IOP PUBLISHING MEASUREMENT SCIENCE AND TECHNOLOGY

Meas. Sci. Technol. 19 (2008) 015105 (9pp) doi:10.1088/0957-0233/19/1/015105

Wavelet spectrum analysis for bearing

fault diagnostics
Jie Liu1, Wilson Wang2, Farid Golnaraghi1 and Kefu Liu2
1
Department of Mechanical and Mechatronics Engineering, University of Waterloo, Waterloo, ON, N2L
3G1, Canada
2
Department of Mechanical Engineering, Lakehead University, Thunder Bay, ON, P7B 5E1, Canada
E-mail: j53liu@uwaterloo.ca, mfgolnar@mecheng1.uwaterloo.ca,
wilson.wang@lakeheadu.ca, kliu@lakeheadu.ca

Received 16 August 2007, in final form 22 September 2007

Published 30 November 2007
Online at stacks.iop.org/MST/19/015105

Abstract
A new signal processing technique, wavelet spectrum analysis, is proposed in this paper for
incipient bearing fault diagnostics. This technique starts from investigating the resonance
signatures over selected frequency bands to extract the representative features. A novel
strategy is suggested for the deployment of the wavelet centre frequencies. A weighted
Shannon function is proposed to synthesize the wavelet coefficient functions to enhance
feature characteristics, whereas the applied weights are from a statistical index that quantifies
the effect of different wavelet centre frequencies on feature extraction. An averaged
autocorrelation spectrum is adopted to highlight the feature characteristics related to bearing
health conditions. The performance of this proposed technique is examined by a series of
experimental tests corresponding to different bearing conditions. Test results show that this
new signal processing technique is an effective bearing fault detection method, which is
especially useful for non-stationary feature extraction and analysis.

Keywords: bearing fault detection, wavelet spectrum analysis, resonance feature

(Some figures in this article are in colour only in the electronic version)

1. Introduction Frequency-domain analysis may be the most commonly used

Rolling element bearings are widely used in various types
of machines ranging from simple electric fans to complex
manufacturing facilities. Bearing faults, in fact, are a common
cause of machinery failures. Therefore, an effective bearing
fault diagnostic technique is critically needed for a wide
array of industries for early detection of bearing defects
so as to prevent machinery performance degradation and
malfunction. Several methods have been proposed in the
literature for bearing fault detection, in which the analysis
can be performed in the time domain [1, 2], frequency domain
[3, 4] or time–frequency domain [5, 6]. In the time-domain
analysis, for example, a bearing fault is detected by monitoring
the variation of some statistical indices such as a root-mean-
square value, crest factor or kurtosis. A bearing is believed to
be damaged as long as monitoring indices exceed thresholds;
unfortunately, it is a challenging task to determine robust
thresholds because they may vary in different applications.
approach currently for bearing fault diagnostics. Bearing
health conditions are assessed by examining the fault-related
characteristic frequency components in a spectrum [7] or
in some extended spectral expressions such as bispectrum
or cepstrum maps [8, 9]. Frequency-based techniques,
however, are not suitable for the analysis of non-stationary
signals that are generally related to machinery defects. Non-
stationary or transient signals can be analysed by applying
time–frequency domain techniques such as the short-time
Fourier transform (FT) [10], Wigner–Ville distribution [11]
or wavelet transform (WT) [12]. In fault diagnostics, the
WT is the most favourite time–frequency domain technique,
because it does not contain cross terms such as those in the
Wigner–Ville transform, and can provide a more flexible multi-
resolution solution than the short-time FT. According to signal
decomposition paradigms, the WT can be classified as the
continuous WT [13], discrete WT [14], wavelet packet analysis
[15], and those WT with post-processing schemes such as

0957-0233/08/015105+09$30.00 1 © 2008 IOP Publishing Ltd Printed in the UK

Meas. Sci. Technol. 19 (2008) 015105
the singularity analysis [16], the FT [17, 18] and the energy
density analysis [19]. Even though a wavelet packet map can
provide more information, it is usually difficult to explain the
results especially for machinery condition analysis because
most mechanical systems are usually complex in structure and
operate under noisy or uncertain environment. In machinery
condition monitoring, the bearing fault diagnostics is one of the
most challenging tasks because it is not a simple mechanical
component, but a composite system that consists of inner and
outer rings as well as many rolling elements; each bearing
rotary component generates a vibratory signal. The vibration
signatures generated by an incipient bearing fault are usually
weak in magnitude and non-stationary in nature. As a result,
it is usually difficult to extract and enhance characteristic
features related to bearing health conditions. The objective
of this work is to develop a new signal processing technique
to approach this challenge; the goal is to provide a wide array
of industries a more effective technique and tool for incipient
bearing fault diagnostics. The proposed technique is new in
the following aspects: (1) a strategy is proposed to effectively
demodulate resonance features related to bearing faults; (2) a
weighted Shannon function is suggested to synthesize wavelet
(coefficient) functions and to enhance feature characteristics;
and (3) an averaged autocorrelation spectrum is employed to
highlight bearing fault characteristic frequencies.
The rest of this paper is organized as follows. The
proposed bearing fault detection technique is systematically
described in section 2. A comparison study is taken in section 3
to verify the effectiveness of the proposed technique based
on the experimental tests. Some concluding remarks are
summarized in section 4.
2. The theory of the wavelet spectrum analysis
Whenever a fault happens on a bearing component, stationary
and/or non-stationary impacts are generated, which excite the
bearing and its support structures. As a result, resonance
features are generated, which are usually buried in other
high-amplitude vibration signals. In the proposed wavelet
spectrum analysis technique, the first step is to demodulate
these resonance features; the next step is to synthesize
the resulting wavelet coefficient functions by applying the
proposed weighted Shannon function; the third step is to
construct the autocorrelation spectrum of the synthesized
function to highlight bearing characteristic frequencies. An
averaging process is taken finally in the frequency domain to
eliminate some random noise related to variations in operating
conditions.
It should be stated that even though the analysis in this
work is focused on bearing fault diagnostics, the proposed
technique can also be applied for fault detection in other
machinery components, such as gears and shaft systems.
2.1. Determination of wavelet functions
For a rolling element bearing, usually the defect occurs on the
fixed ring race first because the fixed ring material is subjected
to more dynamic load cycles. Consider a general bearing
2
J Liu et al
with a fixed outer ring, and suppose that a defect (e.g., a
fatigue pit) has happened on the outer ring race. Each time
a rolling element rolls over the pit, an impulse is generated
due to the impact. This impulse excites vibration resonance
of the bearing and the surrounding structures. In theory, the
excited transient modes due to an outer race defect do not
vary because the defect angular position remains the same as
each impact occurs, as long as there exists no slip among the
bearing components. On the other hand, for a rolling element
fault or an inner race defect, the generated impulse transient
modes may change in properties because the impact occurs at
a different angular position as the bearing components rotate.
As a result, the magnitudes of the impulse transients and the
excited resonance modes vary over time. Therefore, it is more
suitable to apply the WT that has a better time resolution at
high frequencies than at low ones [20], to decompose these
resonance features.
Given a time signal x(t), the wavelet coefficients are
determined by

+∞
√
W Tx (t, s) = x(τ ) sw ∗ (−s(t − τ )) dτ (1)
−∞
where w ∗ (t) denotes the complex conjugation of mother
wavelet function w(t); s and t are the scale and time variables,
respectively, which produce dilation and translation [12].
The choice of an appropriate mother wavelet depends on
the signal itself and the purpose of the analysis. In bearing
fault detection, the interest is to analyse the resonance features
induced by a localized bearing fault. Therefore, a mother
wavelet should possess a similar characteristic as in a fault-
related transient. Through a series of tests and comparisons
of the generally used wavelet functions, it has been found that
the Morlet wavelet gives superior results in this application,
which is a modulated Gaussian function [12]
 
t2
w(t) = exp − 2 exp(j2πf0 t) (2)
2b0
where b0 is the spread of the Gaussian function and f0 is the
wavelet centre frequency. Given a basic wavelet, the following
admissibility should be satisfied,

+∞
|W (f )|2
0< df < ∞ (3)
−∞ |f |
where W (f ) is the FT of w(t).
Because w(t) decays over time, the admissibility (3) is
+∞
equivalent to the requirement −∞ w(t) dt = 0 [20]. Strictly
speaking, the Morlet wavelet does not satisfy this zero-mean
requirement. However, the mean can become infinitely small
if the term b0 f0 is sufficiently large. As b0 f0 increases, the
duration of the wavelet expands, and the time resolution will
decrease correspondingly. As a result, the obtained mother
wavelet w(t) may not be appropriate for localizing fast-
decaying transient signatures. Based on the above reasoning,
the product of the spread and the scaled centre frequency is
kept as a constant in this work, that is,
b0 2πb0 f0 1
bi fi = (f0 si ) = b0 f0 = =√ (4)
si 2π 2 ln 2
√
where 2πb0 f0 = π 2/ln 2 was given in [21]; si represents the
ith selected scale; bi and fi are, respectively, the corresponding
Meas. Sci. Technol. 19 (2008) 015105
ith spread and centre frequency. Based on the relation
between bi andfi as in (4), the mean of the obtained mother
wavelet w(t) remains at a level as low as 10−12 , and the
effective support will vary with the scaled centre frequency
to accommodate the variation of the signatures of interest.
The FT of the mother wavelet w(t) in (2) is given by
√  
W (f ) = b0 2π exp −2b02 π 2 (f − f0 )2 . (5)
Thus, the FT of the dilated wavelet w(st) is
√   discretized points of signal x(t); β is the selected logarithm
Ws (f ) = bi 2π exp −2bi2 π 2 (f − fi )2 . (6)
For the chosen mother wavelet, the following equation
holds, w ∗ (−st) = w(st). Therefore, (1) can be implemented
by using
√
W Tx (t, s) = sF −1 [X(f )Ws (f )] (7)
−1
where F [·] denotes the inverse FT, X(f ) is the FT of x(t),
and Ws (f ) is given by (6). In the implementation of (7),
X(f ) is calculated only once, while Ws (f ) is calculated for
all values of the scale variable s.
2.2. The strategy for locating wavelet centre frequencies
In order to implement the WT over a designated frequency
band, a proper deployment of centre frequencies of the wavelet
is required. Based on (6), the 3 dB bandwidth, BWi , for the
ith centre frequency fi is derived as
BWi = [1 − λ, 1 + λ]fi (8)
where λ = ln 2
√
2π
is a constant. The overall frequency band
of the interest can be chosen as [Nft , fs /2.56], whereft is the
shaft speed, N is the order of shaft harmonics to be considered,
fs is the sampling frequency, and the constant 2.56 is related to
the power of 2. In order to reduce the interference effects from
the shaft harmonics, N should be sufficiently large (e.g., N =
35 in this work). Beginning with the lower bound frequency
Nft , the centre frequencies can be recursively calculated by
(1 + λ)i−1
fi = Nft , i = 1, 2, . . . , m − 1 (9)
(1 − λ)i
 
1 fs
fi = + fi−1 (1 + λ) , i=m (10)
2 2.56
where m is the number of selected scales. The bandwidth
between fm−1 and the upper bound frequency fs /2.56 may be
smaller than the 3 dB bandwidth offm , thus a middle point is
utilized in (10).
2.3. A weighted Shannon function for synthesizing wavelet
coefficient functions
Once wavelet coefficient functions for a cascade of dilated
wavelets are obtained, the next step is to properly synthesize
these coefficient functions for a low-dimension feature
representation to facilitate real-time monitoring applications.
In this work, a new technique is proposed, based on an
extended Shannon entropy function, to synthesize the resulting
wavelet coefficients among different bands. The Shannon
entropy in information theory is a measure of the degree of the
3
J Liu et al
uncertainty associated with a random variable. In this study, a
modification is made to this function in order to highlight
the fault-related features and to minimize the interference
information,
m
h(tk ) = [W Tx (tk , si ) logβ W Tx (tk , si )] (11)
i=1
where i = 1, 2, . . . , m; k = 1, 2, . . . , n; n is the number of
base that may vary from 2 to 10 (β = 2 in this study); and the
normalized wavelet coefficient W Tx (tk , si ) is given by
|W Tx (tk , si )|
W Tx (tk , si ) = (12)
σi
where σi represents the spread of |W Tx (t, si )|.
This technique highlights the coefficients W Tx (t, si )
whose magnitudes are larger than βσi , and diminishes the
others. The high magnitudes of coefficients, namely the fault-
related features, are derived from the resonance signatures
that closely ‘match’ the dilated wavelets. Usually, the fault-
related features can be reflected over several wavelet bands.
Therefore, the synthesis of these wavelet coefficient functions
by applying the extended Shannon function can integrate the
contributions from different frequency bands and enhance the
fault feature characteristics. The criterion for determining such
a contribution depends on the selection of the logarithm base
β; a wavelet coefficient with a magnitude larger than βσi is
considered as a contribution to the feature extraction process.
Because the effect of the WT on the feature extraction
differs from one centre frequency to another, an appropriate
indicator is needed to represent such an effect. The
characteristic defect frequency of a bearing is usually carried
by resonance features generated by the impact as a rolling
element encounters a defect. However, similar types of
features may also be generated by other vibratory sources
such as inherently varying compliance, misalignment and
imbalance. When applying the WT to process the overall
vibration signal from a faulty bearing, the coefficients
|W Tx (t, s)| obtained from the non-defect-related resonance
signatures (or noise) usually exhibit the characteristics of lower
amplitude and longer duration (counting for the major part of
the total energy). On the other hand, the WT coefficients
from the defect-related resonance signatures usually exhibit
the characteristics of higher amplitude but with a much shorter
duration (accounting for a minor part of the overall energy).
Fault detection is to extract the representative features related
to machinery health conditions, from a collected signal. From
a signal property standpoint, if a bearing is damaged, the
magnitude distribution of the wavelet coefficients should have
a longer tail, which indicates that fault-related features buried
in the resonance signatures have been enhanced. Therefore,
the effect of a dilated wavelet (with a varying centre frequency)
to the feature extraction can be characterized by a skewness
measure of the probability distribution function (pdf) of the
resulting wavelet coefficients, that is,
E[(|W Tx (t, si )| − µi )3 ]
γi = (13)
σi3
Meas. Sci. Technol. 19 (2008) 015105
Figure 1. Experimental setup. (1) encoder; (2) variable speed control; (3) enclosure; (4) motor; (5) flexible coupling; (6) ICP
accelerometer; (7) bearing housing; (8) tested bearing; (9) hardened shaft; (10) load disc; (11) belt drive; (12) alignment adjustor; (13) base;
(14) magnetic load system; (15) bevel gearbox.
where E[·] denotes the expectation function, µi and
σi represent the mean and the spread of |W Tx (t, si )|,
respectively.
The suggested index γi is employed to measure the
capability of the WT in extracting fault-related features, which
can be implemented into the extended Shannon function (11),
m SET 3 (mm)
H (tk ) = [γi W Tx (tk , si ) logβ W Tx (tk , si )]
i=1
2.4. Averaged autocorrelation spectrum
The proposed autocorrelation spectrum analysis represents
two processes: autocorrelation for the synthesized wavelet
coefficients H (t) to enhance the involved periodic features,
and spectral analysis (FT) for periodic feature extraction.
Specifically,
rxx (l) = E[H (t)H ∗ (t + l)], l = 0, 1, 2, . . . , n − 1. (15)
R(f ) = F [rxx (l)]
where l is the lag index and F [·] denotes the FT.
Autocorrelation spectrum is defined as
(f ) = R(f )R ∗ (f ).
In implementation, an averaged spectrum, over five
segments of measured signals, should be used in order
to eliminate random noise. The proposed techniques are
implemented in MATLAB environment. Using a Pentium
IV computer (with 1.8 GHz and 512 MB RAM), the total
computational time is around 3.25 min for one segment of
vibration signature (491 520 sample points). Bearing health
conditions are estimated by analysing the related characteristic
frequency components of the resulting spectrum, as illustrated
in the next section.
3. Performance evaluation
The proposed wavelet spectrum analysis technique will be
applied in this section for bearing fault detection, although
J Liu et al
Table 1. Fault type and size.
Outer Inner Rolling
race race element
defect defect defect
SET 1 (mm) 0.35 0.41 0.29
SET 2 (mm) 0.39 0.47 0.35
0.44 0.56 0.38
(14)
it can also be used for fault detection in other machinery
components such as gears and shafts. Its performance will
be compared with several related classical methods, such as
the one-scaled WT [17], the envelope analysis [7] and the max-
envelope approach [18]. In the max-envelope approach, only
the maximum wavelet coefficient is selected as the resultant
value.
3.1. Experimental setup and measurement
In order to verify the effectiveness of the proposed technique
(16) in bearing fault detection, a large number of tests has been
conducted. The experimental setup employed for the tests is
shown in figure 1. The system is driven by a 3-hp induction
motor, with the speed range from 20 to 4200 rpm. The
(17) shaft rotation speed can be controlled by a speed controller.
An optical sensor is used for shaft speed measurement. A
flexible coupling is utilized to damp out the high-frequency
vibration generated by the motor. Two ball bearings are fitted
in the solid housings. Accelerometers are mounted on the
housing of the tested bearing to measure the vibration signals
along two directions. Considering the structure properties,
the signal measured vertically is utilized for analysis in this
work, whereas the information of the signal measured from
the horizontal direction is used for verification. The static
load is applied by two disks, and a variable load is applied by
a magnetic brake system through a bevel gearbox and a belt
drive. A data acquisition board (NI PCI-4472) is employed for
signal collection, which has built-in anti-aliasing filters with
the cut-off frequency set at half of the sampling rate.
In the tests, ball bearings of type MB ER-10K are
utilized. Four bearing health conditions are considered:
4
Meas. Sci. Technol. 19 (2008) 015105 J Liu et al

healthy bearings, bearings with outer race defects, bearings

(a )
with inner race defects, and bearings with rolling element 0.3

Acceleration (V)
faults. Three sets of faulty bearings with different fault sizes
are prepared. The fault types and the corresponding fault 0
dimension (in diameter) are listed in table 1. Each bearing is
tested under seven different shaft speeds (900 rpm, 1200 rpm, -0.3
0 1 2 3
1500 rpm, 1800 rpm, 1920 rpm, 2100 rpm and 2400 rpm) and
(b)
two load levels (1.2 Nm and 2.3 Nm), respectively. When a 0.4

Acceleration (V)
bearing is running at a specific shaft speed and load level, five
segments of vibration signals are collected at a time interval of 0
2 h, and the time span for each measurement is 24 s. Based on
the comprehensive investigation, the proposed technique has -0.4
0 1 2 3
demonstrated its superiority to other three classical techniques.
A few examples from two study cases (cases I and II) are given (c)
0.4

Acceleration (V)
in the following section to illustrate the examination results,
where case I refers to bearings in SET 1 operating at shaft 0
speed 32 Hz with load 1.2 Nm, and case II represents bearings
in SET 2 operating at shaft speed 40 Hz with load 2.3 Nm. -0.4
0 1 2 3
As stated in section 2, the bearing fault is diagnosed
(d )
by examining the related characteristic frequencies and their 0.4
harmonics. For a bearing with a fixed outer ring, the Acceleration (V)

characteristic frequency for an inner race defect [8] is 0

 
Zft d
fid = 1 + cos α (18) -0.4
2 D 0 1 2 3
Time Sample Span (X 10000)
where Z is the number of rolling elements, d is the rolling
element diameter, D is the bearing pitch diameter, α is the Figure 2. Vibration signals: (a) from a healthy bearing; (b) from an
contact angle (α = 0 in this case) and ft is the shaft speed. outer race defect bearing; (c) from an inner race defect bearing; (d)
from a rolling element damaged bearing.
The characteristic frequency for a rolling element defect
is
  coefficients H (t) corresponding to different bearing health
Dft d2
fed = 1 − 2 cos α .
2
(19) conditions. It can be seen that the bearing condition-related
d D features are extracted from those weak vibration signatures.
The characteristic frequency for an outer race defect is In figure 3(a), for a healthy bearing, the features that are
  related to the shaft rotation speed (power input) are extracted.
Zft d
fod = 1 − cos α . (20) The consecutive features are spaced by an interval of 640
2 D
sample data, or with the repetition rate of 32 Hz for a 20 480
For the tested bearings (MB ER-10K) with fixed outer Hz sampling frequency; this characteristic defect frequency
races, if the shaft speed is ft = 32 Hz, for example, the can be accurately obtained by calculating the averaged
characteristic defect frequencies are fod = 97.67 Hz, fid = autocorrelation spectrum from the synthesized wavelet
158.33 Hz and fed = 127.48 Hz. Likewise, if the shaft speed coefficients, which is discussed in the following section. It
is ft = 40 Hz, the characteristic defect frequencies are fod = is also seen that the features from outer race defects are
122.09 Hz, fid = 197.91 Hz and fed = 159.34 Hz. stationary and easy to extract, as clearly shown in figures 3(b)
and (c). However, as seen in figures 3(d)–(g), extracting
3.2. Performance evaluation the representative features from inner race defects or rolling
element defects is very challenging; the features are non-
In the following comparison study, the envelope analysis is stationary, and the magnitudes of the features vary significantly
applied to the signal bandpass filtered around the system’s due to the nature of the corresponding bearing faults. It is also
resonant frequency [7], [1500 2500] Hz, while the one-scaled worth mentioning that for a ball bearing, the defect on a ball
WT is performed with the centre frequency at 2000 Hz, encounters the inner/outer race randomly; correspondingly, it
the middle of the system’s resonant frequency band. In is necessary to take a sufficient time period in data collection;
implementing the proposed technique, N is chosen to be 35 to in this work, the time span for each measurement is 24 s.
reduce the effects from the shaft harmonics, and the frequency
band for the proposed technique is designated as [1000 8000]
Hz in which the upper bound frequency is computed from 3.2.1. Healthy bearing. Bearing components generate
fs /2.56. vibratory signals that are related to the speed of shaft rotation
Figure 2 shows parts of measured vibration signals at (the power input). Figure 4 shows the processing results for a
shaft speed 1920 rpm with the sampling frequency 20 480 Hz. healthy bearing by using the related data analysis techniques.
Figure 3 plots some examples of the synthesized wavelet It is seen from figure 4(a) that, based on the proposed wavelet

5
Meas. Sci. Technol. 19 (2008) 015105 J Liu et al

information can be used not only as a health indicator of the

500
(a ) 640 samples (32 Hz) bearing of interest, but also as a digital encoder in system
condition monitoring and control applications.
0 This shaft speed indicator can also be recognized on the
maps from the max-envelope approach (figure 4(b)), even
500
210 samples (97.7 Hz) though its resolution is lower than in figure 4(a) due to the
(b)
presence of some other spectral components. The results from
0 the one-scaled WT (figure 4(c)) and the envelope analysis
(figure 4(d)), however, may lead to an incorrect diagnosis
500 (i.e., a false alarm) because the third harmonic of shaft speed
(c) 168 samples (122.1 Hz)
(96 Hz) is very close to the outer race defect characteristic
0
frequency (fod = 97.67 Hz).

500
(d ) 129 samples (158.3 Hz) 3.2.2. Outer race fault detection. If the bearing is damaged,
H (tk )

in theory, the corresponding characteristic defect frequency

0 and/or its harmonics appear in the spectrum. As stated earlier,
outer race fault detection is a relatively easier task because the
500
103 samples (197.9 Hz) defect resonance modes do not change dramatically. Figure 5
(e)
shows the processing results in test case I using the related
0 techniques. It is seen that each technique can detect the
presence of this bearing fault, but the proposed wavelet
500 spectrum analysis (figure 5(a)) and the max-envelope approach
(f ) 161 samples (127.5 Hz)
(figure 5(b)) provide more positive fault detection information
0
in case I because of their feature enhancement effects. It is
also seen that the spectral component related to shaft speed
500 disappears in this case because its spectral magnitude is much
128 samples (159.3 Hz)
(g) lower than the defect frequency components. In addition,
0
the second harmonics of the defect characteristic frequency
0 500 1000 1500 are highlighted in case I by the one-scale WT (figure 5(c))
Time Sample Span and the envelope analysis (figure 5(d)), respectively.
Figure 3. Synthesized coefficients H (t) where the repetition of the
Figure 6 shows the processing results corresponding to test
bearing condition related features is indicated using sampling case II. Similar comparison results can be reached as those in
intervals and the features are pointed by arrows: (a) from a healthy figure 5.
bearing; (b) from an outer race defect bearing in case I; (c) from an
outer race defect bearing in case II; (d) from an inner race defect
bearing in case I; (e) from an inner race defect bearing in case II; 3.2.3. Inner race fault detection. As discussed in section 2,
( f ) from a rolling element damaged bearing in case I; (g) from a the detection of a fault on an inner race and a rolling element
rolling element damaged bearing in case II. is more challenging than on a fixed ring because the modes
of the generated impulse resonance features vary over time.
spectrum technique, the dominant spectral component is the Figures 7 and 8 show the processing results corresponding
shaft speed (ft = 32 Hz). This high-resolution shaft speed to the bearing signals with inner race defects. The

-4
(a ) x 10 (c )
400 4
shaft speed
Spectrum

shaft speed
200 2

0 0

-4 -12
x 10 (b) x 10 (d )
3 1.5
Spectrum

shaft speed
2 1 shaft speed

1 0.5

0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Frequency (Hz) Frequency (Hz)

Figure 4. Healthy bearing processing results: (a) using the proposed method; (b) using the max-envelope approach; (c) using the one-scaled
WT; (d) using the envelope analysis.

6
Meas. Sci. Technol. 19 (2008) 015105 J Liu et al

-3
(a ) x 10 (c)
1000 3
defect frequency

Spectrum
defect frequency 2
500
1

0 0

-11
(b) x 10 (d )
0.01 4
defect frequency
Spectrum

defect frequency
0.005 2

0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Frequency (Hz) Frequency (Hz)

Figure 5. Outer race fault detection in case I: (a) using the proposed method; (b) using the max-envelope approach; (c) using the one-scaled
WT; (d) using the envelope analysis.

-3
(a ) x 10 (c )
100 3
defect frequency
Spectrum

defect frequency 2
50
1

0 0

-4 (b) -9 (d )
x 10 x 10
4 1
defect frequency
Spectrum

defect frequency
2 0.5

0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Frequency (Hz) Frequency (Hz)

Figure 6. Outer race fault detection in case II: (a) using the proposed method; (b) using the max-envelope approach; (c) using the
one-scaled WT; (d) using the envelope analysis.

-3
(a ) x 10 (c)
10000 2
defect frequency defect frequency
Spectrum

5000 1

0 0

-4 -10
x 10 (b) x 10 (d )
4 3
defect frequency defect frequency
Spectrum

2
2
1

0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Frequency (Hz) Frequency (Hz)

Figure 7. Inner race fault detection in case I: (a) using the proposed method; (b) using the max-envelope approach; (c) using the one-scaled
WT; (d) using the envelope analysis.

characteristic frequency of inner race defect (fid = 158.33 Hz
in case I; fid = 197.91 Hz in case II) can be identified
clearly in figures 7(a) and 8(a) using the proposed wavelet
spectrum technique. However, when applying the max-
envelope approach for the synthesis of the wavelet coefficient
functions, the defect frequency magnitude, as shown in
figures 7(b) and 8(b), becomes the third or fourth highest
spectral component; this makes it difficult to detect the bearing
fault based only on the spectral analysis. Similar results can be
seen in figures 7(c), (d), and figures 8(c), (d) by using the one-
scaled WT and the envelope analysis, respectively. Unclear
processing results may lead to false alarms or missed alarms
in real-time monitoring applications.

7
Meas. Sci. Technol. 19 (2008) 015105 J Liu et al

-3
(a ) x 10 (c)
4000 3

Spectrum
defect frequency
2 defect frequency
2000
1

0 0
-4 -9
x 10 (b) x 10 (d )
4 1.5
Spectrum

defect frequency
1 defect frequency
2
0.5

0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Frequency (Hz) Frequency (Hz)

Figure 8. Inner race fault detection in case II: (a) using the proposed method; (b) using the max-envelope approach; (c) using the one-scaled
WT; (d) using the envelope analysis.

-4
(a ) x 10 (c )
200 1.5
shaft speed
Spectrum

defect frequency 1 defect frequency

100
0.5

0 0

-3 -12
x 10 (b) x 10 (d )
1 2
Spectrum

defect frequency
0.5 defect frequency 1

0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Frequency (Hz) Frequency (Hz)

Figure 9. Rolling element fault detection in case I: (a) using the proposed method; (b) using the max-envelope approach; (c) using the
one-scaled WT; (d) using the envelope analysis.

-5
(a ) x 10 (c)
600 6
Spectrum

400 defect frequency 4 defect frequency

200 2

0 0

-4 -12
x 10 (b) x 10 (d )
4 4
Spectrum

defect frequency defect frequency

2 2

0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Frequency (Hz) Frequency (Hz)

Figure 10. Rolling element fault detection in case II: (a) using the proposed method; (b) using the max-envelope approach; (c) using the
one-scaled WT; (d) using the envelope analysis.

3.2.4. Rolling element fault detection. Figures 9 and
10 illustrate two processing results corresponding to rolling
element defects in test case I and case II, respectively. It is
clear that the proposed wavelet spectrum technique is the only
method that can recognize the existence of the bearing faults
from weak resonance signatures (fed = 127.48 Hz in case I;
fed = 159.34 Hz in case II). The other three methods have
missed the bearing faults under these test conditions.
In general, 42 different test cases (3 fault dimensions,
7 shaft speeds and 2 load levels) have been examined. Also
in each case, three bearing fault conditions are considered: an
outer race defect, an inner race defect and a rolling element
defect. Based on this comprehensive investigation, it is found

8
Meas. Sci. Technol. 19 (2008) 015105
that it is usually difficult for classical methods to clearly
recognize characteristic defect frequencies from complex and
non-stationary features such as those generated by a bearing
incipient inner race defect or an incipient rolling element fault.
The proposed wavelet spectrum technique is superior in this
aspect to the related classical techniques for bearing fault
diagnostics due to its feature enhancement characteristics.
4. Conclusion
A new signal processing technique, wavelet spectrum analysis,
has been proposed in this paper for incipient bearing
fault diagnostics. A novel strategy is suggested for the
deployment of wavelet centre frequencies. To enhance feature
characteristics, a weighted Shannon function is proposed to
synthesize the wavelet coefficient functions obtained over
the designated frequency bands, whereas the applied weights
are from a statistical index to quantify the effect of wavelet
centre frequencies on the feature extraction. An averaged
autocorrelation power spectrum is adopted to highlight
bearing characteristic features in the spectrum. A series of
experimental tests corresponding to different bearing health
conditions have demonstrated the superior capability of the
proposed wavelet spectrum technique to the related classical
bearing fault detection methods, especially in the aspect of
non-stationary feature extraction and analysis.
Acknowledgments
This project was partly supported by MC Technologies Inc. Jie
Liu is supported by the Graduate Scholarship of the Natural
Sciences and Engineering Council of Canada (NSERC).
References
[1] Tandon N 1994 A comparison of some vibration parameters
for the condition monitoring of rolling element bearings
Measurement 12 285–9
[2] Heng R B W and Nor M J M 1998 Statistical analysis of sound
and vibration signals for monitoring rolling element bearing
condition Appl. Acoust. 53 211–26
[3] Randall R B, Antoni J and Chobsaard S 2001 The
relationship between spectral correlation and envelope
analysis in the diagnostics of bearing faults and other
cyclostationary machine signals Mech. Syst. Signal Process.
15 945–62
J Liu et al
[4] Stack J R, Habetler T G and Harley R G 2006 Fault-signature
modeling and detection of inner-race bearing faults IEEE
Trans. Indust. Appl. 42 61–8
[5] Du Q and Yang S 2006 Improvement of the EMD method and
applications in defect diagnosis of ball bearings Meas. Sci.
Technol. 17 2355–61
[6] Luo G Y, Osypiw D and Irle M 2003 On-line vibration
analysis with fast continuous wavelet algorithm for
condition monitoring of bearing, J. Vib. Control
9 931–47
[7] Mcfadden P D and Smith J D 1984 Vibration monitoring of
rolling element bearings by the high-frequency resonance
technique—a review Tribol. Int. 17 3–10
[8] Stack J R, Harley R G and Habetler T G 2004 An amplitude
modulation detector of fault diagnosis in rolling element
bearings IEEE Trans. Indust. Electron. 51 1097–102
[9] Choi Y-C and Kim Y-H 2007 Fault detection in a ball bearing
system using minimum variance cepstrum Meas. Sci.
Technol. 18 1433–40
[10] Kaewkongka T, Au Y H J, Rakowski R T and Jones B E 2003
A comparative study of short time Fourier transform and
continuous wavelet transform for bearing condition
monitoring Int. J. COMADEM 6 41–8
[11] Kim B S, Lee S H, Lee M G, Ni J, Song J Y and Lee C W
2007 A comparative study on damage detection in speed-up
and coast-down process of grinding spindle-typed
rotor-bearing system J. Mater. Process. Technol. 187 30–6
[12] Wang W, Ismail F and Golnaraghi F 2001 Assessment of gear
damage monitoring techniques using vibration
measurements Mech. Syst. Signal Process. 15 905–22
[13] Wang W, Ismail F and Golnaraghi F 2004 A neuro-fuzzy
approach for gear system monitoring IEEE Trans. Fuzzy
Syst. 12 710–23
[14] Li X and Yao X 2005 Multi-scale statistical process monitoring
in machining IEEE Trans. Indust. Electron. 52 924–27
[15] Eren L and Devaney M J 2004 Bearing damage detection via
wavelet packet decomposition of the stator current IEEE
Trans. Instrum. Meas. 53 431–6
[16] Sun Q and Tang Y 2002 Singularity analysis using continuous
wavelet transform for bearing fault diagnosis Mech. Syst.
Signal Process. 16 1025–41
[17] Wang C and Gao R X 2003 Wavelet transform with spectral
post-processing for enhanced feature extraction IEEE
Trans. Instrum. Meas. 52 1296–301
[18] Nikolaou N G and Antoniadis I A 2002 Demodulation of
vibration signals generated by defects in rolling element
bearings using complex shifted morlet wavelets Mech. Syst.
Signal Process. 16 677–94
[19] Cheng J, Yu D and Yang Y 2005 Time-energy density analysis
based on wavelet transform NDT&E Int. 38 569–72
[20] Daubechies I 1990 The wavelet transform, time-frequency
localization and signal analysis IEEE Trans. Inf. Theory
36 961–1005
[21] Strang G and Nguyen T 1996 Wavelets and Filter Banks
(USA: Wellesley-Cambridge Press) pp 69–79