Mechanical Systems and Signal Processing 94 (2017) 464–481

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Comparative investigation of vibration and current monitoring

for prediction of mechanical and electrical faults in induction
motor based on multiclass-support vector machine algorithms
Purushottam Gangsar, Rajiv Tiwari ⇑
Department of Mechanical Engineering, Indian Institute of Technology Guwahati, Guwahati, Assam 781 039, India

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents an investigation of vibration and current monitoring for effective fault
Received 7 April 2016 prediction in induction motor (IM) by using multiclass support vector machine (MSVM)
Received in revised form 27 January 2017 algorithms. Failures of IM may occur due to propagation of a mechanical or electrical fault.
Accepted 12 March 2017
Hence, for timely detection of these faults, the vibration as well as current signals was
Available online 21 March 2017
acquired after multiple experiments of varying speeds and external torques from an exper-
imental test rig. Here, total ten different fault conditions that frequently encountered in IM
Keywords:
(four mechanical fault, five electrical fault conditions and one no defect condition) have
Induction motor (IM)
Mechanical and electrical faults
been considered. In the case of stator winding fault, and phase unbalance and single phas-
Multi-fault classification ing fault, different level of severity were also considered for the prediction. In this study,
Support vector machine (SVM) the identification has been performed of the mechanical and electrical faults, individually
Radial basis function kernel (RBF) and collectively. Fault predictions have been performed using vibration signal alone, cur-
One-versus-one (OVO) multiclass method rent signal alone and vibration-current signal concurrently. The one-versus-one MSVM
has been trained at various operating conditions of IM using the radial basis function
(RBF) kernel and tested for same conditions, which gives the result in the form of percent-
age fault prediction. The prediction performance is investigated for the wide range of RBF
kernel parameter, i.e. gamma, and selected the best result for one optimal value of gamma
for each case. Fault predictions has been performed and investigated for the wide range of
operational speeds of the IM as well as external torques on the IM.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction and literature review

Induction motors (IM) are most widely used industrial electro-mechanical drives in modern industries owing to their
high power capacity, robust performance, low manufacturing cost, and adaptability. Due to various stresses, such as the ther-
mal, mechanical, electrical and environmental acting upon the stator, rotor, bearings and shaft, the performance of IM can
drastically diminishes, sometime rapid wear and finally catastrophic failures may occur, if faults are not timely detected. The
catastrophic failure leads to disruptive production losses, monetary losses and serious human injuries. Therefore, for timely
fault detection to avoid the chances of catastrophic failure, numerous condition monitoring techniques have been developed
in last four to five decades based on the vibration, current, voltage, temperature, main air-gap flux and acoustic noise [1].

⇑ Corresponding author.
E-mail address: rtiwari@iitg.ernet.in (R. Tiwari).

http://dx.doi.org/10.1016/j.ymssp.2017.03.016
0888-3270/Ó 2017 Elsevier Ltd. All rights reserved.
 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481 465

Among several condition monitoring methods, the current and vibration based methods are the most widely preferred
owing to their reliability, non-intrusiveness and easy measurability. The effectiveness of these methods depends on the load-
ing of machines and signal-to-noise ratio of instruments [2]. The current monitoring can provide a unique fault patterns for
the effective prediction of mainly electrical faults, i.e. the stator winding fault, broken rotor bar fault, and phase unbalance
and single phasing. The vibration monitoring can be useful for the effective prediction of mechanical faults, i.e. bearing faults,
unbalance rotor, bowed and misaligned rotor [3]. But the entire IM can be failed due to the propagation of electrical or
mechanical faults. Hence, it would be beneficial to perform the condition monitoring of IM based on the vibration as well
as current signals.
In recent years, many artificial intelligent methods such as neural networks, fuzzy logic, neural-fuzzy, immune genetic
system and support vector machine algorithms have been employed to improve the effectiveness of prediction and diagnosis
of IM faults particularly during the maintenance judgment process [4]. Among several artificial intelligent methods for faults
prediction of IMs, nowadays the SVM is extensively gaining popularity owing to their best prediction performance, and less
training and testing time [5]. The best property of SVM algorithms is that it can perform well even with the small number of
training and testing data hence reduce the computational load [6].
The statistics confirms that half. i.e. 50% of the IM failures occur due to bearing failures, about 40% due to winding failure
of stator, and less than 10% due to rotor and shafts failures [7]. Nguyen and Lee [8] and Nguyen et al. [9] investigated
mechanical faults diagnosis of IMs based on vibration using SVM, decision tree and GA. Baccarini et al. [10] presented prac-
tical industrial application of the SVM for mechanical fault diagnostics of IM based on frequency-domain signals. Chattopad-
hyay and Konar [11] presented feature extraction using wavelet transform (CWT and DWT) and feature selection using
greedy-search technique for the multi-class fault detection of IM using the RBF neural network, multilayer perception
(MLP) neural network and SVM classifiers. Widodo and Yang [12] utilized transient current CBM and wavelet SVM for the
fault diagnosis of IM. The building of Wavelet kernel function using Haar, Daubechies and Symlet were successfully per-
formed. Bacha et al. [6] and Salem et al. [13] proposed fault condition monitoring of IM based on Hilbert-Park transform
using the SVM. Zhou et al. [14] investigated fault diagnosis of IM based on invariant character vectors using a single
class-SVM.
Widodo et al. [15] presented the fault diagnosis of IM using combination of independent component analysis (ICA) and
SVM based on the vibration and current signatures. The combination of ICA and SVM can serve as an encouraging alternative.
Morales et al. [16] introduced the data fusion by using multi-class SVM to detect mechanical faults in IM using the vibration
and line current signatures. Tran et al. [17] presented fault detection of IM based on classification and regression tree (CART)
decision trees method and adaptive neuro-fuzzy inference based on vibration signatures as well as the MCSA. Garcia et al.
[18] presented the high-resolution spectral analysis for diagnosis of multiple faults in IMs using the multiple signal classi-
fication (MUSIC) algorithm based on vibration signatures and the MCSA. Tran et al. [19] investigated the transient current
CBM based IM fault diagnosis by using Fourier–Bessel expansion and simplified fuzzy ARTMAP, a combination of fuzzy logic
and neural network architecture based on Adaptive Resonance Theory (ART). Generalized discrimination analysis (GDA) was
used to solve the high dimensionality of feature sets. Ergin et al. [20] proposed fault diagnosis in IM, i.e. the stator, bearing
and rotor using common vector approach (CVA). Nakamura et al. [21] presented diagnosis of the electrical and mechanical
faults of IM based on the Hidden Markov Model, which is widely used in the field of speech recognition. Park et al. [22]
attempted fault diagnosis of IM based on a mixed algorithm of PCA and LDA (linear discriminant analysis). From literature
survey, it is evident that faults prediction of IMs using multi-class SVM algorithms is still uncommon and has lot of potential,
especially of the mechanical and electrical faults prediction together. Hence, it can be explored further for the perfect multi-
class fault prediction in the IM.
In this paper, a comprehensive study on the prediction of faults (mechanical and electrical) in IMs has been attempted
using the artificial intelligent method called the MSVM. In Section 2, a brief introduction of the binary and multiclass
SVM algorithms is described. In Section 3, the experimental set-up and various faults in the IM are introduced. Procedure
of data acquisition and feature extractions are described. In Section 4, the IM fault predictions have been performed for var-
ious cases and Section 5 concludes the paper.

2. Brief introduction of SVM algorithms

The SVM, a statistical learning based machine learning/pattern recognition algorithm was first introduced by Vapnik in
1995. The basic SVMs are originally developed to handle the binary classification problems. For illustration, a binary classi-
fication problem is considered here as shown in Fig. 1. The circle and square represent two classes. A number of linear clas-
sifier (hyperplane) are possible to separate the two classes, however there is one that achieves maximum separation known
as optimal separating hyperplane. The goal of SVMs is to construct an optimal hyperplane in the space by using available
training samples. The hyperplane can separate the set of vectors without any errors, if the margin is maximal. Overall
aim of SVMs is to recognize test samples very well. For a binary classification problem, the process of SVM is described
as follows [23].
Now consider a binary classification problem with data points xi ; i ¼ 1; 2; . . . ; n:; belong to class I or class II, and their
associated labels yi ¼ 1, with the following decision function
f ðxÞ ¼ signðw  x þ bÞ ð1Þ
466 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481

y The optimal class I

separating
hyperplane class II
(W)
SV

SV A separating
SV
hyperplane

SV
w, x b 1

b Maximum w, x b 0
margin w, x b 1
x
Fig. 1. Illustration of basic SVM.

For the linearly separable case, the following condition should be satisfied:

yi ðw  x þ bÞ > 0 8i ð2Þ
where the weighting vector w defines a normal vector to the hyperplane that defines the boundary, and b (bias) defines the
hyperplane’s distance from the origin. The optimal hyperplane is one that has maximum margin, i.e. the distance between it
and the nearest data point of each class. The optimal hyperplane and margin are presented in Fig. 1. Points that lie on the
margin or within the margin are called support vectors. The optimal hyperplane maximizes the margin. Thus the problem
of constructing optimal hypeplane is transformed into the following quadratic optimization problem
1
min; UðwÞ ¼ kwk2 ð3Þ
2

Subject to; yi ðw  xi þ bÞ P 1; i ¼ 1; 2; . . . ; n ð4Þ

However, in general case, the data available could be non-linearly separable, therefore some training data points do not
satisfy the constraint conditions. Thus to find the optimal hyperplane, one can solve the following constrained optimization
problem:

1 Xn
min; UðwÞ ¼ kwk2 þ C ni ð5Þ
2 i¼1

subject to; yi ðw:xi þ bÞ P 1  ni ni P 0; i ¼ 1; 2; . . . ; n ð6Þ

where the variables ni defines the positive slack variables, which is necessary to allow misclassification for some data points
in order to reduce the calculation complexity. The variable C defines the generalization parameter or soft margin classifier,
which is a trade-off between the misclassification and boundary complexity. This formulation is called as the soft-margin
SVM. To handle the high number of attributes of data examples, the above problem with complex constraint can be easily
converted into the equivalent Langrange dual problem. Thus the dual problem becomes:
X
n
1X n
MaxfDðaÞg ¼ ak  fy ai y aj ðxi ; xj Þg ð7Þ
k¼1
2 i;j¼1 i j

X
n
constraint; yj aj ¼ 0 and ai P 0; i ¼ 1; 2; . . . ; n ð8Þ
j¼1

The SVM is extended to non-linear case by using kernel tricks for handling the non-linearly separable data. The kernel
function, kðx; x0 Þ is used to map the data from an input space to a higher dimension space by using a nonlinear mapping
 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481 467

function, /ðxÞ. Then, within the higher dimension space, a linear optimal separating hyperplane is constructed for separating
two classes [24,25]. After solving the optimization problem, the decision function will be:
X
m
f ðxÞ ¼ fyi ai Kðx; xi Þg þ b ð9Þ
i¼1

SVMs originally developed for binary class problem, however, in reality more than two classes can exist, hence, the SVM is
extended to a multi-class problem. The multiclass problem can be solved by disintegrating this into several binary problems.
Methods like one-versus-one, one-versus-all and direct acyclic graph have been developed for handling the multiclass situ-
ation. Among these one-versus-one is most effective method for the practical situation owing to their less training period and
good generalization ability [26]. Thus in this study, one-versus-one method is adopted to predict the IM faults. For k-event,
one-versus-one method, constructs N ¼ kðk  1Þ=2 binary SVMs, where each SVMs is trained on data from two classes.
Now, the following optimization problem of binary classification can be solved for training example from ith and jth classes
1 ij T X T
min ðw Þ ðwÞij þ C nijt ðwij Þ ð10Þ
wij ;bij ;nij 2 t

T ij
ðwij Þ /ðxt Þ þ b P 1  nijt ; if yt ¼ i ð11Þ

T ij
ðwij Þ /ðxt Þ þ b P 1 þ nijt ; if yt ¼ j ð12Þ

nijt P 0; ð13Þ
where similar to the binary SVM problem, /ðxt Þ is a mapping function, ðxt ; yt Þ is the training sample, w and b are the weight-
ing factors, nij is the slack variable, and C is the penalty parameter. Now, for the future testing with unseen examples, a fol-
T ij
lowing voting approach is adopted here, where votes can be casted for each data samples, x. If f ðxÞ ¼ signððwij Þ Kðxn Þ þ b Þ
decides x to be in ith class, then vote for this class is added by one. Otherwise the jth class is increased by one. Then, it pre-
dicts x to be in the class with the maximum votes. This approach is also known as Max Wins strategy. If two classes have
identical votes, then it chooses a class with smaller index.

3. Experimental setup and feature extraction

Experimental setup and instrumentation for measurement of the vibration and current signals are described in this sec-
tion. The procedure of vibration and current data collection from the healthy and faulty IMs has also been presented.

3.1. Experimental setup

The experimental test rig consisted of a Machine Fault Simulator (MFS), a data acquisition system with signal monitor,
constant DC power source, a tri-axial accelerometer, and three AC current probes. The pictorial view and line diagram of
experimental test rig is shown in Figs. 2 and 3, respectively. The basic set-up of MFS consist of a test IM (0.37 kW, 50 Hz,
4-pole, and rated RPM-3450) coupled with a shaft using a flexible coupling, the shaft is coupled with a gear box by using
pulley-belt drive and a magnetic brake clutch is attached with the gear box to load the IM. Eight IMs were installed one

Test-motor Signal
monitor
Accelerometer
Constant
DC Source
MFS
Current
probes
DAQ

Fig. 2. The view of experimental test-rig.

468 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481

6
13
4 5
2
1
3

10
9 11
8 7

12

Fig. 3. A schematic diagram of experimental test-rig. (1 – Test IM, 2 – Coupling, 3 – Shaft, 4 – Bearings, 5 – Pulley-belt drive, 6 –Tri-axial accelerometer, 7 –
Gear box, 8 – Torque controller, 9 – Current probe, 10 – DAQ, 11 – Signal monitor, 12 – Base-plate, and 13 – Switch board).

by one on MFS to generate the data for ten faulty conditions. Since bearings are the only load carrying component of the
motor, the tri-axial accelerometer was placed on the top of the test motor close to one of the bearing and at motor shaft
end, to measure both axial and radial vibrations. Three AC current probes were attached with power leads of motor which
was easily accessible. Sensors were connected to the DAQ thorough proper instrumentation, which was used to acquire the
vibration as well as current signals. A constant DC power source was used, which was required for the tachometer in MFS. A
signal monitor with NI-Lab VIEW data acquisition software was used to analyze the data.
Accelerometers and current sensors were used to acquire vibration signals in three orthogonal directions and three-phase
motor current signals, respectively. Signals were recorded in time domain from nine faulty and one healthy IM (or no defect
motor, i.e. ND), the latter was used a benchmark for comparison with faulty ones. The raw data were acquired at the sam-
pling rate of 2000 Hz; hence 2000 points per second were collected to cover the high frequency band in which IM faults can
be identified. Total 300 raw data-sets (300  2000 sample points) were collected for each IM faulty condition.
In total, ten fault conditions of motor are considered in this work (illustrated in Table 1). Four motors were considered
with mechanical faults, i.e. bearing fault (BF), rotor misalignment fault (RMF), bowed rotor fault (BRF), unbalanced rotor fault
(URF). Five motors were considered with electrical faults, i.e. the broken rotor bar fault (BRBF), stator winding fault with
maximum and minimum resistance (MSWF and SWF, respectively), phase unbalance and single phasing fault with maxi-
mum and minimum resistance (MPUSPF and PUSPF, respectively. Here, two different severity levels of stator winding fault
and phase unbalance-single phasing was introduced by varying the resistance of winding. An external control box was con-
nected to one phase of the winding to vary the resistance (0–1 Ohm) of the same. Measurements were taken for various
operating conditions of IM, i.e. for a range of angular speeds, i.e. 10–Hz in 5 Hz interval, and also for three different external
load (or torque) on the motor, i.e. no, light and high-load (0 N m i.e., 0% of rated torque, 0.113 N m i.e., 0.11% of rated torque
and 0.565 N m i.e., 0.55% of rated torque). Raw data sets were stored in the DAQ at individual speeds and loads for various IM
faults for further processing. Fig. 4 shows the vibration in three orthogonal directions and the current in three phases at
40 Hz for no load from all fault conditions in the IM.

3.2. Signal feature extraction

In order to perform the perfect fault prediction of IM, input vectors (or features) must be selected appropriately. Input
features should contain the critical information of each fault condition. The raw data or signals directly from sensors contain

Table 1
The discription of motor fault conditions in this work.

No. Motor fault condition Fault description

1 No defect (ND) Healthy motor
2 Bearing fault (BF) Fault on outer race way of bearing near shaft end
3 Unbalanced rotor fault (URF) Motor with unbalanced rotor
4 Bowed rotor fault (BRF) Motor with centrally bent rotor
5 Rotor misalignment fault (RMF) Angular misalignment
6 Broken-rotor bar fault (BRBF) Broken rotor bar or end ring
7 Phase unbalance and single phasing fault with high resistance (MPUSPF) Severity level-1
8 Phase unbalance and single phasing fault with low resistance (PUSPF) Severity level-2
9 Stator winding fault with high resistance (MSWF) Severity level-1
10 Stator winding fault with low resistance (SWF) Severity level-2
 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481 469

(a) ND vibration signals (b) ND current signals

20 5
x-axis A-phase
0 y-axis 0 B-phase
z-axis C-phase
-20 -5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0. 1
(c) BF vibration signals (d) BF current signals
50 5

0 0

-50 -5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
(e) URF vibration signals (f) URF current signals
50 5

0 0

-50 -5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1

(g) BRF vibration signals (h) BRF current signals

50 5

0 0

-50 -5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
(i) RMF vibration signals (j) RMF current signals
20 5

0 0

-20 -5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1

(k) BRBF vibration signals (l) BRBF current signals

20 5

0 0

-20 -5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0. 1
(m) MPUSPF vibration signals (n) MPUSPF current signals
50 5

0 0

- 50 -5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
(o) PUSPF vibration signals (p) PUSPF current signals
50 5

0 0

-50 -5
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1

(q) MSWF vibration signals (r) MSWF current signals

20 10

0 0

-20 -10
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
(s) SWF vibration signals (t) SWF current signals
20 10

0 0

-20 -10
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
Time (in s) Time (in s)

Fig. 4. Time domain vibration (in m/s2) and current (in Amp) signals at 40 Hz for all faults.
470 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481

high noise and redundant information, and create the problem of dimensionality. Also it is not possible to feed the raw data
directly into the SVM model for training or testing purpose because handling of large data is a challenging task. In order to
overcome the problem with raw data, three effective statistical features, i.e. the standard deviation (r), skewness (v) and
kurtosis (j) are extracted from the raw time domain vibration and current signals. These features are very sensitive to
any changes in IM faults as per literatures [27,28]. In total, 300 feature data sets are extracted from the raw data sets.
Fig. 6 shows the scatter plot of three statistical features of acquired vibration and current signals from all faults (at 40 Hz
for high load). It shows that features are well clustered in case of mechanical faults using vibration signals; however, when
these faults are considered with current signals the features are not clustered well. In case of electrical fault using vibration
signals these features are not clustered well; however, when these faults are considered with current signals the features are
well clustered.

4. Induction motor faults prediction using MSVM

In order to perform the fault prediction, first total feature data sets are divided into training and testing data in a proper
ratio, i.e. 80% and 20%, respectively. The SVM has been trained with the help of 80% data sets and radial basis function (RBF)
kernel. The trained SVM is tested with the 20% data sets and the final results come in the form of testing or prediction per-
formance, i.e. the percentage of successfully classified or predicted data as shown in Fig. 5. The RBF kernel used as a kernel
function of MSVM, which is associated with the two parameters, i.e. the bound of Langrange multiplier (c) and the kernel
parameter (c) [15]. These parameters play crucial role in the SVM prediction performance. Inappropriate selection of these
parameters may create problem of under-fitting or over-fitting, hence reduce the prediction performance of SVM. The SVM
performance is also dependent upon the input vector, i.e. statistical features. In this study, c and c are chosen by hit-trial
method. We have provided a wide range of these parameter in the SVM faults prediction and values with the best prediction
performance are selected. The optimum value of c is equal to 1 for all case. In this section, various cases have been discussed
for prediction of the mechanical and electrical faults based on vibration, current, and vibration-current signals. Finally, the
prediction results are compared to show the best performance of MSVM.

4.1. Prediction of mechanical faults in IM

Here, following three cases are considered

Case A. Mechanical fault prediction based on vibration signal alone: Fig. 7 shows the prediction of mechanical faults
for various operating conditions of IM using vibration. The prediction performance for BF is 100% for all speeds as well as
mechanical load that means BF is perfectly predicted for any operating conditions of IM. The reason for this is, the generation
of intense vibration from BF due to repetitive impacts of the moving elements on the race and subsequently very different
vibration signatures or features than other faults as shown in Fig. 6(a). It is noted that, BF is one of the most critical IM fault.
In addition, the MSVM could able to predict the ND, BRF, URF and RMF successfully (average performance more than 90%) for
all operating conditions. In overall, mechanical faults are classified successfully using vibration signals. The reason for this is,
these faults generate independent and uncorrelated signatures or features (r; v; and j) with each other as shown in Fig. 6
(a). Table 2 shows the maximum, minimum and average value (over the considered speeds) of overall accuracy. It shows the
prediction performance of MSVM improves with the mechanical loading on IM but does not depend over the speed. Because
at higher loads, signal to noise ratio increases due to the amplification of vibration components. That leads to good vibration
signature from each faulty condition especially at high loads. Thus, it is beneficial to perform the prediction of mechanical
fault at high mechanical loads using vibration. For this case, the optimum value of gamma and computational time are 0.3
and 12 second, respectively.

Testing Final SVM Fault

Induction
Load data classifier prediction
motor

Optimal
gamma
Parametric
study for
Statistical Parameter gamma
Data estimation
feature
processing data
extraction
SVM
Training classifier
data with RBF
kernel

Fig. 5. Flow chart of the proposed fault prediction.

 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481 471

(a) Mechanical fault with vibration (b) Mechanical fault with current

(c) Electrical fault with vibration (d) Electrical fault with current
Fig. 6. Typical scatter plots for three features (r; v; and j) at 40 Hz for high load.
472 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481

(a) ND BF URF BRF RMF

performance,
100

Prediction
80
60

%
40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Average
(b) 100
80
performance,
Prediction

60
40
%

20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Average
(c) 100
performance,

80
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Average

Fig. 7. Mechanical fault prediction based on vibration signals alone.

Table 2
Mechanical fault prediction for various cases.

Fault type Signals Load Overall prediction, %

Maximum Minimum Average
Mechanical Vibration No 98.30 (at 20 Hz) 84.4 (at 10 Hz) 93.69
Light 99.32 (at 25 Hz) 89.93 (at 10 Hz) 95.44
High 100 (at 25 Hz) 94.91 (at 10 Hz) 97.48
Mechanical Current No 95.25 (at 20 Hz) 46.1 (at 15 Hz) 77.28
Light 93.89 (at 40 Hz) 47.11 (at 15 Hz) 74.23
High 87.11 (at 40 Hz) 35.25 (at 15 Hz) 69.68
Mechanical Vibration and current No 98.64 (at 20 Hz) 85.08 (at 10 Hz) 93.74
Light 99.32 (at 25 Hz) 89.49 (at 10 Hz) 95.39
High 100 (at 40 Hz) 93.55 (at 10 Hz) 96.84

Case B. Mechanical fault prediction based on current signal alone: Fig. 8 shows the prediction of mechanical faults for
various operating conditions of IM using current. Table 2 illustrate the maximum, minimum and average value (over the con-
sidered speeds) of overall accuracy. The prediction performance of MSVM get reduces (approximately 8% from the high to no
load) with the mechanical loading on IM unlike Case A. Thus, mechanical fault prediction using current should be perform
with no external load on motor. The performance does not depend over the speed, like Case A. The overall accuracy averages
decreased by 16%, 21%, and 27% corresponding to no, light, and high load, respectively, in comparison to Case A. In overall,
mechanical faults are not classified successfully with current signals because there is not much variation in the current sig-
nals of these faults as shown in Fig. 4 and due to this, features (r; v; and j) get overlapped as shown in the scatter plot in
Fig. 6(b). For this case; the optimum value of gamma and computational time are 300 or 3000 and 12 s, respectively.
Case C. Mechanical fault prediction based on vibration as well as current signals concurrently: Fig. 9 shows the pre-
diction of mechanical faults for various operating conditions of IM using both signal. The BF is perfectly predicted (nearly
100%) for all operating conditions of IM like Case A. Other faulty conditions like ND, URF, RMF and BRF could also be able
to predicted well at all operating conditions. Table 2 shows the maximum, minimum and average of overall accuracy over
the considered speeds. It shows that, the prediction performance improves with the mechanical loading on IM and does not
depend over speeds, similar to Case A. For this case; the optimum value of gamma and computational time are 0.3 and 16 s,
respectively.
Comparison of the performance of MSVM based on vibration, current and vibration-current signal for mechanical
faults: Fig. 10 shows the comparison of investigation of Cases A, B and C. Results show that the maximum of average pre-
diction achieves up to 97.48% based on the vibration signal alone at high load condition. Moreover, the MSVM with vibration
and with vibration-current signals, performs excellent although both of them exhibit nearly similar prediction performance
corresponding to all operating conditions. However, the MSVM performance get reduces with the current signal alone (i.e.,
 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481 473

(a) ND BF URF BRF RMF

100

performance,
Prediction
80
60

%
40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Average
(b) 100
performance,

80
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Average
(c) 100
80
performance,
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Average

Fig. 8. Mechanical faults prediction based on current signals alone.

(a) ND BF URF BRF RMF

performance,

100
Prediction

80
60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Average
(b) 100
80
performance,
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Average
(c) 100
80
performance,
Prediction

60
40
%

20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Average

Fig. 9. Mechanical faults prediction based on vibration as well as current signals concurrently.

Case B) than other two cases. Hence, it is not beneficial to perform the SVM prediction for mechanical faults based on current
signal alone. In both Case A and Case C, prediction improves by 3–4% at higher load with respect to no load condition.
Because higher signal-to-noise ratio occurs in the vibration at high loads. High signal-to-noise ratio means the good quality
signature, subsequently easy distinction of the signatures among different faults. In addition, the prediction performance
does not improve much by adding the current with the vibration signal. The unnecessary computation burden and time
of the training and prediction can be reduced by not considering the current signal in conjunction with the vibration. In over-
all, the mechanical fault can be predicted effectively with vibration signal alone using the MSVM. The reason for this is,
mechanical faults are physically related with vibration signals and different mechanical faults generate distinct vibration sig-
nals in different directions (axial or two orthogonal directions) as shown in Fig. 4. On the other hand, there is not much vari-
ation in the current signal due to mechanical faults as shown in Fig. 4.
474 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481

4.2. Prediction of electrical faults in IM

Here, following three cases are considered here:

Case D. Electrical fault prediction based on vibration signal alone: Fig. 11 shows the prediction of electrical faults for
various operating conditions of IM using vibration. The MSVM could predict all electrical faults satisfactorily (prediction
more than 75%) at all operating conditions, except for the MSWF and SWF, at the light and high loads. The reason for
satisfactory prediction is, these faults generate independent faulty vibration signatures and features; however, due to some

100 100
performance, %

90 90

performance, %
Prediction

Prediction
80 80
70 70
60 60
50 50
40 40

(a) Speed (b) Speed

100
Vibration Signal
performance, %

90
Prediction

80 Current signal
70
60 Vibration-current
50 signal
40
30

(c) Speed
Fig. 10. Mechanical fault predictions for (a) no load, (b) light and (c) high load.

(a) ND BRBF MPUSPF PUSPF MSWF SWF

100
performance,
Prediction

80
60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.
(b) 100
80
performance,
Prediction

60
40
%

20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.
(c) 100
performance,

80
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.

Fig. 11. Electrical faults prediction based on vibration signals alone.

 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481 475

overlapping in the fault information (as shown in Fig. 6(c)), the performance of SVM reduced. The rotor or stator winding
fault generally excite the torsional vibration but not necessarily the transverse vibration, still these fault could be predicted
satisfactorily (prediction more than 80%) using the axial and transverse vibration signals in this work. The average
predictions for the light and high loads are 74.57% and 58.35%, respectively for the MSWF, and 74.33% and 69.25%, respec-
tively, for the SWF. That means these faults cannot be classified successfully at higher loads using vibration due to overlap-
ping of information, and thus current signals is also necessary. Table 3 shows the maximum, minimum and average of overall
accuracy over the considered speeds. It shows that, the prediction performance gets reduced with the mechanical loading on
IM. The prediction is independent of speed. Thus, it is beneficial to perform the electrical fault prediction for this case, when
motor is running without any load. For this case, the optimum value of gamma and computational time are 0.3 and 14 s,
respectively. Figs. 11–17 do not display any result for the MPUSPF for 10 Hz at no load, and for 10 and 15 Hz at the light
and high loads, because the IM with the MPUSPF could not run at such a low speed, so raw data is not generated.
Case E. Electrical fault prediction based on current signal alone: Fig. 12 shows the prediction of electrical faults for
various operating conditions of IM using current. The electrical fault conditions could be predicted successfully (prediction
more than 90%) for all considered operating conditions except for the ND at some speeds. The overall prediction accuracy for
all electrical fault comes 100% for no load at all speeds, except for 10, 15 and 30 Hz. The SVM could able to predict the
MPUSPF, PUSPF, MSWF, and SWF perfectly (i.e., with 100% prediction) for all operating conditions, unlike Case D. The reason
is, these faults generate very independent and uncorrelated current signals and features as shown in Figs. 4 and 6(d), hence
no overlapping occurs in the features information of these faults. Different levels of severity in the stator winding fault, and
in phase unbalance and single phasing could be predicted perfectly by current signal alone. However, prediction perfor-
mance for the ND and the BRBF reduces at no-load and high load due to some overlapping of features of the ND and the BRBF
as shown in Fig. 6(d). Table 3 shows the maximum, minimum and average of overall accuracy over the considered speeds. It

Table 3
Electrical fault prediction for various cases.

Fault type Signals Load Overall prediction, %

Maximum Minimum Average
Electrical Vibration No 94.35 (at 35 Hz) 82.84 (at 40 Hz) 87.82
Light 90.50 (at 15 Hz) 70.33 (20 Hz) 80.84
High 83.61 (at 35 Hz) 70.90 (at 20 Hz) 79.29
Electrical Current No 100 (at 15, 20, 35, and 40 Hz) 73.44 (at 15 Hz) 92.96
Light 99.71 (at 40 Hz) 87.45 (at 15 Hz) 95.76
High 98.87 (at 40 Hz) 76.61 (at 15 Hz) 91.05
Electrical Vibration and current No 98.30 (at 20 Hz) 87.79 (10 Hz) 94.05
Light 91.86 (at 15 Hz) 79.66 (10 Hz) 86.2
High 88.70 (at 40 Hz) 76.55 (at 20 Hz) 82.83

(a) ND BRBF MPUSPF PUSPF MSWF SWF

performance,

100
Prediction

80
60
%

40
20
0
10 Hz 15 Hz 20 Hz 25 Hz 30 Hz 35 Hz 40 Hz Avg.
(b) 100
performance,

80
Prediction

60
%

40
20
0
10 Hz 15 Hz 20 Hz 25 Hz 30 Hz 35 Hz 40 Hz Avg.
(c) 100
80
performance,
Prediction

60
%

40
20
0
10 Hz 15 Hz 20 Hz 25 Hz 30 Hz 35 Hz 40 Hz Avg.

Fig. 12. Electrical faults prediction based on current signals alone.

476 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481

(a) ND BRBF MPUSPF PUSPF MSWF SWF

100

performance,
Prediction
80
60

%
40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.
(b) 100
80
performance,
Prediction

60
40
%

20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.
(c) 100
80
performance,
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.

Fig. 13. Electrical faults prediction based on vibration as well as current signals concurrently.

100 100
90
performance, %

90
performance, %

Prediction
Prediction

80 80
70 70
60 60
50 50

(a) Speed (b) Speed

100
performance, %

90
Prediction

Vibration Signal
80
70 Current signal
60 Vibration-current
50 signal

(c) Speed

Fig. 14. Electrical faults prediction for (a) no load, (b) light load and (c) high load.

is noted that, the prediction performance is higher at light and no load as compared to the full load. Thus, it is possible to
perform fault diagnosis at no and light load, which is the main challenge of traditional techniques. For this case; the optimum
value of gamma and computational time are 300 and 14 s, respectively.
Case F. Electrical fault prediction based on the vibration as well as current signals concurrently: Fig. 13 shows the
prediction of electrical faults for various operating conditions of IM. The SVM could able to predict all electrical fault success-
fully (prediction is more than 90%) at no load, and satisfactorily (prediction is more than 80%) at the light and high load con-
ditions except for PUSPF at 10 and 15 Hz. Also different level of severity in the stator winding fault, and in phase unbalance
and single phasing could be predicted perfectly as like in Case E. Table 3 shows the maximum, minimum and average of
overall accuracy over the considered speeds. The prediction performance of MSVM get reduces with the mechanical loading
 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481 477

(a) ND BRBF MPUSPF PUSPF MSWF SWF BF URF BRF RMF

100

performance,
Prediction
80
60

%
40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.
(b) 100
performance,

80
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.
(c) 100
80
performance,
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.

Fig. 15. Prediction of all fault together based on vibration signal alone.

(a) ND BRBF MPUSPF PUSPF MSWF SWF BF URF BRF RMF

100
performance,
Prediction

80
60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.
(b) 100
performance,

80
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.
(c) 100
performance,

80
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.

Fig. 16. Prediction of all fault together based on current signal alone.

on IM like Case B and Case D and does not depends over speed like any other cases. The percentage improvement is 8–12% at
no load in comparison to light load and high load. Thus, for this case, the fault prediction should be performed when there is
no external load on motor. For this case, the optimum value of gamma and computational time are 0.3 and 18 s, respectively.
Comparison of the performance of MSVM based on vibration, current and vibration-current signal for electrical
faults: Fig. 14 illustrate the comparison of investigation of three cases (i.e., Cases D, E and F) for the no, light and high load
condition. Results show that the maximum of average prediction accuracy of 95.76% achieves using SVM with the current
signal alone (i.e., Case D) at light load condition. For no load condition, the MSVM performs best based on vibration in con-
junction with current signals (case F) than two other cases (Cases D and E) but performance at this is nearly similar to anal-
ysis based on current signals alone (i.e., Case D). For the light and high load conditions, the MSVM performs best based on
current signals alone (Cases E) than two other cases (Cases D and F). The MSVM performance get reduces based on vibration
signal alone (Case D) than other two cases (Cases E and F). Also with vibration, much difference in performance occurs cor-
responding to different load unlike with current. Thus, it is not beneficial to perform the SVM prediction based on vibration
478 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481

(a) ND BRBF MPUSPF PUSPF MSWF SWF BF URF BRF RMF

100

performance,
Prediction
80
60

%
40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.
(b) 100
performance,

80
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.
(c) 100
80
performance,
Prediction

60
%

40
20
0
10Hz 15Hz 20Hz 25Hz 30Hz 35Hz 40Hz Avg.

Fig. 17. Prediction of all fault together based on vibration as well as current signals concurrently.

signal alone for electrical faults. Hence, it is possible to use vibration-current signal as a reliable media to predict electrical
faults only in specific operating condition, i.e. no load. On the other hand, current signals are a best indicator of electrical
faults for all operating conditions. The reason for this is the high variation in current signal (as shown in Fig. 4) depending
on the type of electrical faults. For the electrical fault prediction, there is no need to train and test the SVM with vibration
signal alone or in conjunction with current signal. More signals (i.e., vibration and current) means more feature data sets;
hence this can be an unnecessary burden on SVM.

4.3. Prediction of mechanical as well as electrical faults of IM all together

Here, following three cases are considered here;

Case G. Mechanical as well as electrical fault prediction all together based on vibration signal alone: Fig. 15 shows
the prediction of all faults together for various operating conditions of IM based on vibration signal alone. The MSVM could
be able to predict the BRBF up to 93.22% at no load and high load that means the BRBF also generate significant vibration
signatures. The individual prediction performance of BF is nearly 100% for various speed as well as mechanical loading that
means BF is perfectly predicted for all operating conditions of IM even when all faults are considered simultaneously. The
MSVM could predict the ND, BRBF and all mechanical faults successfully (prediction is more than 90%) at all speeds and
loads. The MSVM could predict the MPUSPF, PUSPF, MSWF and SWF satisfactorily (prediction near about 80%) for no load.
However for the light and high load conditions the MSVM performance for the PUSPF, MSWF, and SWF get reduces. That
means fault severity of the stator winding fault, and phase unbalance and single phasing could be predicted at no load.
Table 4 shows the maximum, minimum and average of overall accuracy over the considered speeds. It is noted that, the pre-
diction performance of MSVM get reduces with the mechanical loading on IM like Case B, D and F and does not depends over
speed like any other case. For this case; the optimum value of gamma and computational time are 0.3 and 20 s, respectively.

Table 4
All fault (mechanical and electrical) prediction for various cases.

Fault type Signals Load Overall prediction, %

Maximum Minimum Average
All faults Vibration No 94.57 (at 35 Hz) 81.54 (10 Hz) 88.86
Light 88.70 (at 15 Hz) 77.96 (at 10 Hz) 84.99
High 87.62 (at 35 Hz) 78.71 (at 10 Hz) 84.38
All faults Current No 90 (at 20 Hz) 57.96 (at 15 Hz) 79.27
Light 84.40 (at 15 Hz) 53.86 (at 40 Hz) 73.78
High 90.5 (at 40 Hz) 45.57 (at 15 Hz) 76.23
All faults Vibration and current No 98.3 (at 20 Hz) 84.18 (at 10 Hz) 93.28
Light 94.4 (at 35 Hz) 82.10 (at 10 Hz) 88.95
High 93.05 (at 40 Hz) 82.67 (at 10 Hz) 88.28
 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481 479

Case H. Mechanical as well as electrical fault prediction all together based on current signal alone: Fig. 16 shows the
prediction of all faults together for various operating conditions of IM based on current signal alone. The MSVM could able to
predict the stator winding fault, the phase unbalance & single phasing with min and max resistance perfectly (or nearly
100%) at all operating conditions of IMs (i.e., varying speed and torque). That means different level of severity in the stator
winding fault, and in the phase unbalance and single phasing could be predicted perfectly at varied load or speed of the IM.
Also MSVM could be able to predict the BRBF satisfactorily at high load (average performance up to 67.31% at high load). The
MSVM could predict considered all electrical faults except the ND, BRBF and BRF successfully (prediction is more than 90%)
at all speeds and loads; however, in some cases the ND also could be classified successfully. The average individual predic-
tion performance of the BRF is nearly 40–50% at any operating conditions that means the MSVM based on current signal
alone could not able to predict BRF well when all faults are considered together unlike the Case G, where it could classified
successfully (or more than 90%). Other fault conditions, like the URF and RMF, could be able to predict satisfactorily (predic-
tion is more than 75%) at each operating conditions. The individual performance for BF is nearly 75%-for the no and high
mechanical load conditions, that mean BF is predicted satisfactorily based on the current signal alone even when all faults
are considered simultaneously, unlike Case A, Case C and Case G where BF could be classified perfectly (or 100%) at all oper-
ational conditions. Table 4 shows the maximum, minimum and average of overall accuracy over the considered speeds. It
shows that, the prediction performance of MSVM is independent of mechanical loading on IM (like Case E) and does not
depend over the speed, like any other cases. Also the overall performance reduces as compared to previous case, i.e. Case
G. For this case; the optimum value of gamma and computational time are 300 and 20 s, respectively.
Case I. Mechanical as well as electrical fault prediction all together based on vibration as well as current signals con-
currently: Fig. 17 shows the prediction of all faults together for various operating conditions of IM based on the vibration as
well as current signals, concurrently. The individual prediction performance of BF is nearly 100% for each considered speeds
as well as loading that means BF is perfectly predicted even when all faults are considered simultaneously. The MSVM could
able to predict the MPUSPF, PUSPF, MSWF, and SWF successfully (prediction is more than 90%) at each operating conditions
and sometime perfectly. The prediction of phase unbalance and single phasing is better than stator winding faults at most of
operating conditions. Also the MSVM could be able to predict the BRBF well at various loads (average performance up to
92.98% at the light and high loads) unlike the Case H. The average individual prediction performance of the BRF increases
up to 96.61% at high load, unlike Case H (prediction was 40–50% at each operating conditions based on current signal alone).
Hence, for the ND, BRBF and BRF the prediction of all faults together, it is necessary to train the SVM with the vibration signal
in conjunction with current signals. Other faulty conditions, like ND, URF and RMF, could be able to predict successfully (pre-
diction is more than 90%) at each of the operating conditions. Table 4 shows the maximum, minimum and average of overall
accuracy over the considered speeds. Hence, the prediction performance of MSVM reduces with the mechanical loading on
IM like Case B, Case D, Case E and Case G. For this case; the optimum value of gamma and computational time are 0.3 and
25 s, respectively.
Comparison of the performance of MSVM based on vibration, current and vibration-current signal for all faults:
Fig. 18 illustrate the comparison of Case G, Case--H and Case I for the no, light and high load condition, respectively. The
results show that for prediction of all fault together, the maximum of average prediction accuracy achieves up to 93.28%

100 100
performance, %

performance, %

90 90
Prediction

Prediction

80 80
70 70
60 60
50 50

(a) Speed (b) Speed

100
Prediction perforance,

90
80
70 Vibration Signal
%

60 Current signal

50 Vibration-current signal

(c) Speed

Fig. 18. Prediction of all fault together for (a) no load, (b) light load and (c) high load.
480 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481

at no load using vibration-current signals (Case I). The MSVM performs best based on the vibration signal in conjunction with
current i.e., Case I, at each considered operating conditions of IM. The MSVM also perform better for the fault prediction
based on the vibration signal alone (Case G), but performance gets reduced much for the prediction based on the current
signal alone (Case H). Thus, it is not beneficial to perform the prediction of all faults together using the MSVM based on cur-
rent signal alone (Cases H). Hence, it is effective to predict all faults together using the MSVM with vibration signals in con-
junction with current signals. The reason is mechanical faults generate distinct vibration signals in three orthogonal
directions with each other and electrical fault creates high variation in three phases of current with each other. So it is effec-
tive to collect vibration as well as current signals when both type of faults are considered for prediction.
From the results presented based on the SVM, it can stated that, for any operating conditions of IM, the vibration signa-
ture alone is sufficient and sensitive to the mechanical fault prediction and the current signature alone is sufficient and sen-
sitive to electrical fault prediction. However, both vibration and current signatures are necessary for the successful
prediction of the mechanical and electrical faults, together.

5. Conclusions

In the present work, the fault prediction methodology of the induction motor using the MSVM classifiers has been suc-
cessfully developed. Firstly, three important statistical features (i.e., r; v; and j) are extracted from time domain vibration
and current signals to reduce the redundant information and avoid the problem of dimensionality. After the feature extrac-
tion, the prediction of faults is done for various cases using one-versus-one MSVM with the RBF kernel. The optimal kernel
parameters (c and c) are chosen by hit-trial method. Then by using this kernel and with optimal c and c, a classification
model is built for further fault prediction.
The investigation concludes that the MSVM could be able to predict all mechanical faults perfectly based on vibration
signal alone, hence there is no need to train and test the SVM with current signal alone or in conjunction with vibration
for the mechanical fault prediction. For effective electrical fault predictions, it is more advantageous to perform SVM predic-
tion with current signal alone or vibration in conjunction with current signals, concurrently, when the IM is running without
any mechanical load. For the light and high mechanical loads, current signals alone could be able to predict electrical faults
perfectly. Hence for any load, there is no need to train and test the SVM with vibration signal alone or in conjunction with
current signal for electrical faults prediction. And when it comes to the prediction of all faults together (mechanical and elec-
trical), the MSVM performs effectively based on the vibration-current signal at all considered operating conditions of IM,
hence it is advantageous to train and test the SVM with vibration signal in conjunction with current signal for prediction
of all faults together. The fault prediction can be extended in the frequency, and time-frequency domain. The prediction per-
formance should be load independent so the investigation can be further extended because it is not always possible to find
the test data at the same load as the training data.

Acknowledgements

The support from LIBSVM tool is gratefully acknowledged, which is freely available online at http://www.csie.ntu.edu.tw/
_cjlin/libsvm [29]. The authors would also like to thank Dr. Dhruba Jyoti Bordoloi, Jr. Technical Officer at IIT Guwahati for
extending his support in carrying out the relevant experimentation.

References

[1] M.R. Mehrjou, N. Mariun, M.H. Marhaban, N. Misron, Rotor fault condition monitoring techniques for squirrel-cage induction machine—a review, Mech.
Syst. Signal Process. 25 (8) (2011) 2827–2848.
[2] M. Eftekhari, M. Moallem, S. Sadri, A. Shojaei, Review of induction motor testing and monitoring methods for inter-turn stator winding faults, in: IEEE
21st Iranian Conference on Electrical Engineering (ICEE), 2013, pp. 1–6.
[3] H. Henao, G.A. Capolino, M. Fernandez-Cabanas, F. Filippetti, C. Bruzzese, E. Strangas, S. Hedayati-Kia, Trends in fault diagnosis for electrical machines:
a review of diagnostic techniques, IEEE Ind. Electron. Mag. 8 (2) (2014) 31–42.
[4] V.T. Tran, B.S. Yang, M.S. Oh, A.C.C. Tan, Fault diagnosis of induction motor based on decision trees and adaptive neuro-fuzzy inference, Exp. Syst. Appl.
36 (2) (2009) 1840–1849.
[5] K. Bacha, S..Ben. Salem, A. Chaari, An improved combination of Hilbert and Park transforms for fault detection and identification in three-phase
induction motors, Int. J. Electr. Power Energy Syst. 43 (1) (2012) 1006–1016.
[6] X. Li, A. Zheng, X. Zhang, C. Li, L. Zhang, Rolling element bearing fault detection using support vector machine with improved ant colony optimization,
Measurement 46 (2013) 2726–2734.
[7] L. Li, C.K. Mechefske, Induction motor fault detection & diagnosis using artificial neural networks, Int. J. COMADEM 9 (3) (2006) 15.
[8] N.T. Nguyen, H.H. Lee, An application of support vector machines for induction motor fault diagnosis with using genetic algorithm, in: Advanced
Intelligent Computing Theories and Applications. With Aspects of Artificial Intelligence, Springer, Berlin Heidelberg, 2008, pp. 190–200.
[9] N.T. Nguyen, H.H. Lee, J.M. Kwon, Optimal feature selection using genetic algorithm for mechanical fault detection of induction motor, J. Mech. Sci.
Technol. 22 (3) (2008) 490–496.
[10] L.M.R. Baccarini, V.V. Rocha e Silva, B.R. De Menezes, W.M. Caminhas, SVM practical industrial application for mechanical faults diagnostic, Exp. Syst.
Appl. 38 (6) (2011) 6980–6984.
[11] P. Chattopadhyay, P. Konar, Feature extraction using wavelet transform for multi-class fault detection of induction motor, J. Inst. Eng. (India): Ser. B
(2014) 1–9.
[12] A. Widodo, B.S. Yang, Wavelet support vector machine for induction machine fault diagnosis based on transient current signal, Exp. Syst. Appl. 35 (1)
(2008) 307–316.
 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 94 (2017) 464–481 481

[13] S.B. Salem, K. Bacha, A. Chaari, Support vector machine-based decision for induction motor fault diagnosis using air-gap torque frequency response, Int.
J. Comput. Appl. 38 (2012).
[14] Z. Zhou, J. Zhao, F. Cao, A novel approach for fault diagnosis of induction motor with invariant character vectors, Inf. Sci. 281 (2014) 496–506.
[15] A. Widodo, B.S. Yang, T. Han, Combination of independent component analysis and support vector machines for intelligent faults diagnosis of induction
motors, Exp. Syst. Appl. 32 (2) (2007) 299–312.
[16] J.D. Martínez-Morales, E. Palacios, D.U. Campos-Delgado, Data fusion for multiple mechanical fault diagnosis in induction motors at variable operating
conditions, in: IEEE 2010 7th International Conference on Electrical Engineering Computing Science and Automatic Control (CCE), 2010, September, pp.
176–181.
[17] V.T. Tran, B. Yang, M. Oh, A. Chit, C. Tan, Fault diagnosis of induction motor based on decision trees and adaptive neuro-fuzzy inference, Exp. Syst. Appl.
36 (2) (2009) 1840–1849.
[18] A. Garcia-Perez, R. de Jesus Romero-Troncoso, E. Cabal-Yepez, R.A. Osornio-Rios, The application of high-resolution spectral analysis for identifying
multiple combined faults in induction motors, IEEE Trans. Ind. Electron. 58 (5) (2011) 2002–2010.
[19] V.T. Tran, F. Althobiani, A. Ball, B. Choi, Expert systems with applications an application to transient current signal based induction motor fault
diagnosis of Fourier – Bessel expansion and simplified fuzzy ARTMAP, Exp. Syst. Appl. 40 (13) (2013) 5372–5384.
[20] S. Ergin, A. Uzuntas, M.B. Gulmezoglu, Detection of stator, bearing and rotor faults in induction motors, Proc. Eng. 30 (2012) 1103–1109.
[21] H. Nakamura, Y. Yamamoto, Y. Mizuno, Diagnosis of electrical and mechanical faults of induction motor, in: 2006 IEEE Conference IEEE Electrical
Insulation and Dielectric Phenomena, 2006, pp. 521–524.
[22] W.J. Park, S.H. Lee, W.K. Joo, J.I. Song, A mixed algorithm of PCA and LDA for fault diagnosis of induction motor, in: Advanced Intelligent Computing
Theories and Applications, With Aspects of Artificial Intelligence, Springer, Berlin Heidelberg, 2007, pp. 934–942.
[23] V. Vapnik, E. Levin, C.Y. Le, Measuring the VC-dimension of a learning machine, Neural Comput. 6 (5) (1994) 851–876.
[24] V.N. Vapnik, An overview of statistical learning theory, IEEE Trans. Neural Netw. 10 (5) (1999) 988–999.
[25] C.J. Burges, A tutorial on support vector machines for pattern recognition, Data Min. Knowl. Discov. 2 (2) (1998) 121–167.
[26] C.W. Hsu, C.J. Lin, A comparison of methods for multiclass support vector machines, IEEE Trans. Neural Netw. 13 (2) (2002) 415–425.
[27] R. Tiwari, D.J. Bordoloi, Optimization of support vector machine based multi-fault classification with evolutionary algorithms from time domain
vibration data of gears, Proc. Inst. Mech. Eng., Part C: J. Mech. Eng. Sci., 2013, pp. 0954406213477777.
[28] B.S. Yang, T. Han, Z.J. Yin, Fault diagnosis system of induction motors using feature extraction, feature selection and classification algorithm, JSME Int. J.
Ser. C 49 (3) (2006) 734–741.
[29] C.C. Chang, C.J. Lin, LIBSVM: a library for support vector machines, ACM Trans. Intell. Syst. Technol. (TIST) 2 (3) (2011) 27:1–27:27.