You are on page 1of 12

ARTICLE IN PRESS

Mechanical Systems
and
Signal Processing
Mechanical Systems and Signal Processing 19 (2005) 259–270
www.elsevier.com/locate/jnlabr/ymssp

Application of EMD method and Hilbert spectrum to the


fault diagnosis of roller bearings
Dejie Yu*, Junsheng Cheng, Yu Yang
College of Mechanical and Automotive Engineering, Hunan University, Changsha 410082, People’s Republic of China
Received 25 February 2003; received in revised form 8 July 2003; accepted 16 July 2003

Abstract

Based upon empirical mode decomposition (EMD) method and Hilbert spectrum, a method for the fault
diagnosis of roller bearings is proposed in this paper. The local Hilbert spectrum and local Hilbert marginal
spectrum are introduced. The orthogonal wavelet bases are used to translate vibration signals of a roller
bearing into time-scale representation. Then, an envelope signal can be obtained by envelope spectrum
analysis of wavelet coefficients of high scales. By applying EMD method and Hilbert transform to the
envelope signal, we can get the local Hilbert marginal spectrum from which the faults in a roller bearing can
be diagnosed and fault patterns can be identified. Practical vibration signals measured from roller bearings
with out-race faults or inner-race faults are analysed by the proposed method. The results show that the
proposed method is superior to the traditional envelope spectrum method in extracting the fault
characteristics of roller bearings.
r 2003 Elsevier Ltd. All rights reserved.

Keywords: Roller bearing; EMD; Hilbert spectrum; Local Hilbert marginal spectrum; Wavelet bases; Envelope
analysis

1. Introduction

While operating a roller bearing with local faults impulse is created, the high-frequency shock
vibration is then generated and the amplitude of vibration is modulated by the pulse force. The
envelope analysis method provides an important and effective approach to analyse the fault
signals of high-frequency impact vibration, it has been applied to the fault diagnosis of roller
bearings successfully [1–3]. However, in the traditional envelope analysis method, the fault is
identified through the peak value of envelope spectrum. Thus, this traditional method has two
aspects of disadvantages. On the one hand, FFT method is widely used in the spectrum analysis of

*Corresponding author.
E-mail address: cjssignal@163.net (D. Yu).

0888-3270/$ - see front matter r 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0888-3270(03)00099-2
ARTICLE IN PRESS

260 D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270

envelope signals; however, it could only give the global energy-frequency distributions and fail to
reflect the details of a signal. So it is hard to analyse a signal effectively when the fault signal is
weaker than the interfering signal [4–6]. At the same time, it is easy to diffuse and truncate the
signal’s energy as FFT regards harmonic signals as basic components, which will lead to energy
leakage and cause lower accuracy. On the other hand, the central frequency of the filter is
determined with experience while forming an envelope signal, which will make great subjective
influence on the results [6–8]. Aiming at these two disadvantages, we combined the envelope
analysis method with empirical mode decomposition (EMD, as defined in Section 1) and Hilbert
spectrum (as defined in Section 2) in the fault diagnosis of roller bearings, and the concepts of
local Hilbert spectrum and local Hilbert marginal spectrum based on EMD are introduced. EMD
method is based on the local characteristic time scales of a signal and could decompose the
complicated signal function into a number of intrinsic mode functions (IMFs, as defined in
Section 1) [9,10]. Frequency components contained in each IMF not only relates to the sampling
frequency, but also changes with the signal itself. Furthermore, the corresponding Hilbert
spectrum will not lead to energy diffusion and leakage. Therefore, EMD is a self-adaptive signal
processing method that can be applied to non-linear and non-stationary process perfectly.
In this paper, a method for fault diagnosis of roller bearings is proposed. To extract the fault
characteristics of the fault vibration signal of a roller bearing, the proposed method uses the
following four procedures. First, in order to decrease the influence of lower frequency noise, the
fault vibration signal of a roller bearing is decomposed by wavelet packet; second, a envelope
signal can be obtained with analysing wavelet coefficients of high-frequency band by using Hilbert
transform; third, IMFs can be acquired by applying EMD to the envelope signal; finally, we
choose some special IMFs to obtain local Hilbert marginal spectrum from which the faults in a
roller bearing can be diagnosed and fault patterns can be identified. Practical vibration signals
measured from roller bearings with out-race faults or inner-race faults are analysed by the
proposed method. The results show that the proposed method is superior to the traditional
envelope spectrum method in extracting the fault characteristics of roller bearings.

2. EMD method

EMD method is developed from the simple assumption that any signal consists of different
simple intrinsic modes of oscillations. Each linear or non-linear mode will have the same number
of extrema and zero-crossings. There is only one extremum between successive zero-crossings.
Each mode should be independent of the others. In this way, each signal could be decomposed
into a number of IMFs, each of which must satisfy the following definition [9]:
(1) In the whole data set, the number of extrema and the number of zero-crossings must either
equal or differ at most by one.
(2) At any point, the mean value of the envelope defined by local maxima and the envelope
defined by the local minima is zero.

An IMF represents a simple oscillatory mode compared with the simple harmonic function.
With the definition, any signal xðtÞ can be decomposed as follows [9]:
ARTICLE IN PRESS

D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270 261

(1) Identify all the local extrema, and then connect all the local maxima by a cubic spline line as
the upper envelope.
(2) Repeat the procedure for the local minima to produce the lower envelope. The upper and
lower envelopes should cover all the data between them.
(3) The mean of upper and low envelope value is designated as m1 ; and the difference between the
signal xðtÞ and m1 is the first component, h1 ; i.e.
xðtÞ  m1 ¼ h1 : ð1Þ
Ideally, if h1 is an IMF, then h1 is the first component of xðtÞ:
(4) If h1 is not an IMF, h1 is treated as the original signal and repeat (1), (2), (3); then
h1  m11 ¼ h11 : ð2Þ
After repeated sifting, i.e. up to k times, h1k becomes an IMF, that is
h1ðk1Þ  m1k ¼ h1k : ð3Þ
Then it is designated as
c1 ¼ h1k : ð4Þ
The first IMF component is obtained from the original data. c1 should contain the finest scale
or the shortest period component of the signal.
(5) Separate c1 from xðtÞ: We get
r1 ¼ xðtÞ  c1 ; ð5Þ
where r1 is treated as the original data and repeat the above processes. The second IMF
component c2 of xðtÞ could be got. Let us repeat the process as described above n times. Then
n-IMFs of signal xðtÞ can be got. Then,
r1  c2 ¼ r2
^ : ð6Þ
rn1  cn ¼ rn

The decomposition process can be stopped when rn becomes a monotonic function from which
no more IMFs can be extracted. By summing up Eqs. (5) and (6), we finally obtain
X
n
xðtÞ ¼ cj þ rn : ð7Þ
j¼1

Thus, one can achieve a decomposition of the signal into n-empirical modes and a residue rn ;
which is the mean trend of xðtÞ: The IMFs c1 ; c2 ; y; cn include different frequency bands
ranging from high to low. The frequency components contained in each frequency band are
different and they change with the variation of signal xðtÞ; while rn represents the central tendency
of signal xðtÞ:
To clarify the decomposition processes, Fig. 1 shows the procedure of EMD. Let us consider
the case of the sum of two sine waves and one amplitude modulation wave as
1
xðtÞ ¼ 2 sinð2p15tÞ þ 4 sinð2p10tÞ sinð2p10 tÞ þ sinð2p5tÞ tA½0; 1 ð8Þ
ARTICLE IN PRESS

262 D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270

Start

Input signal x(t )

r = x(t ), n = 1

x(t )

Identify all the local extrema in x(t )

Construct the upper envelope E1


and the lower envelope E2

m = E1 − E 2 x(t ) = h

h = x(t ) − m
x(t ) = r
h Is an IMF? No

Yes

n = n + 1, c(n) = h, r = r − c(n)

Is r a monotonic No
function?

Yes

End

Fig. 1. Flow chart for EMD.

Fig. 2. The waveform of xðtÞ:

with the waveform given in Fig. 2. The sample frequency is 1024 Hz. The four IMF components
after exerting EMD to xðtÞ are given in Fig. 3. The first components, c1 ; indeed shows 15 waves;
the second components, c2 ; is an amplitude modulation signal; the third components, c3 ;
ARTICLE IN PRESS

D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270 263

Fig. 3. The first four IMF components of xðtÞ:

shows 5 waves; while, the last components, c4 ; is of negligible energy. While the EMD
succeeded in extracting the intrinsic time scale of the processes, the components,
especially the second components, failed to give the exact amplitudes and frequencies, a flaw
due to the accumulation of errors from the imperfect spline fitting in the decomposition
processes.

3. Local Hilbert marginal spectrum

For one IMF ci ðtÞin Eq. (7), we can always have its Hilbert transform as
Z
1 N ci ðt0 Þ 0
H½ci ðtÞ ¼ dt : ð9Þ
p N t  t0
With this definition, we can have an analytic signal as
zi ðtÞ ¼ ci ðtÞ þ jH½ci ðtÞ ¼ ai ðtÞejFi ðtÞ ð10Þ
in which
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ai ðtÞ ¼ c2i ðtÞ þ H 2 ½ci ðtÞ; ð11Þ

H½ci ðtÞ
Fi ðtÞ ¼ arctan : ð12Þ
ci ðtÞ
From Eq. (12), we can have the instantaneous frequency as
dFi ðtÞ
oi ðtÞ ¼ : ð13Þ
dt
ARTICLE IN PRESS

264 D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270

After performing the Hilbert transform to each IMF component, the original signal can be
expressed as the real part (RP) in the following form:
X n X
n R
jFi ðtÞ j oi ðtÞ dt
xðtÞ ¼ RP ai ðtÞe ¼ RP ai ðtÞe : ð14Þ
i¼1 i¼1

Here we left out the residue rn on purpose, for it is either a monotonic function or a constant.
Eq. (14) gives both amplitude and frequency of each component as functions of time. This
frequency-time distribution of the amplitude is designated as the Hilbert spectrum Hðo; tÞ:
X n R
Hðo; tÞ ¼ RP ai ðtÞej oi ðtÞ dt : ð15Þ
i¼1

With the Hilbert spectrum defined, we can also define the marginal spectrum, hðoÞ; as
Z T
hðoÞ ¼ Hðo; tÞ dt; ð16Þ
0
where T is the total data length. The Hilbert spectrum offers a measure of amplitude contribution
from each frequency and time, while the marginal spectrum offers a measure of the total
amplitude contribution from each frequency.
After performing the Hilbert transform on some of the IMF components that we are interested
in, we can also define the local Hilbert spectrum by
R R
0 j oi ðtÞ dt j ok ðtÞ dt
H ðo; tÞ ¼ Reð? þ ai ðtÞe þ ? þ ak ðtÞe þ ?Þ ð17Þ
and the local Hilbert marginal spectrum by
Z T
0
h ðoÞ ¼ H 0 ðo; tÞ dt: ð18Þ
0
The H 0 ðo; tÞ offers a measure of the amplitude contribution from each time and some
frequencies, while the local marginal h0 ðoÞ spectrum offers a measure of the total amplitude
contribution from some frequencies that we are interested in.
The Hilbert spectrum of the signal given in expression (8) is presented in contour lines in Fig. 4,
in which the energy frequency distribution of the signal is shown. Fig. 5 gives the Hilbert marginal
spectrum. From Fig. 5, one can see that the peaks of the three components are clearly separated

Fig. 4. The Hilbert spectrum (in contour lines) of the signal given in expression (8).
ARTICLE IN PRESS

D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270 265

Fig. 5. The Hilbert marginal spectrum of the signal given in expression (8).

Fig. 6. The Fourier spectrum of the signal given in expression (8).

and the resolution ratio is higher. The Fourier spectrum of the same signal is presented in Fig. 6. It
is clear that there is severe energy leakage and the resolution ratio is lower.

4. Local Hilbert marginal spectrum approach based on the envelope signal of wavelet coefficients

The process of local Hilbert marginal spectrum approach based on the envelope signal of
wavelet coefficients is given below:
(1) In order to decrease the influence of lower frequency noise, wavelet coefficient dt in high-
frequency band is obtained by applying wavelet decomposition to the original signal xðtÞ:
(2) According to Eq. (9), H½dt is acquired by applying Hilbert transform to dðtÞ; and then the
envelope signal BðtÞ based on wavelet coefficients is obtained as
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
BðtÞ ¼ d2 ðtÞ þ H 2 ½dðtÞ: ð19Þ

(3) BðtÞ is decomposed by using EMD and IMF components as c1 ; c2 ; y; cn are obtained.
(4) Apply Hilbert transform to all IMF components according to Eqs. (9)–(13).
(5) Compute the instantaneous frequencies of all IMF components and choose useful
components cj ; y; ck from above IMF components according to the fault characteristic
frequency of the roller bearing.
ARTICLE IN PRESS

266 D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270

(6) According to Eqs. (17) and (18), analyse IMF components cj ; y; ck and obtain local marginal
spectrum h0 ðf Þ from which fault patterns can be identified.

Fig. 7 shows the flow chart of the process of local Hilbert marginal spectrum approach based on
the envelope signal of wavelet coefficients.

Start

Input signal x(t )

Apply wavelet decompose to x(t )

Extract the wavelet coefficient


d (t ) in high scale

1 ∞ d (τ )
H [ d ( t )] = π ∫ −∞ t − τ dτ

B (t ) = d 2 (t ) + H 2 [d (t )]

IMF components c1 , c 2 , ... , c n are


obtained after applying EMD to B (t )

Apply Hilbert transform to c1 , c 2 , ... , c n

Compute the fault characteristic frequency

Choose useful IMF components cj , ... , ck

Construct the local marginal


spectrum ( f , h ′( f ))

Extract the fault characteristic from


( f , h ′( f )) and identify the fault model

End

Fig. 7. The flow chart of the process of the proposed method.


ARTICLE IN PRESS

D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270 267

5. Applications

The high-frequency vibration amplitude of operating roller bearings with local faults was
modulated by pulse force. In order to obtain the fault characteristic frequencies, the vibration
signals of roller bearings need to demodulate. Fig. 8 illustrated the time domain waveform of an
acceleration signal picked up from the bearing seat of a 6311-type rolling bearing with out-race
fault. The experimental rotating frequency is 25 Hz; the sample frequency is 4096 Hz and the
characteristic frequency of the roller bearing with out-race fault is 76 Hz. In order to decrease the
influence of lower frequency noise, the vibration signal of the roller bearing is decomposed with
Daubechies 10 (D10) wavelet. The envelope signal BðtÞ is obtained by using Hilbert transform to
the wavelet coefficients of high scales. After EMD is applied to the envelope signal BðtÞ; BðtÞ is
decomposed into 20 IMF components, c1 ; c2 ; y; c20 ; which include different components from
lower to high frequency. All IMF components are exerted the Hilbert transform according to
Eqs. (9)–(13). After the instantaneous frequencies of IMF components are computed, we can
know that the fault characteristic frequency locates at c7 ; c8 ; c9 ; c10 ; c11 : Therefore, the local
marginal spectrum h0 ðf Þ corresponding to IMF components c7 ; c8 ; c9 ; c10 ; c11 as Fig. 9 illustrated
can be obtained according to Eqs. (17) and (18). In Fig. 9, one can see spectrum lines at the
characteristic frequency foc½1 ð76 HzÞ of the roller bearing with out-race fault and its double
frequency 152 Hz. Figs. 10 and 11 illustrated the envelope spectrum of the same data by the
traditional Hilbert envelope analytic method with band filtering (the pass band is [1000, 2000]Hz
in Fig. 10, while [1500, 2000]Hz in Fig. 11). None of them shows the spectrum line at the
characteristic frequency of the roller bearing with out-race fault.

Fig. 8. Accelerative vibration signal of a roller bearing with outer-race fault.

Fig. 9. The local marginal spectrum of the signal shown in Fig. 8.


ARTICLE IN PRESS

268 D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270

Fig. 10. The envelope spectrum of the signal shown in Fig. 8 by the traditional Hilbert envelope analytic method with
band filtering (the pass band is [1000, 2000]Hz).

Fig. 11. The envelope spectrum of the signal shown in Fig. 8 by the traditional Hilbert envelope analytic method with
band filtering (the pass band is [1500, 2000]).

Fig. 12. Accelerative vibration signal of a roller bearing with inner-race fault.

Fig. 12 illustrated the time domain waveform of the acceleration vibration signal of a
roller bearing with inner-race fault. The rotating frequency is 20 Hz and the characteristic
frequency of the roller bearing with inner-race fault is 99.2 Hz. Analysing the original signals
with the same procedure, we can get 21 IMFs c1 ; c2 ; y; c21 : After the instantaneous
frequencies of IMFs are computed, we can know that the fault characteristic frequency
and its multiplication frequencies locate at c4 ; c5 ; c6 ; c7 ; c8 ; c9 : The local marginal spectrum h0 ðf Þ
as Fig. 13 illustrated can be obtained after c4 ; c5 ; c6 ; c7 ; c8 ; c9 are analysed. Fig. 13 illustrated
clearly a spectrum line at characteristic frequency fic½2 ð99:2 HzÞ of the roller bearing with inner-race
fault.
ARTICLE IN PRESS

D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270 269

Fig. 13. The local marginal spectrum of the signal shown in Fig. 12.

6. Conclusions

The envelope analysis method is a practical analytic approach for the fault diagnosis of roller
bearings, but the traditional way has its limitations. In order to overcome its limitations, the
envelope analysis method and EMD method are combined and applied to the fault diagnosis of
roller bearings, and better analytic results are obtained. In the proposed method, the vibration
signal of a roller bearing with faults is translated into time-scale representation by using the
wavelet bases. Hilbert transform is then used to make an envelope analysis of wavelet coefficients
of high scales that represent the high-frequency components. By applying EMD and Hilbert
transform to the envelope signal, we can obtain the local marginal spectrum from which the faults
in a roller bearing can be diagnosed and fault patterns can be identified. Practical vibration signals
measured from roller bearings with out-race faults or inner-race faults are analysed by the
proposed method. The results show that the proposed method is superior to the traditional
envelope spectrum method in extracting the fault characteristics of roller bearings.

Acknowledgements

The support from Chinese National Science Foundation Grant (No.50275050) for this research
is gratefully acknowledged.

References

[1] G.A. Radcliff, Condition monitoring of rolling element bearings using the enveloping technique. Machine
Condition Monitoring. Mechanical Engineering Publication Ltd., London, 1990, pp. 55–67.
[2] R.B. Randall, Hilbert transform techniques in machine diagnostics. IFToMM International Conference on
Rotordynamics, Tokyo, 1986.
[3] D.N. Brown, Envelope analysis detects bearing faults before major damage occurs, Pulp and Paper 63 (1989)
113–117.
[4] D. Ho, R.B. Rand, Optimization of bearing diagnostic techniques using simulated and actual bearing fault signal,
Mechanical Systems and Signal Processing 14 (5) (2000) 763–788.
[5] R.B. Randall, Development in digital analysis techniques for diagnosis of bearings and gears, Proceedings of the
Fifth International Congress on Sound and Vibration, Adelaide, Australia, 1997.
[6] P.D. McFadden, J.D. Smith, Vibration monitoring of rolling element bearings by the high frequency resonance
technique: a review, Tribology International 17 (1984) 3–10.
ARTICLE IN PRESS

270 D. Yu et al. / Mechanical Systems and Signal Processing 19 (2005) 259–270

[7] P.W. Tse, Y.H. Peng, R. Yam, Wavelet analysis and envelope detection for rolling element bearing for rolling
element bearing fault diagnosis—their affectivities and flexibilities, Journal of Vibration and Acoustic 123 (2001)
303–310.
[8] R.B. Randall, Y. Gao, Masking effects in digital envelope analysis of faulty bearing signals, Proceedings of the
Sixth International Conference on Vibrations in Rotating Machinery, ImechE, Oxford, 1996, pp. 351–359.
[9] N.E. Huang, et al., The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary
time series analysis, Proceedings of the Royal Society of London A 454 (1998) 903–995.
[10] N.E. Huang, Z. Shen, S.R. Long, A new view of nonlinear water waves: the Hilbert spectrum, Annual Review of
Fluid Mechanics 31 (1999) 417–457.

You might also like