Identification of Significant Intrinsic Mode Functions For The Diagnosis of Induction Motor Fault

Cho et al.
: JASA Express Letters
[http://dx.doi.org/10.1121/1.4885541]
Published Online 11 July 2014
Identification of significant intrinsic mode

functions for the diagnosis of induction
motor fault
Sangjin Cho, Md. Rifat Shahriar, and Uipil Chonga)
School of Electrical Engineering, University of Ulsan, 93 Daehak-ro, Nam-gu, Ulsan,
680-749, South Korea
sjcho75@ulsan.ac.kr, m.rifat@ieee.com, upchong@ulsan.ac.kr
Abstract: For the analysis of non-stationary signals generated by a

non-linear process like fault of an induction motor, empirical mode
decomposition (EMD) is the best choice as it decomposes the signal
into its natural oscillatory modes known as intrinsic mode functions
(IMFs). However, some of these oscillatory modes obtained from a
fault signal are not significant as they do not bear any fault signature
and can cause misclassification of the fault instance. To solve this issue,
a novel IMF selection algorithm is proposed in this work.
C 2014 Acoustical Society of America
V
PACS numbers: 43.60.Bf, 43.60.Np, 43.60.Wy, 43.60.Hj [CG]

Date Received: March 31, 2014
Date Accepted: June 12, 2014
1. Introduction
Induction motors play an important role in plant productivity and efficiency. As it is
directly related to business profit, efficient maintenance of these motors must be
ensured. Existing fault diagnosis techniques can be classified into three main types
depending on their diagnosis procedure (Ref. 1): Model based, signal based, and data
based. Signal processing is a crucial component for these three types but with different
impacts and roles.2 The main purpose of the signal processing step in a fault diagnosis
system is to reveal fault signatures from the measured quantities obtained from a
motor in operation. For this purpose, timefrequency analysis tools are popular as
they can provide both time and frequency resolution simultaneously. Most of the existing time-frequency analysis methods decompose the signal based on a priori bases with
stationary assumption of the signal. But these techniques are not suitable enough for
the analysis of fault signals as they can be non-linear and non-stationary at the same
time. In contrast, empirical mode decomposition (EMD), proposed by Huang et al.,3
is a signal decomposition method that decomposes the signal into intrinsic mode
functions (IMFs) based on the local characteristic time scale of the data. These IMFs
represent natural oscillatory modes embedded in the signal and work as the basis functions, which are derived from the signal itself, rather than any pre-determined kernel.
Therefore EMD is a data adaptive decomposition technique that overcomes limitations
of other similar tool such as STFT or wavelet.
Up to now, a number of fault diagnosis approaches based on EMD appeared
in different literature. Fan et al.4 reported that the change in velocity of the amplitude
of IMFs in a particular unit of time is increased when vibrations are caused by faulty
components. Li et al.5 exploited EMD to decompose vibration signal obtained from
roller bearings into IMFs and determined the marginal spectrum of each IMF. Yu
et al.6 proposed the concept of EMD energy entropy and showed that its value for
vibration signals differs in case for different bearing fault types. Junsheng et al.7
proposed a fault diagnosis method for gear and bearing signals that utilized singular
a)
Author to whom correspondence should be addressed.
EL72 J. Acoust. Soc. Am. 136 (2), August 2014
C 2014 Acoustical Society of America

V
Redistribution subject to ASA license or copyright; see http://acousticalsociety.org/content/terms. Download to IP: 103.246.106.10 On: Thu, 26 Nov 2015 09:54:48
Cho et al.: JASA Express Letters
[http://dx.doi.org/10.1121/1.4885541]
values of the matrices constituted by the IMFs as feature vectors for support vector
machine (SVM) classifiers. Yan et al.8 proposed two criteria, energy measure and correlation measure, for identifying the most representative IMF after the EMD process.
Envelop spectrum of the selected IMF is subsequently analyzed for detecting existence
and location of structural defect within a bearing.
Being motivated by the success of EMD in fault diagnosis of individual rotating
parts, we present a fault diagnosis system that is capable of identifying a number of mechanical and electrical faults of an induction motor such as faulty bearing, rotor unbalance,
broken rotor bar, bowed shaft, misalignment, and phase unbalance. The main objective of
this research is to identify an efficient signal processing technique that helps to efficiently
extract fault signature. To achieve this objective, an efficient decomposition algorithm,
EMD is utilized. It decomposes the signal adaptively without a priori bases. Therefore natural oscillatory modes of a signal are obtained. To identify the oscillatory modes that represent fault signature, a novel IMF selection algorithm is proposed. This IMF selection
technique exploited characteristics of fault signals to identify significant IMFs.
2. Background: EMD process
The EMD algorithm employed here incorporates the modifications proposed by Wang
et al.,9 to overcome limitations of the original EMD.3 A summary of the implementation process of this improved EMD process is given in the following text.
(1) For any given data x(t), data validity is checked. Here, validity refers to the condition that, data x(t) has more than one extremum.
(2) To hold the adaptive dyadic filter bank characteristics of EMD, number of iteration
for each sifting process is set at 10. The expected number of IMF components are
determined as log2N, where N total number of data points.10,11
(3) All the local extrema for data x(t) are identified.
(4) All the maxima and minima are separately connected with natural cubic spline lines
to form the upper, u(t), and lower, l(t), envelopes.
(5) The mean of the envelopes is determined as m(t) [u(t) l(t)]/2.
(6) The difference between the data and the mean is taken as the proto-IMF, h(t) x(t)
m(t).
(7) Check the number of iterations in the sifting process and repeat steps 3 to 6 until its
value reaches 10.
(8) When the iteration number reaches 10, assign the proto-IMF as an IMF component,
c(t).
(9) Repeat steps 1 to 8 on the residue, r(t) x(t)c(t).
(10) The operation ends when (log2N)1 number of IMFs are obtained and the latest
r(t) is taken as the final residue.
3. Proposed method
The proposed system utilizes vibration signal obtained from an induction motor through
accelerometers. Natural oscillatory modes hidden in the signal are then obtained through
the EMD process. To eliminate spurious IMFs insignificant for our diagnosis purpose,
an IMF selection process is performed. A novel IMF selection algorithm is incorporated
here for this purpose. Features extracted from these significant IMFs are then fed to the
classifier for fault diagnosis purpose. In this work, SVM is used as a classifier.
3.1 IMF selection algorithm
The IMFs, which are considered significant due to their contribution in the exposure
of fault signature, usually hold some unique characteristics. These characteristics are
the basis of our proposed IMF selection process. The first characteristic is that they
are usually of higher power, which is also supported by the objective of the EMD process. The EMD efficiently extracts the natural oscillatory modes from a given signal.
Now the fact is that oscillations that occur due to the fault should have higher power
J. Acoust. Soc. Am. 136 (2), August 2014
Cho et al.: Diagnosis of induction motor fault EL73
[http://dx.doi.org/10.1121/1.4885541]
in comparison to other oscillations that do not represent a fault situation. As a consequence, these oscillatory modes are characterized by a higher power. The second characteristic pertains to harmonic contents of the IMFs. Usually, most of the fault signatures appear as peak amplitude at several harmonics of some fundamental frequency.
Due to the dyadic filter bank nature of EMD process, few of these harmonic peaks
will be observed in the Fourier spectrum of lower index IMFs, whereas fundamental
fault frequency peak can be found in the higher index IMFs. Considering the preceding two juxtaposed facts, an index labeled power-harmonic ratio (PHR) is calculated
for each IMF. The higher value of PHR helps us to identify the IMFs with higher average power containing fault related frequency peaks. A low value of PHR indicates
that the IMF may be of low power or contain many low amplitude harmonics. A summary of this IMF selection algorithm is given here.
P
2
(1) For each of the IMFs cj, where IMF index j 1, 2,,n, energy Ej N
i1 jcj i j is
calculated; here N number of samples
each
Pnof
IMF.
l
(2) Calculate P
threshold energy Eth
El =nl , l 1, 2,,nl and nl < n, where
l1
El 1=n nj1 Ej .
(3) Identify only those IMFs cj, for which Ej > Eth , i.e., IMF energy based thresholding
is performed which eliminates the physically insignificant IMFs.
(4) Evaluate Fourier spectrum of IMFs Cj FT{cj} and fault signal X FT{x(t)}.
(5) Exclude any cj containing only dc value in spectrum Cj.
(6) Calculate frequency peaks fp, p 1, 2,,r, which occur above the average amplitude
of the spectrum X. Then, determine frequency peaks fk, k 1, 2,,s (s < r), in the
spectrum X which have amplitude higher than the average amplitude of fp.
(7) Calculate energy contained in the determined peak frequencies and their harmonics
in case of both IMFs (Cj) and fault signal (X), which are represented by Hj and HX,
respectively.
(8) Again, for each cj selected previously, calculate PHR, PHRj Hj =HX .
(9) Now, IMFs are rearranged according to descending values of PHRj.
The essence of the proposed IMF selection technique can be utilized in two
ways. A first way is to select a fixed and minimum m number of significant IMF(s)
based on the PHR values in case of every signal record. The value of m has to be set
empirically as it is application specific. Another way is to select all the IMFs obtained
after the energy thresholding step and consider their significance in the classification
process according to the PHR values. In this paper, we considered the first approach
for illustration.
3.2 Fault classification
IMFs selected through the above-described process contain fault signatures, therefore,
they can be utilized for the detection and diagnosis of induction motor faults. After
obtaining the desired m number of IMFs, a simple statistical parameter variance is calculated for each of the IMFs that are later utilized for classification of faults. Due to
higher correlation of the selected IMFs with fault signatures, these small number of
feature parameters are sufficient to classify different fault instances efficiently.
For the classification of motor faults, SVM is employed as a classifier. SVM
has the capability of solving the learning problem with a smaller number of samples.12
Moreover, SVM supports multi-class classification by adopting the oneagainstall
(OAA) or several two-class problems. In this work, the OAA technique is exploited,
which trains a binary classifier for each fault instance to discriminate one fault case
from all others and outputs the class with largest outputs.
4. Experiments and results
In the experiment, six 0.5 kW, 60 Hz, two-pole induction motors are used to produce
the fault data under full load conditions.2 One of the motors operated under normal
condition as a benchmark for comparison with other faulty motors. The others faulty
Cho et al.: Diagnosis of induction motor fault
[http://dx.doi.org/10.1121/1.4885541]
motors had faults such as bowed rotor shaft, broken rotor bar, bearing outer race
fault, rotor unbalance, adjustable eccentricity motor (misalignment), and phase unbalance. Therefore there are eight kinds of vibration signal categories called faulty categories: angular misalignment (AMIS), bowed rotor shaft (BRS), broken rotor bar
(BRB), faulty bearing (outer raceway) (FBO), rotor unbalance (RUN), normal motor
(NOM), parallel misalignment (PMIS), and phase unbalance (PUN). Three accelerometers are used to measure vibrations in horizontal, vertical, and axial directions. These
vibration signals are used to evaluate diagnosis capability of the proposed fault diagnosis systems. Sampling frequency of the data acquisition unit was 7.68 kHz, and this
experiment provided us with 12 signal records for each fault category. Each signal was
of 1 s duration and contained 7680 samples. Therefore a total of 96 signal records,
each of 1 s length, are obtained and the number of IMF components obtained after the
EMD process is fixed, as log2(7680) 12. Among these 12 components, the first 11
are IMFs and the last one is residue. For illustration, the result of the proposed IMF
selection algorithm and associated parameters for AMIS fault signal are shown in
Table 1.
After obtaining the significant IMFs, variances of the IMFs are calculated as
feature parameters. Because three most significant IMFs for axial and horizontal sensor and two most significant IMFs for vertical sensor are considered, a total of eight
feature parameters are calculated from the three sensors data, and a feature vector is
constructed for each fault signal instance. As the feature vector size is small, no feature
selection step is required in the proposed system, which is a major benefit of proposed
fault diagnosis method.
Multi-class SVM is applied to perform the classification process using OAA
methods. The process of this method was clearly explained by Hsu et al.13 The crossvalidation procedure can prevent the over-fitting problem. Therefore in this work, fourfold cross-validation was performed to get the exact classification. The classification
process is briefly described here. Using fourfold cross-validation, the input feature
vectors are randomly partitioned into four subsets for training data and testing data.
For any iteration, 72 training data and 24 testing data are given. Sequentially one of
the subsets is tested using the SVM classifier, trained on the remaining three subsets.
Thus each instance of the whole training set is predicted; therefore the cross-validation
accuracy is the percentage of data that are correctly classified. The classification result
obtained for the proposed method is given in Table 2. To justify effectiveness of the
proposed IMF selection method, the classification without the IMF selection process is
performed, i.e., using all the IMFs and residue for three different sensor signals and
taking variance as the feature parameters, i.e., a total of 72 features. The classification
performance is also shown in Table 2 in parentheses. From the results, it is evident
Table 1. PHR value calculation for different IMFs in the case of AMIS fault.
IMF index
1
2
3
4
5
6
7
8
9
10
11
Ej (106)
Eth (106)
Hj (106)
PHRj
3826
34 482
44 753
39 803
36 906
9623
1580
1080
243
52.8
9.36
2300
4.44
218
195
167
269
8.74
0.0035
0.1696
0.1521
0.1301
0.2092
0.0068
[http://dx.doi.org/10.1121/1.4885541]
Table 2. Classification results for the proposed fault diagnosis system; numbers in the parentheses refer to
results excluding the proposed IMF selection technique.
Category
Signal class
AMIS
BRB
NOM
FBO
RUN
PMIS
PUN
BRS
Classification
accuracy (%)
AMIS
BRB
NOM
FBO
RUN
PMIS
PUN
BRS
12 (11)
0 (0)
0 (0)
0 (0)
0 (1)
0 (1)
0 (0)
0 (0)
0 (0)
12 (12)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
12 (11)
0 (0)
0 (1)
0 (0)
2 (3)
0 (0)
0 (0)
0 (0)
0 (0)
12 (12)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
12 (8)
0 (1)
0 (0)
0 (0)
0 (1)
0 (0)
0 (0)
0 (0)
0 (1)
12 (10)
0 (0)
0 (0)
0 (0)
0 (0)
0 (1)
0 (0)
0 (1)
0 (0)
10 (9)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
0 (0)
12 (12)
100 (91.67)
100 (100)
100 (91.67)
100 (100)
100 (66.67)
100 (83.33)
83.33 (75)
100 (100)
Cross-validation accuracy (%):
97.92 (88.54)
that only 2 of the 96 signals are misclassified, and a cross-validation accuracy of

97.92% is obtained in case of the proposed technique. On the other hand, if the proposed IMF selection technique is not incorporated in the fault diagnosis system, then
the accuracy drops down to 88.54%. In this case, lower classification accuracy occurs
in all the fault cases except BRB, FBO, and BRS.
5. Conclusion
A fault diagnosis system for induction motors utilizing EMD is proposed in this paper.
The IMFs obtained after EMD are not all significant; therefore a technique for identifying significant IMFs is proposed. This technique is based on finding higher energy
IMFs, which contain fault related frequency peaks, thus ensuring selection of IMFs
with fault signatures. Laboratory experiments, as well as result analysis, have also
been performed to evaluate effectiveness of the proposed technique. Finally, fault classification has also been performed utilizing the selected IMFs by SVM technique,
which produced excellent classification accuracy.
Acknowledgments
This work was supported by the 2013 Research Fund of University of Ulsan.
References and links
1
A. Bellini, F. Filippetti, C. Tassoni, and G.-A. Capolino, Advances in diagnostic techniques for
induction machines, IEEE Trans. Ind. Electron. 55, 41094126 (2008).
2
M. R. Shahriar, T. Ahsan, and U. P. Chong, Fault diagnosis of induction motors utilizing local binary
pattern-based texture analysis, EURASIP J. Image Video Process. 2013, 29 (2013).
3
N. E. Huang, Z. Shen, S. R. Long, M. Wu, H. Shih, N. Zheng, C. Yen, C. C. Tung, and H. H. Liu, The
empirical mode decomposition and the Hilbert spectrum for non-linear and non-stationary time series
analysis, Proc. R. Soc. Lond. A Math. Phys. Sci. 454, 903995 (1998).
4
X. Fan and M. J. Zuo, Machine fault feature extraction based on intrinsic mode functions, Meas. Sci.
Technol. 19, 045105 (2008).
5
H. Li, Y. Zhang, and H. Zheng, HilbertHuang transform and marginal spectrum for detection and
diagnosis of localized defects in roller bearings, J. Mech. Sci. Technol. 23, 291301 (2009).
6
Y. Yu, Y. Dejie, and C. Junsheng, A roller bearing fault diagnosis method based on EMD energy
entropy and ANN, J. Sound Vib. 294, 269277 (2006).
7
C. Junsheng, Y. Dejie, T. Jiashi, and Y. Yu, Application of SVM and SVD technique based on EMD to
the fault diagnosis of the rotating machinery, Shock Vib. 16, 8998 (2009).
8
R. Yan and R. X. Gao, Rotary machine health diagnosis based on empirical mode decomposition,
J. Vib. Acoust. 130, 021007 (2008).
Cho et al.: Diagnosis of induction motor fault
[http://dx.doi.org/10.1121/1.4885541]
G. Wang, X. Chen, F. Qiao, Z. Wu, and N. E. Huang, On intrinsic mode function, Adv. Adapt. Data
Anal. 2, 277293 (2010).
10
P. Flandrin, G. Rilling, and P. Goncalves, Empirical mode decomposition as a filter bank, IEEE
Signal Process. Lett. 11, 112114 (2004).
11
Z. Wu and N. E. Huang, On the filtering properties of the empirical mode decomposition, Adv.
Adapt. Data Anal. 2, 397414 (2010).
12
A. Widodo and B. S. Yang, Support vector machine in machine condition monitoring and fault
diagnosis, Mech. Syst. Signal Process. 21, 25602574 (2007).
13
C. W. Hsu and C. J. Lin, A comparison of methods for multi-class support vector machines, IEEE
Trans. Neural Netw. 13, 415425 (2002).

Identification of Significant Intrinsic Mode Functions For The Diagnosis of Induction Motor Fault

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Identification of Significant Intrinsic Mode Functions For The Diagnosis of Induction Motor Fault

Transféré par

Droits d'auteur :

Formats disponibles

Cho et al.

: JASA Express Letters

Published Online 11 July 2014