Journal of Computational Information Systems 7:3 (2011) 745-753

Available at http://www.Jofcis.com

Incipient Fault Diagnosis of Rolling Element Bearing Based on Wavelet

Packet Transform and Energy Operator

Zhongqing WEI1, Zhinong JIANG1,†, Bo MA1, Xin ZHONG2, Jinji GAO1

1
Diagnosis and Self-Recovery Engineering Research Center, Beijing University of Chemical Technology, Beijing 100029, China
2
The First Research Institute of the Ministry of Public Security, Beijing 100048, China

Abstract
This paper mainly deals with the issue of incipient fault diagnosis of rolling element bearing. Firstly, an envelope
demodulation technique based on wavelet packet transform and energy operator was applied to extract the fault feature of
vibration signal, Secondly, the relative spectral entropy of envelope spectrum and the gravity frequency were applied to
construct two-dimensional features for distinguishing the fault types. Furthermore, the K-nearest neighbors (KNN) was
used to identify the faults automatically. The method could avoid inaccurate diagnosis only depending on the characteristic
frequency of rolling element bearing. The experimental results proved that the method was very effective for incipient fault
diagnosis of rolling element bearing; therefore, this method has certain application value.

Keywords: Wavelet Packet Transform; Energy Operator; Rolling Element Bearing; Incipient Fault; Envelope Spectrum;
KNN

1. Introduction

Rolling element bearing is widely used in industrial equipment. And fault diagnosis of rolling element
bearing is very important to improve reliability and performance of mechanical equipment. Generally
speaking, the faults of rolling element bearing mainly include irregular damage (such as spalling, pitting,
cracks, etc.) of inner race, outer race, balls and cage. Due to the affect of impact force, the vibration signal
often displays a complex modulation wave [1-4].
Currently, resonance demodulation is a widely used method for fault diagnosis of rolling element bearing.
It mainly adopts envelope analysis through the band pass filter, and then extracts the time-domain
characteristics of envelope signal to diagnosis the fault. In the process, one of the key factors to constrain
the success of resonance demodulation is how to select the center frequency and the bandwidth of envelop
analysis reasonably [5-6]. Wu et al. [7] developed an energy operator demodulation method to diagnosis
fault of rolling element bearing. It was proved that the energy operator demodulation method was better
than Hilbert transform, but it still needed to choose rational parameters for band-pass filter. Because the
incipient fault signal of rolling element bearing is weak and it is often disturbed by noise. It is usually
difficult to find out the characteristics of the carrier frequency band for band-pass filter in advance.

†
Corresponding author.
Email addresses: jiangzhinong@263.net (Zhinong JIANG).

1553-9105/ Copyright © 2011 Binary Information Press

March, 2011
 Z. Wei et al. /Journal of Computational Information Systems 7:3 (2011) 745-753 746

Therefore, it is inconvenient for practical application. Considering the high precision and speed of wavelet
multi-resolution, time-frequency localization and energy operator demodulation, Lei et al. [8] developed
the wavelet-energy operator demodulation method to diagnosis fault of rolling element bearing. The energy
concentration of high frequency modulation was extracted by using wavelet transform, and the fault
characteristic frequency was obtained through the energy operator demodulation. This method reduces the
interference of noise to some extent. However, because the results of envelope demodulation have multiple
spectrum lines, the characteristic frequency of fault may not be obvious that cause the difficulty to
precision for diagnosis fault of rolling element bearing. In general, the incipient fault signal of rolling
element bearing is weak and modulated. In view of both the accuracy and real-time of signal envelope
demodulation, this paper studied the demodulation method using the energy operator. In the pre-processing
of signal, wavelet packet transform was used to avoid selecting characteristic bands of demodulation of
signal artificially. Then the relative spectral entropy of envelope spectrum and the gravity frequency were
calculated as the two-dimensional features to distinguish fault type. Finally, the K-nearest neighbors
(K-NN), which was the very simple and effective method of pattern recognition, was used to distinguish
the fault. The experimental results proved that the method could obtain good performance in incipient fault
of rolling element bearing.
This paper is organized as follows. In section 2 we will introduce mathematical model of vibration signal
of rolling element bearing. In section 3 we will introduce the envelope demodulation based on wavelet
packet transform and energy operator, followed by the experimental evaluations in section 4. The
conclusions are given in section 5.

2. Mathematical Model of Vibration Signal of Rolling Element Bearing

The modulation phenomenon of vibration acceleration signal acquired from rolling element bearing is quite
obvious. Impulse response will be generated owing to the fault of rolling element bearing. Assuming that
transmitting path from the impact point of fault to the sensor installation position is unchanged, the function
of impulse response is expressed as h (t ) , the vibration signal acquired from rolling element bearing is
expressed as x (t ) , and then the following formula can be obtained [9]:

+∞
x(t ) = xT (t ) + n(t ) = ∑ Dk h(t − kT ) + n(t ) (1)
k =0

Where xT (t ) is the vibration impact response signal under the fault condition; Dk is the strength coefficient
of impact response; n (t ) is the vibration signal other than the impact response caused by the fault; T is the
period of impact response of the fault.
From the formula (1), when the incipient fault of rolling element bearing occurs, since the strength
coefficient expressed as Dk is weak, it can be seen that the impact response of vibration signals expressed
as xT (t ) is feeble. Therefore, in order to diagnose the incipient fault of rolling element bearing
successfully, the key factor is to eliminate the influence of other vibration response signal expressed as n (t ) ,
and separate the weak impact response from original signal accurately.
 747 Z. Wei et al. /Journal of Computational Information Systems 7:3 (2011) 745-753

3. Envelope Demodulation Based on Wavelet Packet Transform and Energy Operator

3.1. Energy Operator Demodulation

Energy Operator (EO) is a powerful nonlinear operator proposed by Teager, which is able to extract the
signal energy based on mechanical and physical considerations. It has been used in various speech
applications successfully. And energy operator is generally denoted as Ψ , it can be used very well for
analysis and tracking the energy of narrowband signals [10]. Assuming a continuous signal is expressed as:
t
x(t ) = a(t )cos[ ∫ ωi (τ )dτ ] (2)
0

Then the continuous energy operator can be defined as:

x(t ) ≈ [a(t )ωi (t )]2

Ψ c [ x(t )] = [ x (t )]2 − x(t ) ⋅  (3)

According to the literature [10], it can be obtained the two key formula of energy operator:
Ψ c [ x(t )]
a(t ) = (4)
Ψ c [ x (t )]

ωi (t ) = Ψ c [ x(t )] Ψ [ x(t )]

(5)
c

Formula (4) and formula (5) show that the energy operator can be used to demodulate the amplitude and
instantaneous frequency of signal. Relative to the envelope demodulation of Hilbert transform, the speed
and accuracy of envelope demodulation using energy operators is better than using Hilbert transform [11].
Therefore, this paper adopts the method to improve discrete-time energy operator demodulation algorithm.

2.2. Analysis of Envelope Demodulation Based on Wavelet Packet Transform and Energy Operator
Wavelet packet transform is essentially the bandwidth and multi-band filter for the signal, and it can
achieve the good results in denoising [12-15]. The selection indicators of fault diagnosis of rolling element
bearing not only need to consider the optimal criteria of wavelet packet basis, but also need to consider the
best decomposition level of wavelet packet [2, 16].
Supposing there is a vibration signal with the sampling length expressed as N (the value of N should
satisfy two times exponentially), and the sampling frequency expressed as Fs , then the vibration signal is
decomposed by wavelet packet transform with J layers, so the No. J layer wavelet packet coefficient can be

expressed as c J , k (i ) , where k = 0,1," , 2 J − 1 , i = 0,1," , N J − 1 ; N J = N 2 J , the bandwidth of each

band is expressed as Fb = Fs 2 J +1 . If the maximum cut-off frequency of the envelope signal is expressed

as FJ , then the Fs , FJ and Fb must meet the formula (6), which is determined by the characteristics of
envelope signal.
Fs
Fb ≥ 2 FJ ⇔ ≥ 2 FJ (6)
2 J +1
Based on the formula (6), the upper bound of the maximum decomposition level J can be derived; it
also needs to meet the conditions of N J > 2 . Therefore, it needs the formula (7) to calculate the upper
bound of J .
 Z. Wei et al. /Journal of Computational Information Systems 7:3 (2011) 745-753 748

⎧ Fs
⎪ J ≤ log 2 −1
⎨ FJ (7)
⎪
⎩ J < log 2 N − 1
After confirming the maximum decomposition level J , the selection criteria of optimal wavelet packet

basis can be analyzed. The energy value using wavelet packet transform coefficients c J , k (i ) of each

layer is used as the cost function to evaluate the best choice of the wavelet packet basis. Energy expressed

as E J , k of wavelet packet transform coefficients c J , k (i ) is calculated by the formula (8).

N J −1
EJ , k = ∑ cJ ,k (i)2 (8)
i =0

Through the wavelet packet transform, it can be decomposed the layer J , which meets the formula (9);
and the wavelet packet transform coefficients can be reconstructed.
E j max = max( E j 0 , E j1 )
E j min = min( E j 0 , E j1 ) (9)
Ej max
> Kth E j min

Where the K th is generally 1.5 or 2.

Therefore, based on the above criteria of wavelet packet transform, and according to the formula (10), it
can be easily obtained the m th frequency band component of vibration signals expressed as [11]:
+∞
xm (t ) = ∑ Dk hm (t − kT ) (10)
k =0

Where the hm (t ) is the impact response function in the m th sub-band weight, the value of n (t ) is zero in
relation to high frequency sub-band component, using the narrow scale to analysis, It can separate the
signal of bearing from the noise, therefore, it can define the characteristic function expressed as
s (t ) ,which is fault signal of rolling element bearing.

+∞ N
s(t ) = ∑ Dk ∑ hmi (t − kT ) (11)
k =0 i =1

Where the m i is the i th sub-band number of the impact response function distribution. According to
the formula (11), the characteristic function is acquired to avoid the interference of low frequency noise,
which can reflect the degree of impact distribution of the fault signal very well. Therefore, the higher
signal-to-noise ratio (SNR) can be obtained.

4. Experimental Data Analysis

This study uses experimental data from the bearing data center of Case Western Reserve University
(CWRU) (http://www.eecs.case.edu/laboratory/bearing/download.htm). The data acquisition system was
 749 Z. Wei et al. /Journal of Computational Information Systems 7:3 (2011) 745-753

used to acquire the data of the motor drive end bearing by acceleration sensors. The type of drive end
bearing was SKF6205-2RS, the speed was 1796r/mim, the sampling frequency was 12KHz, and the
sampling number which is intercepted was 10240. We studied with normal, the fault of inner race, outer
race and balls operation, which were the four kinds of running modes of rolling element bearing. The
radiuses expressed as r of fault damage included 7mils, 14mils, 21mils, 28mils and 40mils, which
showed the varying degrees of damage. This paper selected the small damage radius to simulate the case of
incipient fault occurred, that is to say, fault data for research was the r with 7mils.
According to calculation, the characteristic frequency of inner race fault was 162.1Hz, the characteristic
frequency of outer race fault was 107.3Hz, and the characteristic frequency of balls fault was 141.1Hz. As
shown in figure 1, they were the spectrograms of vibration signals under the four operating conditions.

Fig.1 The SpectroGrams of Four Running Status of Rolling Bearing

It could be seen from figure 1, the spectrograms of vibration signals under the four operating conditions
were rather complicated, it was difficult to judge and classify the faults. According to the range of
characteristic frequency of bearing fault, when the working speed was 1796r/mim, the maximum frequency
of the envelope signal was 6 times the speed of work frequency; so the maximum of frequency was 180Hz.
According to the formula (6), the maximum decomposition level of wavelet packet transform was 7 by
calculation, where the decomposition level of the wavelet packet energy operator was chose for 4, and this
paper selected the db4 wavelet. The energy operator and the wavelet packet energy operator were both used
to analyze the signal envelope. As shown in figure 2, it was the comparison of the time-domain waveform
for the inner race fault mode respectively using energy operator and the wavelet packet energy operator
envelope.

Fig.2 Comparison of Envelope Waveform with Inner -race Fault

 Z. Wei et al. /Journal of Computational Information Systems 7:3 (2011) 745-753 750

It could be found that there was a lot of clutter using the energy operator envelope, and it brought a lot of
inconvenience to analyze the signal. However, in view of the time-domain waveform, the clutter of
envelope domain waveform using wavelet packet energy operator was rare. It could be also found that the
features were more obvious when the location and extent of decay in the hardest shock under the fault
condition.
Envelope spectrum analysis is often used to analyze the fault of rolling element bearing, because the
envelope analysis is essentially the peak value detector, which is also called peak energy spectrum in the
analysis of fault of rolling element bearing. As the high-frequency amplitude of envelope spectrum was
very small, the low frequency below 180Hz of envelope spectrum would be analyzed. it was relatively
difficult to pinpoint fault characteristic frequency only relying on the characteristic frequency of spectrum
by the envelope spectrum analysis, In order to distinguish the fault better, it was very important to select
the appropriate spectral characteristics of the envelope. Spectral entropy and gravity frequency were
commonly used as the characteristics of fault [17-19]. In the Literature [20], it pointed out that the spectrum
of one-dimensional spectrum entropy had some limitations, and two-dimensional vector used as the feature
parameter was proposed, named spectral entropy and spectral entropy arm, and both of them were
associated with analysis.
Spectral entropy expressed as H was defined by the formula (12), which reflected the concentration of
spectrum, the smaller the spectral entropy, the more concentrated the spectral. To compare the spectral
entropy of signal with different lengths, the relative spectral entropy expressed as H r was used as the
feature parameter; the relative spectral entropy was calculated by the formula (13). The physical
significance of spectral entropy arm was the gravity frequency of the spectral entropy. In this paper, the
gravity frequency of spectrum expressed as Fc was used directly to replace the spectral entropy arm and
the gravity frequency was calculated by the formula (14). The relative spectrum entropy H r and the
gravity frequency Fc of envelope spectrum were used as two-dimensional feature indexes.

H ( X ) = −∑ μ ( X ( K )) log( μ ( X ( K ))) (12)

Where the X was the spectral sequence of the time series of { x ( n )} , in this paper, it was the envelope

spectral sequence. And μ (t ) = X ( K ) / ∑ X ( j ) .

H r ( X ) = H ( X ) / log( N / 2) (13)

Where the N was the length of time series.

N /2
K
Fc ( X ) = ∑ N / 2 μ ( X ( K )) (14)
k =1

Under the four operating conditions of rolling element bearing, including the normal operation, inner
fault, outer fault and balls fault, it was selected the eight sets of data to calculate the characteristic index of
envelope spectrum, the results were shown in table 1.
Through the above analysis, after performing the wavelet packet transform and envelope demodulation
of the energy operator, the two-dimensional vector of the envelope spectrum was constructed as the feature
parameters, and the eigenvalue of two-dimensional vector space had a significant distinction degree under
the four running conditions.
 751 Z. Wei et al. /Journal of Computational Information Systems 7:3 (2011) 745-753

Table 1 Calculation Results of Two-dimensional Vectors under the Four Conditions

Normal Inner Fault Outer Fault Ball Fault

Data
Hr Fc Hr Fc Hr Fc Hr Fc

1 0.8331 0.2155 0.7368 0.1888 0.8511 0.1721 0.9390 0.2070

2 0.8348 0.2144 0.7264 0.1770 0.8396 0.1704 0.9452 0.2184

3 0.8421 0.2072 0.7294 0.1742 0.8487 0.1818 0.9383 0.2026

4 0.8284 0.2112 0.7382 0.1788 0.8428 0.1651 0.9299 0.2130

5 0.8446 0.2248 0.7252 0.1763 0.8543 0.1634 0.9266 0.1861

6 0.8267 0.1997 0.7369 0.1805 0.8507 0.1550 0.9451 0.2153

7 0.8384 0.2108 0.7418 0.1855 0.8456 0.1753 0.9388 0.1915

8 0.8475 0.2223 0.7243 0.1729 0.8410 0.1643 0.9313 0.2053

Mean 0.8370 0.2132 0.7324 0.1793 0.8467 0.1684 0.9368 0.2049

The standard sample expressed as X was constructed by the above four categories of feature vectors.
Then a set of fault data with the spectral envelope characteristics of two-dimensional vector was selected as
the test samples expressed as xTest , the KNN , which was the very simple and effective method of pattern
recognition, was used to identify the test samples xTest . Based on the formula (15), the standard
sample X could be calculated and identified for all sample points in sample space distance, and
its K -nearest neighbors could be found [21, 22].

d ( xi ) = (c1i − ci )2 + (c2i − c2 )2 + " + (cni − cn )2 , i = 1, 2,", N (15)

Where d ( xi ) was distance between the test sample point and standard sample point. In this paper, the
number of nearest neighbor was selected for 5 [23, 24]. According to test and analysis, the identification
result using the KNN was shown in figure 3.
From figure 3, the result of automatic recognition was balls fault, and it was consistent with the actual
fault type. Based on the analysis of multiple data measurement, the satisfactory results showed that the
proposed method was a simple and effective method of the incipient diagnosis fault of rolling element
bearing.
 Z. Wei et al. /Journal of Computational Information Systems 7:3 (2011) 745-753 752

Points of KNN

Fig.3 The Result of Fault Identification using the KNN

5. Conclusion

In this paper, a method of incipient fault diagnosis of rolling element bearing based on wavelet packet
transform and energy operator was thoroughly studied. The following results can be achieved:
(1) The paper introduced the envelope analysis technique which combines the wavelet packet transform
with energy operator demodulation. This technique not only avoids the need of an artificial carrier signal
with a band-pass filtering, but also demonstrates excellent performance of constraining the noise. Through
envelope analysis under the four kinds of status signal of rolling element bearing, the satisfactory results of
distinguishing fault are achieved.
(2) The relative spectral entropy and the gravity frequency of envelope spectrum, which was used as the
two-dimensional vector feature, was proposed to be the characteristics indicator of diagnosis fault of rolling
element bearing. Meanwhile, the KNN was used to identify the faults automatically. The experimental
results show that the method could avoid only relying on characteristic frequency of fault, and could
distinguish incipient fault of rolling element bearing automatically. Therefore, the method could be used for
realization of intelligent fault diagnosis and identification of rolling element bearing, and it has certain
application value.

Acknowledgement

This work is supported by the Fundamental Research Funds of the National 863 High-tech Research
Development Program (No.2007AA04Z433) and National Natural Science Foundation (No. 50635010) of
China.

References
[1] H. Yang, J. Mathew, L. Ma, Fault diagnosis of rolling element bearings using basis pursuit, Mechanical Systems
 753 Z. Wei et al. /Journal of Computational Information Systems 7:3 (2011) 745-753

and Signal Processing, 19(2):341-356, 2005.

[2] N. G. Nikolaou, I. A. Antoniadis, Rolling element bearing fault diagnosis using wavelet packets, NDT & E
International, 35(3):197-205, 2002.
[3] C. Junsheng, Y. Dejie, Y. Yu, Application of an impulse response wavelet to fault diagnosis of rolling bearings,
Mechanical Systems and Signal Processing, 21(2): 920-929, 2007.
[4] T. Doguer, J. Strackeljan, New time domain method for the detection of roller bearing defects, in Proceedings of
International Conference on Condition Monitoring & Machinery Failure Prevention Technologies CM 2008, 1-
11,2008.
[5] Liu Jinzhaoli, Ding Xiawan, Wang Chenguo, Adaptive resonance demodulation method and its application to
fault diagnosis of rolling bearings, Journal of Vibration and Shock, 26(1): 38- 41, 2007.
[6] R. Hao, F. Chu，Morphological undecimated wavelet decomposition for fault diagnostics of rolling element
bearings, Journal of Sound and Vibration, 302(4-5): 1164- 1177, 2009.
[7] Wu Helei, Zhu Shan’an, Lin Ruizhong, Zheng Jianfeng, Fault diagnosis of rolling bearing based on the energy
operator demodulation approach, Transactions of the Chinese Society for Agricultural Machinery, 34(1): 118-
120, 2003.
[8] Lei Wenping, Han Jie, Sun Junjie, DONG Xin-min, Fault diagnosis of rolling bearing by using wavelet and
energy operator demodulation, Journal of wuhan university of technology, 30(5): 128-131, 2008.
[9] Fu Qinyi, Zhang Yicheng, Ying Lijun, Li Guoshun, Extraction of failure character signal of rolling element
bearings by wavelet, Chinese journal of mechanical engineering, 37(2): 30-32, 2001.
[10] P. Maragos, J. F.Kaiser, T. F.Quatieri, Energy Separation in Signal Modulations with Application to Speech
Analysis, IEEE Trans. on Signal Processing, 41(10): 3024-3051, 1993.
[11] Ma Bo, Gao Jinji, Feng Kun, Application of energy operators to fault diagnosis of a reciprocating compressor,
Mechanical Science and Technology for Aerospace Engineering, 26(6): 696-700, 2007.
[12] Tang Guiji, Cai Wei, Rolling Bearings Fault Diagnosis by Using Wavelet Packet and Envelope Analysis, Journal
of vibration, measurement & diagnosis, 29(2): 201- 204, 2009.
[13] V. Purushotham, S. Narayanan,S. A. N. Prasad, Multi-fault diagnosis of rolling bearing elements using wavelet
analysis and hidden Markov model based fault recognition, NDT & E International, 38(8): 654-664, 2005.
[14] J. Lin, M. J. Zuo, K. R. Fyfe, Mechanical Fault Detection Based on the Wavelet De-noising Technique, Journal
of Vibration and Acoustics, 126(1): 9-16, 2004.
[15] X. Wang, Y. Zi, Z. He, Multiwavelet denoising with improved neighboring coefficients for application on rolling
bearing fault diagnosis, Mechanical Systems and Signal Processing, In Press, Corrected Proof
[16] B. Liu, Selection of wavelet packet basis for rotating machinery fault diagnosis, Journal of Sound and Vibration,
284(3-5): 567-582, 2005.
[17] N. Sawalhi, R. B. Randall,H. Endo, The enhancement of fault detection and diagnosis in rolling element bearings
using minimum entropy deconvolution combined with spectral kurtosis, Mechanical Systems and Signal
Processing, 21(6): 2616-2633, 2007.
[18] A. S. Kayhan, Evolutionary maximum entropy spectral analysis of chirps in noise, Signal Processing, 78(2):
151-157, 1999.
[19] D. M. Lin, E. K. Wong, A survey on the maximum entropy method and parameter spectral estimation, Physics
Reports, 193(2): 41-135, 1990.
[20] Liu Hongxing，Zuo Hongfu，Jiang Chenyu，Qu Liangsheng, A New Index-Two-dimensional Vector on a
Frequency Spectrum and Its Applications, China mechanical engineering, 10(5): 537-539, 1999.
[21] K. Fukunaga and P. M. Narendra, A branch and bound algorithm for computing k-Nearest Neighbors, IEEE
Trans. Compute, 24: 750-753, 1975.
[22] K. Yang,C. Shahabi, An efficient k nearest neighbor search for multivariate time series, Information and
Computation, 205(1): 65-98, 2007.
[23] J. Wang, P. Neskovic, L. N. Cooper, Neighborhood size selection in the k-nearest-neighbor rule using statistical
confidence, Pattern Recognition, 39(3): 417-423, 2006.
[24] A. K. Ghosh, On optimum choice of k in nearest neighbor classification, Computational Statistics & Data
Analysis, 50(11): 3113-3123, 2006.