Health Monitoring and Fault Detection Using

Wavelet Packet Technique and Multivariate Process

Control Method

Yu Wang
School of Mechanical Engineering
Xian Jiao Tong University
Xiaohang Jin, Yi Sun, Jihong Shan Xian China
College of Mechanical Engineering
Zhejiang University of Technology
Hangzhou, China Zhengguo Xu
[email protected], [email protected] Department of Control Science and Engineering
Zhejiang University
Hangzhou, China

AbstractGenerally, machines will undergo two stages: normal calculated by using wavelet packet technique to represent the
working stage and irreversible degradation stage, before they health condition of machines in this article. Multivariate
fail. The division of two these stages are helpful for remaining statistical process control (MSPC) [6-8] method is then
useful life (RUL) estimation. With the help of health monitoring employed to analyze the feature data set for detecting the initial
technique, the start point of the degradation stage can be detected point of the degradation stage. These two methods combination
successfully, which facilitates the prognostic activities. This paper make it possible to detect the fault at early stage without
reports the work on the detection of start point of the expertise for fault diagnosis, which could facilitate the
degradation stage in machines. An approach, which combines the prognostics activities.
wavelet packet technique and multivariate statistical process
control method, for fault detection was developed. Motor The rest of this article is organized as follows. In Section II,
bearings and fans data are employed to illustrate the effectiveness the proposed ensemble approach is described. The principles of
of the proposed method. Results show that it is feasible for fault wavelet packet technique and multivariate statistical process
detection. control method are also briefly introduced. In Section III,
validation results based on fan and bearing data are reported.
Keywords- Fault detection; cooling fan; health monitoring; Finally, in Section IV, some conclusions are drawn.
motor bearing; multivariate statistical process control; wavelet
packet.
II. PROPOSED ENSEMBLE APPROACH
I. INTRODUCTION Sensor-based vibration signal is a useful way to indicate the
health conditions of machines. Generally, faulty machine
Prognostics gains increasing interests in industry because of
signals are always different with normal signals. Fig. 1 presents
its effective and efficient role in monitoring the health
vibration signals for a cooling fan at its normal work stage.
conditions of machines, detecting potential failures, and
Notice that the vibration energy is dominated at low
preventing unexpected downtime and economic losses [1-2]. In
frequencies (<5000 Hz). Fig. 2 presents vibration signals for a
practice, we often observe that the life of machines undergoes
degrading bearing. We can see that the amplitude of spectra
two stages: normal working stage and irreversible degradation
increases and the energy distribution at different frequency
stage [3]. The division of these two stages will be helpful for
bands varies with the time index (Sample No.).
predicting the remaining useful life (RUL) of the machines,
since it is meaningless to estimate the RUL in the normal Since energy related features could reflect the health of
working stage. This paper reports the work on the detection of machine, vibration signal was decomposed by wavelet packet
start point of machines degradation stage. technique and energy corresponding to the approximation and
details were calculated. The constructed multivariate energy
When there is a fault in machine, fault signatures will lead
feature data were then input into MSPC method for detecting
the energy distribution to vary from the normal ones. Yan and
the start point of degradation stage. The proposed approach,
Gao presented an energy-based approach for diagnosing
which was based on wavelet packet technique and multivariate
bearing faults using wavelet technique [4]. The energy features
statistical process control for fault detection, is illustrated in
corresponding to wavelet coefficients of vibration signals were
Fig. 3 and Fig. 4. In this article, we define that there is a fault in
widely used for diagnosis and prognosis of the health condition
the machine when more than 10 continuous samples are out of
s in rotary machines [5]. Therefore, energy related features are

978-1-4799-7958-5/14/$31.00 2014 IEEE 257

control or the total percentage of samples above control limit is Energy for wavelet
over 8%. packet coefficients

Energy-based feature data samples

0.1
Amplitude (g)

Phase I
Phase II
1. Obtain a healthy machine data set
0.05 Calculate the Hotelling T2 statistic
2. Calculate the sample mean vector
and draw T2 control chart of future
and covariance matrix
health condition of machine
3. Calculate the UCL
0
Quantification
50
10000 1. Count the number of samples above control limit
30
5000 2. Calculate total percentage of samples above control limit on the chart
10
Sample no. 0
Frequency (Hz)
Figure 1. Vibration spectra of a healthy cooling fan.
More than 10 continuous
No
samples above control limit
1.5 or total percentage > 8% ?
Amplitude (g)

1 Yes
Start point of degradation stage
0.5
Figure 4. Block diagram of MSPC for fault detection.

0
A. Wavelet packet technique
2000 10000 The discrete wavelet transform (DWT) of a signal provides
1000 5000 shifted and scaled wavelet filters to examine the signal locally
Sample no. 0 0 in the time domain at different levels of scale. As a result, a tree
Frequency (Hz)
structure with approximation and detail coefficients is
Figure 2. Vibration spectra of a degrading bearing. constructed. Wavelet packet decomposition (WPD) is a
generalized DWT. In the WPD of a signal, both approximation
and detail coefficients are decomposed at each level [9]. Figs. 5
Vibration signal and 6 show an example of DWT and WPT of a signal at level 2,
respectively.
Wavelet packet Calculation energy
decomposition for vibration signal

Calculate energy for Wavelet packet

coefficients respectively
Figure 5. DWT of a signal at level 2.
Multivariate statistical process control processor

Fault detection

Figure 3. Flowchart of the proposed strategy.

Figure 6. WPT of a signal at level 2.

B. Multivariate control charts using individual observations

Control charts are tools used to determine if a process is
under statistical control [8]. Similar idea is borrowed in this
article to monitor the health condition of a machine. That is to
say, control charts are used to monitor a machine if it is under
statistical health. Generally, there are two or more variables
indicating the health condition of machine, and machine will
undergo two stages. Hence, a reasonable approach is to chart
Hotelling T2 statistic for individual observation from the

258
machine to detect the initial point of the degradation stage. A 2
10
Hotelling Ti 2 - statistic is of the form as follows [6-8].

( ) (
Ti 2 = Zi Z S1 Zi Z ) (1)
10
1

where Zi = ( Z i1 , Z i 2 , , Z ip ) , p is the number of features, Z is

'

T2
the mean vector, and S is the covariance matrix.
0
10
There are two different phases in the application of control
chart: Phase I and Phase II. In Phase I, it aims to find the
control limits which based on controlled observations.
-1
10
(m 1)2 0 20 40 60
UCL phase I = , p / 2,( m p 1) / 2
m (2) Sample no.
LCL phase I = 0
Figure 7. Hotelling T2 control chart for Fan #1.

Phase II aims to monitor future or unknown processes to 2

estimate whether they are in control or not. The control limits 10
for this statistic are
p (m + 1)(m 1)
UCL phase II = F , p , m p 10
1

m 2 mp (3)
LCL phase II = 0

T2
When the sample size is big ( m > 100 ), and the covariance 0
matrix is assumed to be known, the chi-square limit then could 10
be used as an approximate control limit [8]. In this article,
upper percentage point of the distributions is set as = 0.01 .
-1
UCL = 2 , p (4) 10
0 20 40 60
Sample no.

Figure 8. Hotelling T2 control chart for Fan #2.

III. RESULTS AND DISCUSSION
The proposed methodology is validated on two B. Fault detection of bearings
experimental data. Specifically, healthy cooling fan data are
used to test whether the method will generate false alarms, and The run-to-failure bearing data are acquired on
the run-to-failure bearing data are used to test whether the PRONOSTIA platform [11] under below operation conditions:
degradation stage could be detected effectively. 4000 N (bearings maximum dynamic load) and 1800 rpm.
Seven same types of bearings are tested to failure totally.
Vibration signals are acquired every 10 seconds. Signal
A. Health monitoring of cooling fans
acquisition is stopped when the amplitude of the vibration
Two healthy cooling fans, named Fan #1 and Fan #2, are signal is bigger than 20 g. The data-sampling rate is 25.6 kHz.
powered by 12 V DC power supply to run under free airflow In this article, results are got just based on the horizontal
conditions continuously. Their vibration signals are acquired vibration signals. Since these seven bearings show similar
every 15 minutes. Seventy samples were recorded for each fan. results, only bearing #1s Hotelling T2 control chart is shown in
The data-sampling rate is 25.6 kHz [10]. Normally, cooling fan Fig. 9 and Fig. 10. The upper control limit is 13.3. As we can
L10 life is 30,000 hours, so about one-days continuous running see that the T2 statistics of the beginning samples of bearing
of fan will not degrade the fans. This information is also overpass the UCL. This scenario happens due to the installation
demonstrated by the Hotelling T2 control chart in Fig. 7 and uncertainty and bearings are still in break-in period. The
Fig. 8. The UCL is 15.5. Both of these two fans are under sample no. is 1129 when more than 10 continuous samples
health condition, although two samples in Fan #2 overpass the overpass the control limit, while sample no. is 1147 when the
UCL. These two anomaly conditions may due to the total percentage of samples above control limit is over 8%, as
measurement uncertainty during the data acquisition. shown in Fig. 11. Hence, fault in bearing #1 can be detected at
sample no. 1129 based on our proposed method. Other bearings
detection time is reported in Table 1. Bearing faults can be
detected prior to the failures happen, and the duration of the
irreversible degradation stage of bearings varied greatly. As we
can see, there are 1334 time index left in bearing #1 while there
are only 109 time index left in bearing #2.

259
 10
2 IV. CONCLUSION
In this article, an ensemble approach is proposed to detect
the early stage of machine faults has been presented. By
1
decomposing the vibration signals using wavelet packet
10 technique, energy of wavelet packet coefficients are extracted
to represent the health condition of machines. Such health-
related feature data are further processed by MSPC method to
T2

0 monitor the machine is under statistical health or not. Healthy

10 cooling fan data and bearing data are used to illustrate the
effectiveness of our proposed approach. The results show that
cooling fans are statistical health and early stage of bearing
-1 faults can be detected prior to the failure happen. With this
10 method, the health condition of machine can be monitored and
0 500 1000 1500
Sample no.
early stage of machine fault can be detected. Hence, our work
could help to prevent the unexpected downtime and economic
Figure 9. Hotelling T2 control chart for Bearing #1. loss due the machine failure.

ACKNOWLEDGMENT
20 This work was supported in part by National Natural
Science Foundation of China under Grant No. 51275474 and
T2

Zhejiang Provincial Natural Science Foundation of China

10 under Grant No. LZ12E05002.
8
1000 1050 1100 1150 1200 REFERENCES
Sample no. [1] X. Jin, E. W. M. Ma, L. L. Cheng, and M. Pecht, Health monitoring of
cooling fans based on Mahalanobis distance with mRMR feature
Figure 10. Close up of Fig. 9. selection, IEEE Transactions on Instrumentation and Measurement, vol.
61, no. 8, pp. 2222-2229, Aug. 2012.
[2] N. Gebraeel, M. Lawley, R. Liu, and V. Parmeshwaran, Residual life
12 prediction from vibration-based degradation signals: a neural network
approach, IEEE Transactions on Industrial Electronics, vol. 51, no. 3,
10 pp. 694-700, June 2004.
[3] W. Wang, A two-stage prognosis model in condition based
8 maintenance, European Journal of Operational Research, vol. 182, no. 3
Percentage (%)

pp. 1177-1187, 2007.

6 [4] R. Yan and R. X. Gao, Energy-based feature extraction for defect
diagnosis in rotary machines, IEEE Transactions on Instrumentation
4 and Measurement, vol. 58, no. 9, pp. 31303139, Sep. 2009.
[5] X. Jin, M. Zhao, T. W. S. Chow, and M. Pecht, Motor bearing fault
2 diagnosis using trace ratio linear discriminant analysis, IEEE
Transactions on Industrial Electronics, vol. 61, no. 5, pp. 2441-2451,
2014.
0 [6] N. D. Tracy, J. C. Young, and R. L. Mason, Multivariate control charts
500 600 700 800 900 1000 1100 1200
for individual observations, Journal of Quality Technology, vol. 24, no.
Sample no. 2, pp. 88-95, 1992.
Figure 11. Total percentage of samples above UCL for Bearing #1. [7] J. H. Sullivan and W. H. Woodall, A comparison of multivariate quality
control charts for individual observations, Journal of Quality
Technology, vol. 28, no. 4, pp. 398-408, 1996.
TABLE I Sample no. corresponding to different conditions of bearing. [8] D. C. Montgomery, Introduction to Statistical Quality Control. 6th ed.,
Wiley, 2009
10 continuous [9] J. Daubechies, Orthonormal bases of compact supported wavelets,
Bearing Total Fault
sample above Failure Communications on Pure and Applied Mathematics, vol. 41, no. 7, pp.
no. percentage >8% detection
the UCL 909-996, Oct. 1988.
#1 1147 1129 1129 2463 [10] X. Jin, T. W. S. Chow, and K-L Tsui, Online anomaly detection of
#2 762 779 762 871 brushless dc motor using current monitoring technique, International
#3 891 892 891 2375 Journal of Performability Engineering, vol. 10, no. 3, pp. 263-271, May
#4 1083 1088 1083 1428 2014.
#5 1141 1148 1141 2803 [11] P. Nectoux, R. Gouriveau, K. Medjaher, E. Ramasso, B. Morello, N.
#6 1941 1641 1641 2448 Zerhouni, and C. Varnier, PRONOSTIA: an experimental platform for
#7 900 885 885 2259 bearings accelerated life tests, IEEE International Conference on
Prognostics and Health Management, Denver, CO, USA, 2012.

260