Professional Documents
Culture Documents
Life Cycle Vibration Analysis Based On EMD of Rolling Element Bearing Under ALT by Constant Stress
Life Cycle Vibration Analysis Based On EMD of Rolling Element Bearing Under ALT by Constant Stress
Life Cycle Vibration Analysis Based On EMD of Rolling Element Bearing Under ALT by Constant Stress
Abstract—Bearing is one of critical rotation components in based on expert system of fuzzy knowledge base[9], Markov
mechanical systems. Many grievous accidents of mechanical model[10], machine learning of support vector machine[11],
systems are chiefly caused by bearing failure. So it’s significant Kalman filter[12], pattern recognition in artificial
meaningful for practical application to research failure feature intelligence[13] and so on. In addition, there are also some
detection and residual life prediction of key rotation components methods by information entropy and combined methods for
such as bearings. This paper has designed experimental scheme prediction. But no matter any methods, statistics or signal
of Accelerated Life Test of rolling element bearings and gained analysis, the key is to detect Anomaly Characteristics. Only
all vibration signals of bearings through their total life, from when anomaly feature is distinguished from normal, prediction
normal to fault. Then we extracted failure feature from those
can be underway.
signals and analyzed its evolving rule in the bearing life cycle
based on Empirical Mode Decomposition(EMD). Finally, we have Nowadays there many fault diagnosis methods for rolling
analyzed bearing failure feature in each stage through its life element bearings, the core of most of above methods is means
cycle by comparing Frequency Spectrums with Hilbert Marginal for fault diagnosis such as time series model, ferrography
Spectrums. analysis, wavelet decomposition and so on. Because of
nonlinear and nonstationary when bearings run, the
Keywords- EMD; fault diagnosis; IMF; Hilbert transform; performance of many fault diagnosis methods is not all roses.
fault prediction
In this paper, we designed experiment of Constant Stress
I. INTRODUCTION Accelerated Life Testing for rolling element bearings and
obtained vibration signals in each phases of rolling element
Bearings are key rotation components in mechanism system. bearings life cycle. Then we decomposed the vibration signals
Rolling element bearings are usually adopted when there is no by the method of Empirical Mode Decomposition(EMD)[14]
space or cost to use hydraulic System. Because of the feature which was proposed by Norden E. Huang in 1998. For the
of bearings structure, rolling element bearing is one of easier adaptability of EMD, the result of decomposition is dependent
fault parts in mechanism system, the total system performance on signal itself. It shows that EMD be suitable for the analysis
is affected by the condition of rolling element bearings. In the of nonlinear and nonstationary signals. Intrinsic Mode
past, the common method to deal with fault is to change a new Functions(IMFs) which are decomposed by EMD basically
one. Along with the increasing of system complexity more and satisfy the demand of completeness and orthogonality.
more difficulties are appearing when bearings are assembled
and disassembled. The use life and function precision of other
parts in relation to bearings have been affected by assembly II. EMD
and un-assembly of them. For the increasing of use life and the Vibration signals of bearings are typically nonlinear and
developing of reliability and availability, bearings are the nonstationary. They do not satisfy the qualification of directly
bottleneck of improving the use life of mechanism system. using Hilbert Transformation and Fourier Transformation to
Moreover, other faults in mechanism system usually are felt deal with them. EMD is a signal processing method proposed
through move parts, so we need to distinguish faults which by Norden E. Huang for decomposing signals and signals
occur in bearings or not. Hence the research of feature decomposed are called Intrinsic Mode Functions(IMFs) who
extraction and prediction of remains life of typical key move are linear and stationary and satisfy basal demands of signal
parts such as bearings have a critical practical value for safely processing methods. IMFs need to satisfy two conditions[15]:
working of the complex mechanism system.
(1). Number of extremum points and zero-crossing points
At present, there are many methods for fault prediction. should be equal or no more than one point;
Such as the trend of history data fitting curve[1], time series
model of AR & ARMA[2,3], ferrography analysis[4], grey (2). Average value of envelopes of local maximum and
model based on grey system theory[5], learning and training of minimum should be zero at any time.
neural network[6], wavelet decomposition[7,8], fuzzy model
978-1-4244-4905-7/09/$25.00©2009 IEEE
1177
On the practical application, the first condition is satisfied
more easily than the second one. Therefore, we should define
stopping criteria for the second condition.
The algorithm steps as following:
(1). Find the extremum points of the original signal x ( N )
and then connect them by spline curves to obtain envelope
curves. After this we can figure out average curves m1 of the
envelopes maximum and minimum;
(2). Take away m1 ( N ) which are values of
m1 corresponding to x ( N ) from x ( N ) , the remnant is h1 ( N ) , Figure 1. Structure Principle of ABLT-1A automatic control machine of
fatigue life and reliability enhancement experiment of rolling element
that is x( N ) − m1 ( N ) = h1 ( N ) ; bearings
r2 ( N ) = r1 ( N ) − imf 2 ( N )
#
rn ( N ) = rn −1 ( N ) − imf n ( N )
rn ( N ) is either a monotone function or satisfying the Figure 2. Experiment and data acquisition flow
stopping criteria, terminate decompostion process and mark
res ( N ) = rn ( N ) . Then we can represent x ( N ) by Statistical data from large-sized bearing factory showed
n
that failure process of rolling element bearings could be
x( N ) = ∑ imfi ( N ) + res ( N ) . divided into four stages in vibration signal field. There are
i =1 super frequency stage, natural frequency stage, fault frequency
stage and wide band random vibration stage. The flow of our
In this paper, the method of EMD is discretized and it will experiment is organically integrated with failure process of
be suit data process well in practical using. We use it to bearings and we measure vibration signals through the total life
decompose vibration signals which were obtained in our cycle of bearings. Following this we can obtain the feature of
experiment. vibration signals of bearings in each stage. Based on the
constant stress Fr , the machine stopped because the kurtosis
III. EXPERIMENTAL SCHEME was beyond the value set before on 90h. After disassembled the
Experimental object: 6307 deep groove ball bearings; fault type and position we detected was fatigue failure on the
inner raceway in Fig. 3.
Experimental stress: radial load Fr = 11.13kN ;
Experimental method: assemble four 6307 deep groove ball
bearings on the common axis of rotation and then install it into
ABLT-1A which is an automatic control machine of fatigue
life and reliability enhancement experiment of rolling element
bearings. Test bearings are assembled on the position of 14, 15
and 16 in Fig. 1. Then make sure the radial load of each
bearing correctly by hydraulic system. Install accelerometers
on appropriate positions and then start it and do our experiment
following the flow chart in Fig. 2.
1178
analysis methods satisfied by those conditions to deal with
IMFs. The physical meaning of characteristic frequencies in
d
n FFT spectrum is the value they contribute to vibration in whole
Doi
D
io
Di
Do
time and they cannot reflect the real-time change of
Dm
characteristic frequencies of signals. But the spectrum of FFT
can be regarded as a reference. On the other side each IMF
dealt with EMD satisfies the demand of Hilbert Transform and
can be translated into the Hilbert Marginal Spectrum(HMS).
There are not obvious characteristic frequencies under the
condition of non-failure in HMS even power frequency or
Figure 4. 6307 deep groove ball bearings doubling. They are obvious failure feature frequencies only
when there have faults and characteristic frequencies such as
power frequency and doubling will be also obvious because of
IV. CHARACTERISTIC FREQUENCY CALCULATION faults. Figures can show us the visualized trend of change. But
For more effectively analyzing the physical meaning of they cannot be compared quantificationally because natural
feature frequency in each stage, we are necessary for figuring frequencies and other non-failure feature frequencies change as
out several more important characteristic frequency of 6307 well as failure feature frequencies. Therefore normalization
deep groove ball bearings. In Fig. 4, the ball number z is 8. processing should be implemented for characteristic
The diameter of inner ring Di is 35mm and the diameter of frequencies and failure feature frequencies.
outer ring Do is 80mm . The width of bearing B is 21mm . The
B. Result Comparison
diameter of raceway bottom of inner ring Dio is 45.006mm and
Fig. 5 are IMFs which are the results of bearing vibration
the diameter of raceway bottom of outer ring Doi is 72.007mm . signals measured in the beginning dealt with EMD. From the
The diameter of ball d is 13.494mm . The contact angle of figure we can observe that the original signal was decomposed
6307 deep groove ball bearings α is 0o . So the average by EMD into IMFs from high frequency to low. The last one is
diameter Dm = ( Dio + Dio ) / 2 = 58.507 mm . Spindle speed n is a residual signal after EMD processing. We didn’t do it again
because there is less meaning to deal with IMFs in time domain.
3000r / min .
5
nal
0
sig
-5
0
-5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
0
-2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
1
imf3
0
-1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
1
f4
zn d 8*3000 13.494
im
0
-1
cos α ) =
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Fi = (1 + (1 + ) = 246.13Hz
0.5
imf5
0
-0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
2* 60 Dm 120 58.507
0.5
imf6
0
-0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
0.2
imf7
0
-0.2
zn d 8*3000 13.494
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
0.05
imf8
Fo = (1 − cos α ) = (1 − ) = 153.87 Hz
-0.05
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
0.05
imf9
2 * 60 Dm 120 58.507
-0.05
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
0
0.05
f1
0
1 im
-0.05
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
0.02
f1
n * Dm
0
im
Fb = (1 − ( ) 2 cos 2 α ) = (1 − (
0.05
) )
s
re
0
-0.05
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
1000
500
0
0 200 400 600 800 1000 1200 1400
500
0
0 200 400 600 800 1000 1200 1400
200
0
100
0
0 200 400 600 800 1000 1200 1400
after EMD processing satisfy the demand of orthogonality,
completeness and stationarity. Therefore we can use many Figure 6. IMFs in the beginning dealt with FFT
1179
2000 1000
signal
imf1
1000 500
0
0 200 400 600 800 1000 1200 1400 0
0 200 400 600 800 1000 1200 1400
1000 500
imf1 500
imf2
0
0 200 400 600 800 1000 1200 1400
0
1000 0 200 400 600 800 1000 1200 1400
imf2
500 1000
0
imf3
0 200 400 600 800 1000 1200 1400 500
500 0
imf3
imf4
400
imf4
200 0
0 200 400 600 800 1000 1200 1400
0 1000
0 200 400 600 800 1000 1200 1400
imf5
200 500
imf5
100
0
0 0 200 400 600 800 1000 1200 1400
0 200 400 600 800 1000 1200 1400
100
res
50
0
0 200 400 600 800 1000 1200 1400
Figure 12. ES of IMFs on 39min past 65h
imf1
1
0
0 200 400 600 800 1000 1200 1400
1000 4
signal
x 10
500 2
imf2
0 1
0 200 400 600 800 1000 1200 1400 0
400 0 200 400 600 800 1000 1200 1400
imf1
200
10000
0
imf3
0 200 400 600 800 1000 1200 1400 5000
400 0
0 200 400 600 800 1000 1200 1400
imf2
200
0 4000
imf4
0 200 400 600 800 1000 1200 1400 2000
500 0
imf3
imf5
400 500
imf4
200 0
0 200 400 600 800 1000 1200 1400
0
0 200 400 600 800 1000 1200 1400
500
imf5
0
0 200 400 600 800 1000 1200 1400 Figure 13. ES of IMFs on 27min past 89h
100
res
50
0
0 200 400 600 800 1000 1200 1400
Characteristic frequencies are everydayness and the bearing
is health in Fig. 6 and in Fig. 10. A lot of frequencies and
Figure 8. IMFs on39min past 65h dealt with FFT
harmonics appear around 1000Hz in IMF1 of Fig. 7 and Fig.
10000 11 and the value of them cannot be recognized. The bearing is
signal
5000
0
0 200 400 600 800 1000 1200 1400 under the condition of initial damage. But these cannot show
10000
that the failure feature frequency of initial damage is about
imf1
5000
0
10000
0 200 400 600 800 1000 1200 1400
1000Hz, it may be more high frequency than 1000Hz and
imf2
5000
0
4000
0 200 400 600 800 1000 1200 1400 1000Hz may be only one of its harmonics. In Fig. 8 and Fig. 12,
the value of amplitude in IMF1 is contracted and many
imf3
2000
0
0 200 400 600 800 1000 1200 1400
2000 harmonics appear around 246Hz which is the failure feature
imf4
1000
0
500
0 200 400 600 800 1000 1200 1400 frequency of inner raceway we have calculated before. In Fig.
9 and Fig. 13, characteristic frequencies suddenly are very
imf5
0
0 200 400 600 800 1000 1200 1400
1000 obvious in each IMF of Fig. 9 and Fig. 13. There are obvious
res
500
0
0 200 400 600 800 1000 1200 1400 failure feature frequencies in IMF1, IMF2 and IMF3. The
machine stopped because the value of kurtosis was beyond
Figure 9. IMFs on27min past 89h dealt with FFT stopping criteria.
The Envelopment Spectrums(ES) represent more than
Hilbert 边 际 谱 采 样 频 率 Fs=2560Hz
350
2000
200
imf1
1000
幅值
0
0 200 400 600 800 1000 1200 1400 150
2000
imf2
1000
0 100
0 200 400 600 800 1000 1200 1400
1000
imf3
500
50
0
0 200 400 600 800 1000 1200 1400
400
0
imf4
500
0
0 200 400 600 800 1000 1200 1400 Figure 14. HMS in the beginning
700
2000
imf1
1000
0
600
Failure Frequency Point
0 200 400 600 800 1000 1200 1400
2000 500
imf2
1000
0
幅值
500 300
0
0 200 400 600 800 1000 1200 1400
1000 200
imf4
500
0 100
0 200 400 600 800 1000 1200 1400
400
imf5
200
0
0 200 400 600 800 1000 1200 1400
0 频率 (Hz)
0 200 400 600 800 1000 1200 1400
Figure 11. ES of IMFs on 19min past 41h Figure 15. HMS on 19min past 41h
1180
700
Hilbert 边 际 谱 采 样 频 率 Fs=2560Hz
fault in the later period. Feature frequencies in HMS are very
600 obvious. Along with the fault being aggravated, feature
500 Failure Frequency Point frequency is realization of periodicity and the amplitude of it is
400
larger and larger. Therefore it shoots way out of strong noise
300
frequencies and the period looks like piecewise linear and
200
stationary.
100 In the later period of bearing fatigue failure, the failure
0
0 200 400 600 800 1000 1200 1400
feature frequency is more and more obvious. The viscosity of
频 率 (Hz)
lubricating oil in degrading and the amplitude value of fault is
Figure 16. HMS on 39min past 65h
growing and the temperature around the fault bearing is rising
speedly. The contact area of failure surface is spreading
14000
Hilbert 边 际 谱 采 样 频 率 Fs=2560Hz quickly. All these will lead the bearing to accelerating damage.
12000
For analyzing the vibration signal of bearing quantificationally,
10000 Failure Frequency Point
the vibration signal in the beginning and on 27min past 89h are
8000
normalized, then deal with them and the result is showed in
Table 1.
6000
4000 By observing the data in the table, we can find that values
2000 around 246Hz are very different in the beginning and on 27min
0
0 200 400 600 800 1000 1200 1400
past 89h. It shows that the bearing failure feature frequency is
频 率 (Hz)
about 246Hz. By comparing with failure feature frequencies
we have calculated before, we know the fault type is inner
Figure 17. HMS on 27min past 89h raceway fatigue failure and the result we observe after
disassembling has proved it. Furthermore, the amplitude value
Fig. 14-Fig. 17 are Hilbert Marginal Spectrums of bearing
of each later characteristic frequency is larger than the former,
vibration signals in each stage dealt with EMD in total life
this shows that amplitude values of natural frequencies and
cycle. This method highlights characteristic frequencies when
their relative frequencies are increasing with fault growing.
some faults arise. In Fig. 14, the failure feature frequency is
There are many harmonic frequencies appear in the stage of
submerged by other frequencies and there are not obvious
fault growing.
feature frequency components. In Fig. 15 and Fig. 16, the
failure feature frequency of bearing is relatively obvious and it
is in a stabilizing phase. Fig. 17 represents the HMS of bearing
TABLE I. NORMALIZAITON OF VIBRATION SIGNALS IN THE BEGINNING AND ON 27MIN PAST 89H
The Beginning 27min past 89h
Frequency Amplitude Amplitude Amplitude Amplitude
Amplitude Amplitude
/Maximum /Mean /Maximum /Mean
49.75 158.3 0.215 14.77 746.0 0.270 24.03
96.5 41.6 0.058 3.89 252.4 0.091 8.13
99.5 57.5 0.080 5.37 503.5 0.182 16.22
146.25 52.8 0.070 4.92 965.0 0.349 31.09
149.25 106.9 0.145 9.97 412.5 0.149 13.29
152.25 736.6 1.000 68.71 970.5 0.352 31.28
196 140.1 0.190 13.07 1725.2 0.625 55.60
198.75 167.0 0.227 15.60 468.6 0.170 15.10
242.75 28.0 0.038 2.61 383.5 0.139 12.35
245.75 89.8 0.122 8.38 581.0 0.211 18.72
248.5 29.4 0.040 2.75 700.0 0.253 22.60
295.5 114.0 0.155 10.63 2761.8 1.000 89.01
298.25 64.9 0.088 6.06 350.2 0.125 11.28
304.25 56.7 0.077 5.29 468.0 0.170 15.08
305 653.8 0.888 60.99 436.0 0.158 14.05
342.25 27.5 0.037 2.54 1078.5 0.391 34.75
345.25 53.1 0.072 4.95 1921.5 0.696 61.92
392 42.3 0.057 3.94 1440.0 0.521 46.40
395 38.3 0.052 3.57 280.0 0.102 9.11
441.75 27.2 0.037 2.54 203.0 0.073 6.53
Attention: frequencies we compare are less than 450Hz.
IMFs which were the results of vibration signals dealt with
VI. CONCLUSIONS EMD.
This paper has designed experimental scheme of total life By dealt with vibration signals by HHT and analyzing the
cycle of bearings under constant stress and dealt with the results, we can summarize several conclusions as following:
vibration signal we gained from experiment. Then we analyzed
1181
(1). By analyzing EMD and discretization, we know the [7] J.S. Cheng, D.J. Yu, Y. Yang and Q.W. Deng. “Application of Scale-
second condition of IMF is very harsh. So we must choose Wavelet Power Spectrum to Fault Diagnosis of Rolling Bearings,”
Chinese Journal of Vibration Engineering, 2004, No.1, pp.82-85.
proper stopping criteria in practical using. Otherwise the loop
[8] Z. Li, Z.J. He, Y.Y. Zi and Y.X. Wang. “Customized Wavelet Denoising
will be iterated many times or all along and we cannot get Using Intra- and Inter-scale Dependency for Bearing Fault Detection,”
correct IMFs or the result of decomposition is a constant Journal of Sound and Vibration, 2008, No.2008, pp.342-359.
amplitude harmonic frequency, and it loses the physical [9] Q. Wang, J. Zheng, J.Q. Qin and C.Z. Jia. “The Modeling Method Study
meaning of the original signal. of Fault Fuzzy Forecast System,” Chinese Journal of Computer
Measurement & Control, 2002, No.1, pp.23-25.
(2). There are usually several neighboring IMFs that [10] Y.C. Li, X.D. Wu, Z.Q. Li, S.Q. Liu and Y.P. Wang. “Application of
contain feature frequencies because of nonlinear and Gray Markov Prediction Model in Prediction of Progressive-cavity
nonstationary, but they must not appear in nonadjacent IMFs at Pumps Lifetime,” Chinese Journal of Natural Gas Industry, 2006, No.10,
the same time. pp.92-94.
[11] D.J. Yu, M.F. Chen, J.S. Cheng and Y. Yang. “Fault Diagnosis
(3). the vibration signal in the beginning and on 27min past Approach for Rotor Systems Based on Support Vector Machine
89h are normalized and we observed the failure feature Predictive Model,” Chinese Journal of China Mechanical Engineering,
frequency and knew that the fault of the bearing is inner 2006, No.7, pp.696-699.
raceway fatigue failure by comparing the feature frequency [12] M.Z. Chen and D.H. Zhou. “Fault Prediction Techniques for Dynamic
with those we have calculated before. Systems,” Chinese Journal of Control Theory & Applications, 2003,
No.6, pp.819-823.
[13] F.Y. Wang, D.X. Jiang, J. Han and H.J. Sun. “Application of Pattern
REFERENCES Recognition on Faults Prediction,” Chinese Journal of Vibration
[1] Y. Jiang, X. Yang and Q.M. Ruan. “Review of prediction methods of Engineering, 2000, No.5, pp.669-672.
fault rules and trend of machine,” Chinese Journal of Mechatronics, [14] N.E. Huang, Z. Shen and S.R. Long. “The Empirical Decomposition and
2001, No.3, pp.14-17. the Hilbert Spectrum for Nonlinear and Non-stationary Time Series
[2] Z.S. Chen, Y.M. Yang and Z. Hu. “Early Fault Forecasting based on Analysis,” Proceedings of the Royal Society of London, 1998, pp.903-
Almost Periodic Time-Varying Autoregressive Model,” Chinese Journal 995.
of Mechanical Engineering, 2005, No.1, pp.184-188. [15] G. Rilling, P. Flandrin and P. Goncalvès. “On Empirical Mode
[3] J. Han and R.L. Zhang, Methods of Failure Mechanism and Diagnosis in Decomposition and Its Algorithms,” Proceedings of IEEE-EURASIP
Rotation Machine. Beijing: Publishing Company of Mechanical Industry, Workshop on Nonlinear Signal and Image, Grado, Italy, 2003.
1997. [16] Z.J. Liu, S.Q. He and X.L. Li, Application Handbook of Rolling
[4] D.X. Tang. “Application of Ferrography in the Research of Vehicle Element Bearing. Beijing: Publishing Company of Mechanical Industry,
Failure Prediction,” Chinese Journal of Southwest University(Natural 2006.
Science), 2007, No.1, pp.111-115. [17] Hong-Zhong Huang, Zhi-Jie Liu, D.N.P. Murthy, “Optimal reliability,
[5] J.H. Chen and Y.Z. Guo. “Application of GM Fuzzy Optimal Method in warranty and price for new products,” IIE Transactions, 2007, Vol.39,
Fault Forecasting for a Few Sample Mechanical Systems,” Chinese No.8, pp. 819-827.
Journal of China Mechanical Engineering, 2002, No.19, pp.1658-1660. [18] Z.J. Liu, W. Chen, H.Z. Huang, B. Yang, “A diagnostics design decision
[6] Z.Y. Lu, Z.C. Zhao and W. Hao. “Multi-model Ensemble Forecast model for products under warranty,” International Journal of Production
Method Based on ANN,” Chinese Journal of Computer Applications, Economics, 2007, Vol.109, No.1-2, pp. 230-240.
2004, No.4, pp.50-51.
1182