼ 4:9484 f shaft
2 Pd
5. Experimental investigations ¼ 148:45 Hz ð18Þ

In this section, two laboratory experiments on rolling and the Ball spin frequency (BSF) is
  "  2 #
element bearing with inner race and rolling element faults Pd Bd
are conducted to validate the proposed method. The exper- BSF ¼ f 1 cos a ¼ 1:9902f shaft ¼ 59:71 Hz
2
2Bd shaft Pd
imental setup (The machinery fault simulator–magnum,
MFS–MG [57]) is shown in Fig. 6. The test system consists ð19Þ
J. Xiang et al. / Measurement 75 (2015) 180–191 185

15 0.7

10 0.6

0.5
5

Amplitude
Amplitude

0.4 G
0
0.3
-5
0.2

-10
0.1

-15 0
0 0.05 0.1 0.15 0.2 0 2000 4000 6000 8000 10000 12000
t /s f /Hz

Fig. 3. The waveform and frequency spectrum of the PPCA denoising signal (n = 20 and k = 9).

Fb-kurt2-Kmax=0.3; Level 4, Bw=640Hz, fc=4160Hz


1
0 0.25
1 0.8 200Hz
0.2
1.6 100Hz
0.6
Amplitude

2
Level K

0.15
300Hz
2.6
0.4
0.1
3

3.6 0.05 0.2

4
0
0 2000 4000 6000 8000 10000 0 500 1000 1500 2000
Frequency (Hz) Frequency (Hz)

Fig. 4. Fast kurtogram and envelope spectrum of the purified signal (fc = 4160 Hz, Bw = 640 Hz).

Fb-kurt2-Kmax=0.4; Level 4, Bw=640Hz, fc=4160Hz


0.4 0.2
0
0.35 30Hz
1
0.3 0.15
1.6 100Hz
0.25
Amplitude

2 160Hz
Level K

0.2 0.1
2.6 320Hz
0.15
3
0.1 0.05
3.6
0.05
4
0
0 2000 4000 6000 8000 10000 0 500 1000 1500 2000 2500 3000
Frequency (Hz) Frequency (Hz)

Fig. 5. Fast kurtogram and envelope spectrum of the raw signal (fc = 4160 Hz, Bw = 640 Hz).
186 J. Xiang et al. / Measurement 75 (2015) 180–191

Speed indicator Acceleration sensor

Data acquisition system

Experimental bearing

Signal conditioner Manual speed governor speed sensor

Fig. 6. The machinery fault simulator–magnum (MFS–MG).

fault
fault

D=3.4mm
D=2.2mm

(a) Inner race fault (D=3.4mm) (b) Rolling element fault (D=2.2mm)
Fig. 7. The damaged faults of rolling bearings.

In the experimental processing, the sampling frequency in the 2.5th decomposition layer. Applied Hilbert envelope
f s ¼ 25:6 kHz and the total collected data is 16,384. spectrum to the band-pass filtered signal, we finally obtain
the envelope spectrum of the purified signal for the bear-
5.1. Inner race fault detection ing with inner race fault, as shown in Fig. 10(b). The shaft
rotating frequency is exactly equal to 30 Hz and fault fre-
The raw and purified signals (using PPCA denoising with quency 148.4 Hz and its harmonic frequency 296.9 Hz is
k = 8 and n = 20, the CVCR value is 86.5) and the correspond- clearly demodulated, which is matched with the theoreti-
ing frequency spectrums are shown in Figs. 8 and 9, respec- cal calculation value 148.4 Hz (show in Eq. (18)).
tively. Compared Figs. 8(a) and 9(a), we can see that the Therefore, the bearing fault for inner race is detected.
impact characteristic is not shown in the raw signal By contrast, the same procedures are conducted directly
whereas it is clearly revealed in the purified signal. to the raw signal without PPCA denoising and the result is
However, as shown in Figs. 8(b) and 9(b), the feature fre- shown in Fig. 11. As shown in Fig. 11(a), the center fre-
quency cannot be directly revealed in the corresponding quency is 10,400 Hz, the optimal bandwidth is 533 Hz
frequency spectrums of raw and purified signals. and maximum kurtosis value is 5.5, which are situated in
Therefore, the SK analysis is performed to detect the fre- the 4.5th decomposition layer. Finally, the feature
quency band and further demodulate the feature frequency. frequency is completely distorted and the shaft rotating
As shown in Fig. 10(a), for the purified signal, SK is frequency f shaft ¼ 30 Hz and the unknown frequency com-
calculated using fast kurtosis diagram and the center fre- ponents, such as 118.8 Hz, 170.3 Hz and 243.5 Hz are
quency is 9600 Hz, the optimal bandwidth is 2133 Hz found in Fig. 11(b). Therefore, the bearing with inner race
and maximum kurtosis value is 151.2, which are situated fault will not be detected.
J. Xiang et al. / Measurement 75 (2015) 180–191 187

-3
x 10
0.15 3

2.5
0.1

2
Amplitude

Amplitude
0.05
1.5
0
1

-0.05
0.5

-0.1 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 5000 10000 15000
t /s f /Hz

Fig. 8. The raw signal and frequency spectrum for the bearing with inner race fault.

-3
0.15 x 10
2.5

0.1
2

0.05
Amplitude

Amplitude

1.5

0 1

-0.05 0.5

-0.1 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 5000 10000 15000
t /s f /Hz

Fig. 9. The purified signal and frequency spectrum for the bearing with inner race fault.

Fb-kurt2-Kmax=151.2; Level 2.5, Bw=2133Hz, fc=9600Hz -3


x 10
0 150 2
1
1.6
2 1.5
2.6 100 30Hz
Amplitude

3
Level K

3.6
1 148.4Hz
4 178.1Hz
4.6
5 50 296.9Hz
5.6 0.5
6
6.6
7
0 0
0 2000 4000 6000 8000 10000 12000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz)
(a) Fast kurtogram (b) Envelope spectrum

Fig. 10. Fast kurtogram and envelope spectrum of the purified signal for the bearing with inner race fault (fc = 9600 Hz, Bw = 2133 Hz).
188 J. Xiang et al. / Measurement 75 (2015) 180–191

Fb-kurt2-Kmax=5.5; Level 4.5, Bw=533Hz, fc=10400Hz


-3
x 10
0 2
1 5
1.6
2 4 1.5
2.6
30Hz
3
Level K

Amplitude
3.6 3
1 118.8Hz
4 170.3Hz
4.6 2
5 243.5Hz
5.6 0.5
6 1
6.6
7
0 0
0 2000 4000 6000 8000 10000 12000 0 200 400 600 800 1000
Frequency (Hz) Frequency (Hz)

Fig. 11. Fast kurtogram and envelope spectrum of the raw signal for the bearing with inner race fault (fc = 10,400 Hz, Bw = 533 Hz).

0.15 x 10
-3

3.5

0.1 3

0.05 2.5
Amplitude

Amplitude

2
0
1.5
-0.05
1
-0.1
0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 5000 10000 15000
t /s t /s

Fig. 12. The raw signal and frequency spectrum for the bearing with ball fault.

-3
0.15 x 10
2

0.1
1.5
0.05
Amplitude

Amplitude

0 1

-0.05
0.5
-0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 5000 10000 15000
t /s t /s

Fig. 13. The purified signal and frequency spectrum for the bearing with ball fault.

5.2. Rolling element fault detection fault. The impact characteristic is contaminated by struc-
tural and environmental noise and cannot be revealed in
Figs. 12 and 13 show the raw and purified signals (using raw signal, as shown in Fig. 12(a). Because the noise is
PPCA denoising with k = 6 and n = 20, the corresponding reduced tremendously, the impact characteristic can be
the CVCR value is 85.9) for the bearing with rolling element seen in the purified signal, as shown in Fig. 13(a).
J. Xiang et al. / Measurement 75 (2015) 180–191 189

Fb-kurt2-Kmax=245.6; Level 2.5, Bw=2133.3Hz, fc=9600Hz -3


x 10
0 5
1
1.6 200 4
2
30Hz
2.6
3 150
Level K

Amplitude
3
3.6
4 59.3Hz
4.6 100 2
5
5.6
6 50 1
6.6
7 0
0 0 100 200 300 400 500 600
0 2000 4000 6000 8000 10000 12000
Frequency (Hz) Frequency (Hz)

Fig. 14. Fast kurtogram and envelope spectrum of the purified signal for the bearing with ball fault (fc = 9600 Hz, Bw = 2133.3 Hz).

Fb-kurt2-Kmax=24.5; Level 7, Bw=100Hz, fc=1050Hz


-3
x 10
0 1.5
1
1.6 20
2
2.6 30Hz
1
3 15
Level K

Amplitude

3.6
4 40.63Hz
4.6 10
5 0.5
5.6
6 5
6.6
7
0 0
0 2000 4000 6000 8000 10000 12000 0 100 200 300 400 500 600
Frequency (Hz) Frequency (Hz)

Fig. 15. Fast kurtogram and envelope spectrum of the raw signal for the bearing with rolling element fault (fc = 1050 Hz, Bw = 100 Hz).

However, the frequency spectrums both in Fig. 12(b) (raw fault and some reasonably results can be acquired for
signal) and Fig. 13(b) (purified signal) will not provide the vibration signal contaminated by noise.
feature frequency information.
Fig. 14(a) shows fast kurtogram of the purified signal for 6. Conclusions
the bearing with rolling element fault. The optimal param-
eters are as follows: the center frequency is 9600 Hz, the A hybrid approach using PPCA and spectral kurtosis is
bandwidth is 2133.3 Hz and the maximum kurtosis value proposed to detect and diagnose rolling element bearing
is 245.6, which are situated in the 2.5th decomposition fault. The noise and low frequency interference is suffi-
layer. The envelope spectrum of the purified signal for ciently suppressed by PPCA. SNR is improved and the fault
the bearing with rolling element fault is shown in Fig. 14 modulation frequency is protruded around the bearing
(b). The shaft rotating frequency 30 Hz and the fault fre- natural frequency. The principal component number and
quency 59.3 Hz (matched with the theoretical calculation original variable number are optimized with CVCR and
value 59.71 Hz) are detected and then the rolling element denoising effect. The optimal center frequency and band-
fault will be definitely detected. width are established from fast kurtosis diagram. Hilbert
The same procedures are conducted directly to the raw envelope spectrum is calculated for the filtered signal
signal without PPCA denoising. The results of SK analysis and finally the characteristic frequency of the rolling ele-
and envelope spectrum are shown in Fig. 15(a) and (b), ment bearing is verified. Simulation and experimental
respectively. We can see that no useful information will investigations have been carried out to examine the perfor-
be obtain except for the shaft rotating frequency 30 Hz. mance of the proposed approach. The results indicate that
The above numerical investigations demonstrate that it can extract modulation frequency easily and can be used
the proposed approach is effective for detecting bearing for fault diagnosis of rolling element bearing.
190 J. Xiang et al. / Measurement 75 (2015) 180–191

Acknowledgements machinery fault detection, Sci. China Technol. Sci. 56 (5) (2013)
1294–1306.
[22] B.Q. Chen, Z.S. Zhang, Y.Y. Zi, Z.J. He, C. Sun, Detecting of transient
The authors are grateful to the support from the National vibration signatures using an improved fast spatial-spectral
Science Foundation of China (No. 51175097), the Zhejiang ensemble kurtosis kurtogram and its applications to mechanical
signature analysis of short duration data from rotating machinery,
Provincial Natural Science Foundation for Excellent Young
Mech. Syst. Sig. Process. 40 (1) (2013) 1–37.
Scientists of China (No. LR13E050002), the Zhejiang [23] Y.G. Lei, Z.J. He, Y.Y. Zi, Application of the EEMD method to rotor fault
Technologies R&D Program of China (No. 2014C31103) diagnosis of rotating machinery, Mech. Syst. Sig. Process. 23 (4)
(2009) 1327–1338.
and the Project-sponsored by SRF for ROCS, SEM.
[24] Y.G. Lei, J. Lin, Z.J. He, M.J. Zuo, A review on empirical mode
decomposition in fault diagnosis of rotating machinery, Mech. Syst.
Sig. Process. 35 (1–2) (2013) 108–126.
References [25] Y.G. Lei, N.P. Li, J. Lin, S.Z. Wang, Fault diagnosis of rotating
machinery based on an adaptive ensemble empirical mode
[1] V.N. Patel, N. Tandon, R.K. Pandey, Defect detection in deep groove decomposition, Sensors 13 (12) (2013) 16950–16964.
ball bearing in presence of external vibration using envelope analysis [26] J.S. Cheng, D.J. Yu, J.S. Tang, Y. Yang, Application of SVM and SVD
and Duffing oscillator, Measurement 45 (5) (2012) 960–970. technique based on EMD to the fault diagnosis of the rotating
[2] W.C. Tsao, Y.F. Li, D.L. Duc, An insight concept to select appropriate machinery, Shock Vib. 16 (1) (2009) 89–98.
IMFs for envelope analysis of bearing fault diagnosis, Measurement [27] J.S. Cheng, D.J. Yu, J.S. Tang, Y. Yang, Local rub-impact fault diagnosis
45 (6) (2012) 1489–1498. of the rotor systems based on EMD, Mech. Mach. Theory 44 (4)
[3] R. Kuma, M. Singh, Outer race defect width measurement in taper (2009) 784–791.
roller bearing using discrete wavelet transform of vibration signal, [28] Y. Yang, Y.G. He, J.S. Cheng, D.J. Yu, A gear fault diagnosis using
Measurement 46 (1) (2013) 537–545. Hilbert spectrum based on MODWPT and a comparison with EMD
[4] W.Y. Liu, J.G. Han, J.L. Jiang, A novel ball bearing fault diagnosis approach, Measurement 42 (4) (2009) 542–551.
approach based on auto term window method, Measurement 46 (10) [29] J.S. Cheng, Y. Yang, A rotating machinery fault diagnosis method
(2013) 4032–4037. based on local mean decomposition, Dig. Sig. Process. 22 (2) (2012)
[5] M. Singh, R. Kumar, Thrust bearing groove race defect measurement 356–366.
by wavelet decomposition of pre-processed vibration signal, [30] J.S. Cheng, K. Zhang, Y. Yang, An order tracking technique for the gear
Measurement 46 (9) (2013) 3508–3515. fault diagnosis using local mean decomposition method, Mech.
[6] H.F. Xiao, Y.M. Shao, X.J. Zhou, An improved simplex-based adaptive Mach. Theory 55 (2012) 67–76.
evolutionary digital filter and its application for fault detection of [31] Y. Yang, J.S. Cheng, K. Zhang, An ensemble local means
rolling element bearings, Measurement 55 (9) (2014) 25–32. decomposition method and its application to local rub-impact
[7] X.F. Liu, B. Lin, H.L. Luo, Bearing faults diagnostics based on hybrid fault diagnosis of the rotor systems, Measurement 45 (3) (2012)
LS-SVM and EMD method, Measurement 59 (2015) 145–166. 561–570.
[8] C. Li, M. Liang, Continuous-scale mathematical morphology-based [32] Y. Yang, H.H. Wang, J.S. Cheng, K. Zhang, A fault diagnosis approach
optimal scale band demodulation of impulsive feature for bearing for roller bearing based on VPMCD under variable speed condition,
defect diagnosis, J. Sound Vib. 331 (26) (2012) 5864–5879. Measurement 46 (8) (2013) 2306–2312.
[9] C. Li, M. Liang, Y. Zhang, S.M. Hou, Multi-scale autocorrelation via [33] J.S. Cheng, Y. Yang, D.J. Yu, Application of the improved generalized
morphological wavelet slices for rolling element bearing fault demodulation time-frequency analysis method to multi-component
diagnosis, Mech. Syst. Sig. Process. 31 (2012) 428–446. signal decomposition, Sig. Process. 89 (6) (2009) 1205–1215.
[10] C. Li, M. Liang, Time-frequency signal analysis for gearbox fault [34] J.S. Cheng, Y. Yang, D.J. Yu, The envelope order spectrum based on
diagnosis using a generalized synchrosqueezing transform, Mech. generalized demodulation time-frequency analysis and its
Syst. Sig. Process. 26 (2012) 205–217. application to gear fault diagnosis, Mech. Syst. Sig. Process. 24 (2)
[11] A.B. Ming, Z.Y. Qin, W. Zhang, F.L. Chu, Spectrum auto-correlation (2010) 508–521.
analysis and its application to fault diagnosis of rolling element [35] J. Antonia, R.B. Randall, The spectral kurtosis: a useful tool for
bearings, Mech. Syst. Sig. Process. 41 (1–2) (2013) 141–154. characterizing non-stationary signals, Mech. Syst. Sig. Process. 20 (2)
[12] T.Y. Wang, M. Liang, J.Y. Li, W.D. Cheng, Rolling element bearing fault (2006) 282–307.
diagnosis via fault characteristic order (FCO) analysis, Mech. Syst. [36] J. Antonia, R.B. Randall, The spectral kurtosis: application to the
Sig. Process. 45 (1) (2013) 139–153. vibratory surveillance and diagnostics of rotating machines, Mech.
[13] J.Q. Hu, L.B. Zhang, W. Liang, Dynamic degradation observer for Syst. Sig. Process. 20 (2) (2006) 308–331.
bearing fault by MTS–SOM system, Mech. Syst. Sig. Process. 36 (2) [37] N. Sawalhi, R.B. Randall, H. Endo, The enhancement of fault detection
(2013) 385–400. and diagnosis in rolling element bearing using minimum entropy
[14] R.L. Jiang, J. Chen, G.M. Dong, L. Tao, W.B. Xiao, The weak fault deconvolution combined with spectral kurtosis, Mech. Syst. Sig.
diagnosis and condition monitoring of rolling element bearing using Process. 21 (6) (2007) 2616–2633.
minimum entropy deconvolution and envelope spectrum, J. Mech. [38] N. Sawalhi, R.B. Randall, Spectral kurtosis optimization for rolling
Eng. Sci. 227 (5) (2013) 1116–1129. element bearing, in: Processing 8th International Symposium on
[15] H.f. Tang, J. Chen, G.M. Dong, Sparse representation based latent Signal Processing and its applications, ISSPA 2005, Australia, 2005,
components analysis for machinery weak fault detection, Mech. pp. 839–842.
Syst. Sig. Process. 46 (2) (2014) 373–388. [39] J. Antoni, Fast computation of the Kurtogram for the detection of
[16] Y.G. Lei, D.T. Kong, J. Lin, M.J. Zuo, Fault detection of planetary transient faults, Mech. Syst. Sig. Process. 21 (1) (2007) 108–124.0.
gearboxes using new diagnostic parameters, Meas. Sci. Technol. 23 [40] Y.G. Lei, J. Lin, Z.J. He, Y.Y. Zi, Application of an improved kurtogram
(5) (2012). 055605-1-10. method for fault diagnosis of rolling element bearing, Mech. Syst.
[17] Y.G. Lei, J. Lin, M.J. Zuo, Z.J. He, Condition monitoring and fault Sig. Process. 25 (5) (2011) 1738–1749.
diagnosis of planetary gearboxes: a review, Measurement 48 (2014) [41] F. Combet, L. Gelman, Optimal filtering of gear signal for early
292–305. damage detection based on the spectral kurtosis, Mech. Syst. Sig.
[18] J.L. Chen, Y.Y. Zi, Z.J. He, J. Yuan, Compound faults detection of Process. 23 (3) (2009) 652–668.
rotating machinery using improved adaptive redundant lifting [42] C.M. Bishop, M.E. Tipping, Probabilistic principal component
multiwavelet, Mech. Syst. Sig. Process. 38 (1) (2013) 36–54. analysis, J. Roy. Stat. Soc. B 61 (3) (1999) 611–622.
[19] J.L. Chen, M.J. Zuo, Y.Y. Zi, Z.J. He, Construction of customized [43] C.M. Bishop, M.E. Tipping, A hierarchical latent variable model for
redundant multiwavelet via increasing multiplicity for fault data visualization, IEEE Trans. Pattern Anal. Mach. Intell. 20 (3)
detection of rotating machinery, Mech. Syst. Sig. Process. 42 (1–2) (1998) 281–293.
(2014) 206–224. [44] K. Dongsoon, L. In-Beum, Process monitoring based on probabilistic
[20] B.Q. Chen, Z.S. Zhang, C. Sun, B. Li, Y.Y. Zi, Z.J. He, Fault feature PCA, Chemometr. Intell. Lab. Syst. 67 (1) (2003) 109–123.
extraction of gearbox by using overcomplete rational dilation [45] A. Bellas, C. Bouveyron, M. Cottrell, J. Lacaille, Model-based
discrete wavelet transform on signals measured from vibration clustering of high-dimensional data streams with online mixture
sensors, Mech. Syst. Sig. Process. 33 (2012) 275–298. of probabilistic PCA, Adv. Data Anal. Classif. 7 (2013) 281–300.
[21] B.Q. Chen, Z.S. Zhang, Y.Y. Zi, Z.B. Yang, Z.J. He, A pseudo wavelet [46] P. Zuccolotto, Principal component analysis with interval imputed
system-based vibration signature extracting method for rotating missing values, Adv. Statist. Anal. 96 (1) (2012) 1–23.
J. Xiang et al. / Measurement 75 (2015) 180–191 191

[47] A. Utsugi, T. Kumagai, Bayesian analysis of mixtures of factor [53] L. van der Maaten, Matlab Toolbox for Dimensionality Reduction,
analyzers, Neural Comput. 13 (5) (2001) 993–1002. 2015 <http://homepage.tudelft.nl/19j49/Matlab_Toolbox_for_Dim-
[48] F.R. Cardoso, J.A. Achcar, C.L. Piratelli, J.L.G. Hermosilla, J.C. Barbosa, ensionality_Reduction.html>.
Bayesian analysis of employee suggestions in a food company, Int. J. [54] J. Antoni, Fast Kurtogram, 2015 <http://www.mathworks.-
Adv. Manuf. Technol. 70 (2014) 2059–2070. com/matlabcentral/fileexchange/48912-fast-kurtogram>.
[49] M.I. Jordan, R.A. Jacobs, Hierarchical mixtures of experts and the EM [55] Y.X. Wang, M. Liang, An adaptive SK technique and its application for
algorithm, Neural Comput. 6 (2) (1994) 181–214. fault detection of rolling element bearings, Mech. Syst. Sig. Process.
[50] G. McLachlan, T. Krishnan, The EM Algorithm and Extensions, 2nd 25 (5) (2011) 1750–1764.
ed., Wiley, New York, 2008. [56] Y.X. Wang, M. Liang, Identification of multiple transient faults based
[51] I.T. Jolliffe, Principal Component Analysis, 2nd ed., Springer-Verlag, on the adaptive spectral kurtosis method, J. Sound Vib. 331 (2)
New York, 2008. (2012) 470–486.
[52] P.S. Landa, Time series analysis for system identification and [57] MFS-MG User Guides, SpectraQuest Inc, Richmond, 2012.
diagnostics, Phys. D 48 (1991) 232–254.

You might also like