Electric Power Systems Research 127 (2015) 249258

Contents lists available at ScienceDirect

Electric Power Systems Research

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

A comprehensive evaluation of intelligent classiers for fault

identication in three-phase induction motors
Rodrigo H. Cunha Palcios a,b , Ivan Nunes da Silva a , Alessandro Goedtel b, ,
Wagner F. Godoy a,b
a
University of So Paulo (USP), So Carlos School of Engineering, Department of Electrical Engineering, Av. Trabalhador So Carlense, 400, Centro,
13.566-590 So Carlos, SP, Brazil
b
Federal Technological University of Paran (UTFPR), Department of Electrical Engineering, Av. Alberto Carazzai, 1640, Centro, 86.300-000 Cornlio
Procpio, PR, Brazil

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

Article history: Three-phase induction motors are the key elements of electromechanical energy conversion for a vari-
Received 18 August 2014 ety of industrial sectors. The ability to identify motor faults before they occur can reduce the risks in
Received in revised form 8 May 2015 decisions regarding machine maintenance, lower costs, and increase process availability. This article
Accepted 8 June 2015
proposes a comprehensive evaluation of pattern classication methods for fault identication in induc-
Available online 26 June 2015
tion motors. The methods discussed in this work are: Naive Bayes, k-Nearest Neighbor, Support Vector
Machine (Sequential Minimal Optimization), Articial Neural Network (Multilayer Perceptron), Repeated
Keywords:
Incremental Pruning to Produce Error Reduction, and C4.5 Decision Tree. By analyzing the amplitudes
Three-phase induction motor
Pattern recognition
of current signals in the time domain, experimental results with bearing, stator, and rotor faults are
Rotor tested using different pattern classication methods under varied power supply and mechanical loading
Stator conditions.
Bearing 2015 Elsevier B.V. All rights reserved.
Fault

1. Introduction
Three-phase induction motors (TIMs) are frequently used in var-
ious industrial sectors. These machines are characterized by their
low acquisition and maintenance costs, adaptability to different
mechanical loads, and robustness in harsh working environments.
They consume more than 60% of the electricity in the industrial sec-
tor [1] and are the primary means of transforming electrical energy
into mechanical driving [2].
Like all electrical machines, these motors need proper mainte-
nance, as faults can adversely affect production processes and cause
substantial losses for the industries where they are used. According
to [3], the correct diagnosis and early detection of incipient faults
in TIMs results in quick unscheduled maintenance, whereby the
machines are ofine for brief periods only.
Operational faults related to TIMs may be divided into two
major groups: electrical and mechanical faults. Electrical faults are
mainly characterized by problems regarding the stator winding,
rotor winding (found on some motor models), broken rotor bars,
Corresponding author. Tel.: +55 43 3520 4096; fax: +55 43 3520 4010.
E-mail address: agoedtel@utfpr.edu.br (A. Goedtel).
broken rotor rings, and faulty connections, while mechanical faults
include problems with bearings, eccentricity of the shaft or rotor,
and wear coupling misalignment [3]. Of the faults reported in the
literature, the work of [4,5] states that the bearing, rotor, and stator
are responsible for about 7888% of breakdowns in electric motors.
Various intelligent approaches for pattern recognition have
recently been used to identify faults in induction motors. Intel-
ligent techniques such as Naive Bayes (NB), k-Nearest Neighbor
(KNN), Support Vector Machine with Sequential Minimal Optimiza-
tion (SVM), Articial Neural Network (ANN), Repeated Incremental
Pruning to Produce Error Reduction (RIP), and C4.5 Decision Tree
(C4.5) have been effective at pattern recognition in many applica-
tions reported in the literature.
Some works in the literature are related to pattern identication
based on the Bayes theorem. The research conducted by Hajiagha-
jani et al. [6], where the current signals in the frequency domain
were used as input to supply the system, proposed using this
method to detect eccentricity faults in a DC motor. In the research
presented in [7], the NB, KNN and SVM pattern recognition methods
are tested to check induction motor faults such as broken rotor bars,
tilted rotors, unbalanced rotors, short-circuits in the stator winding,
and bearings aws. Analysis is based on the motor current data to
extract the frequency domain (Hilbert transform), statistics (RMS,

http://dx.doi.org/10.1016/j.epsr.2015.06.008
0378-7796/ 2015 Elsevier B.V. All rights reserved.
250 R.H. Cunha Palcios et al. / Electric Power Systems Research 127 (2015) 249258
skewness, kurtosis, entropy, and crest factor), and feature selection
(such as PCA and ICA). The authors specied that these tests were
conducted under different load variations and by using an inverter
fed.
The KNN method has been employed in several pattern recog-
nition studies. For diagnosing faults in TIMs, the work developed
by Ondel et al. [8] proposed a solution to detect broken bars under
various levels of mechanical load, based on voltage and current sig-
nal in the frequency domain. In [9] the authors sought to diagnose
eccentricity faults in synchronous motors by using wavelet decom-
position and principal component analysis. This approach analyzed
the stator current signature in the frequency domain, using a fre-
quency inverter for speed control and torque variation for the bench
tests. The KNN and FSVM (Fuzzy Support Vector Machine) methods
were used to classify the fault.
The SVM method is recognized in the literature as an effective
method for solving problems which require pattern recognition.
In recent years several works have been published about detec-
ting aws in induction motors, such as that presented by Konar
and Chattopadhyay [4], where a system was developed to identify
bearing faults by using SVM plus Wavelet. Motor vibration sig-
nals were used as SVM input, which performs the classication
conditions. The methodological structure comprised a frequency
inverter, while the experiments were conducted under three sepa-
rate loading conditions: no load, half of the rated load, and nominal
load. The experimental results also demonstrated the results com-
pared with SVM and ANN. Another recent work [10] developed a
method for using magnetic signals to identify faults such as unbal-
anced currents and short-circuits in the stator windings of TIMs.
The results used the numerical nite element method and Arti-
cial Neural Networks to generate simulation and experimental data.
The SVM method was also used to collect data from the vibration
signal, with sinusoidal supply, for detecting bearings faults (balls,
inner race, and outer race) [11]. This research did not consider the
imbalance of supply voltages in the machine or load variation in its
experiments.
The development of data-acquisition techniques is also inves-
tigated. The purpose of article [12] was to employ a low-cost
technique using multi-sensor and wireless transmission in order to
monitor the condition of faults in bearings and eccentricity under
motor-fed induction with variable frequency drive. For this, the
authors used the stator current and acoustic signs at frequency
domain (FFT and HilbertHuang transform) and QDA (quadratic
discriminant analysis), where the SVM method performs the clas-
sication. The inner race, outer race, cage, and bearing lubrication
were investigated for bearing failure problems. Other important
studies related to using SVM to identify faults with induction
motors appear in [1315].
Decision trees are simple representations of knowledge. They
can be applied to learning systems and are widely used in classi-
cation algorithms. In the research of TIM faults, some recent studies
[16] describe an investigation of bearings faults in TIMs by using
methods based on decision trees, called Random Forest and C4.5
Decision Tree. Mechanical vibration signals are used as the input of
these methods in the time frequency domain and considering sinu-
soidal supply. To complete the pre-processing system, the Principal
Component Analysis (PCA) and Linear Discriminant Analysis (LDA)
are used to reduce the size and complexity of the feature set. The
results indicate that the Random Forest method is the most efcient
for this work. The research presented in [17] proposed a method
that uses the measurement of a phase current signal, through the
Hilbert transform, to build a representation of a graphic image,
allowing the detection of edges in images and pattern recognition of
faults related to broken rotor bars and connectors in TIMs with sinu-
soidal supply and disregarding loading variation. This same study
evaluated the following classiers methods: C4.5, Articial Neural
Networks, Gaussian mixture models, negative selection algorithm
based on genetic algorithm, and articial immune classier with
learning particle swarm.
ANNs are computational techniques which present a model
inspired by the neural structure of intelligent organisms, and which
acquire knowledge through experience. This method is character-
ized by its exibility and range of architecture. Several studies [18]
regarding fault classication in induction motors have recently
been published. The authors presented two ANN-based schemes
in order to identify and diagnose faults in induction motor rotors
through the use of self-organizing maps. The rst scheme used a
single phase current signal in the frequency domain to diagnose
the load unbalance and shaft misalignment. This neural network
was trained and validated using data generated through simula-
tion. The second scheme was carried out by analyzing the frequency
of active and reactive instantaneous power in order to detect
and classify broken rotor bars and oscillating loads faults. The
research conducted by Ertunc et al. [5], the authors presented a
new methodology for monitoring the condition of the bearings in
a three-phase induction motor. The tests presented in this paper
used the ANN and Neuro-Fuzzy (ANFIS) models, in the time and
frequency domain, to generate data resulting from the vibration
and stator current of an induction motor with sinusoidal supply.
The authors inserted predened bearing aws into the internal
and external tracks in order to detect the severity of the fault. In
[19], proposed a set of hybrid intelligent models for condition-
monitoring of induction motors under the following mechanical
load conditions: 25%, 50%, 75%, and nominal torque, considering the
sinusoidal supply. To classify the faults, the authors used the Neu-
ral Network Fuzzy Min-Max (FMM) and Random Forest (RF) with
inputs fed by the stator current signals through the MCSA (Motor
Current Signature Analysis) method in the frequency domain. The
experimental results also considered noisy environment situa-
tions. One recently employed technique, found in [20], applied
the analysis of acoustic data from an induction motor, captured
simultaneously by multiple microphones. Correlations and analy-
sis based on signal processing by wavelet transform are used in the
current study as the inputs to extract features. The pattern classi-
cation method used is based on Konohen self-organizing maps, and
faults in bearings, broken rotor bars, and short-circuits in the sta-
tor are investigated. Other ANN-related investigations of induction
motors appear in [2123]. Recently in Ref. [24], researchers focused
on the development of a solution to identify stator inter turn short-
circuit in a permanent magnet synchronous motor. A Multilayer
Perceptron network was used for the diagnosis and classication
of different levels of short-circuit and check its severity under vari-
able speed and load torque conditions. The pre-processing in the
frequency domain is processed via FFT.
Furthermore, statistical methods are also studied as an alterna-
tive to intelligent systems as can be observed in the work of [25]
where this tools were tested to diagnose rotor broken bars faults
in three phase induction motors (0.75 and 1.1 kW) by means of
features extracted from the stator current by the power spectral
density (PSD) using the FFT. For the experiments, the motors were
driven by frequency inverters under different load torque condi-
tions.
Image-based methods are also currently being investigated for
fault identication in induction motors such as Ref. [26] which
presents a methodology based on segmentation of thermographic
images for fault diagnosis in induction motors. According to the
authors, this methodology diagnoses broken bars in the rotor,
bearings, mechanical unbalance, misalignment and also unbalance
voltage.
The primary purpose of this study is to analyze the behavior of
the Naive Bayes, C4.5, RIPPER, KNN, SVM/SMO, and ANN/MLP classi-
cation methods under various operating conditions and induction
 R.H. Cunha Palcios et al. / Electric Power Systems Research 127 (2015) 249258
motor faults. These conditions are associated with a varied combi-
nation of mechanical loads through analysis of the amplitudes of
the stator current in the time domain, considering voltage unbal-
ance and torque variation. All intelligent methods are compared
with the same sets of samples gathered from the laboratory test
bench in order to characterize the faults in the stator, rotor, and
bearings of induction motors operating directly from the power
supply at 60 Hz. This article is divided into ve sections. Section 2
discusses the basics concepts of intelligent classiers employed in
this work. Section 3 presents the methodology applied. Results are
given in Section 4, while conclusions and future work are found in
Section 5.
2. Pattern identication
Intelligent systems have been developed as a exible approach
which can be applied to various types of problems with electrical
machines, particularly those related to the identication of fault
patterns. The pattern classiers considered in this work are based
on rules, decision trees, elements of neighborhood support vectors,
Articial Neural Networks, and Bayesian networks. These methods
are characterized by their use of training data and validation to
create a classication model, which is subsequently used to classify
a new instance of related data. A brief description of the algorithms
in this paper is presented in the following subsections.
2.1. k-Nearest Neighbors classier
The KNN classier [27] is a non-parametric method based on the
Nearest Neighbor (NN) algorithm [28]. The KNN proposes a modi-
cation to the original algorithm during testing and classication,
where the algorithm makes use of the k-Nearest Neighbors, unlike
NN, which only uses the nearest neighbor.
The KNN method classies a data sample according to the dis-
tance between the sample and certain pre-labeled training samples.
If the majority of samples closest to the unknown sample are from
a specied class, the sample will be assigned to that class. Using
this method, the independent axes are divided into partitions (i ).
Thus, the probability density function is evaluated for each of these
partitions, as shown by (1):
ni
pi =
Ni
where ni is the number of the data sample in i and N is the total
number of data points.
To evaluate the probability of a test sample, a partition must be
centered at a point ni , xed to changes, so that ni (K) data points
to place it in that partition. Under this procedure, the probability
of the test sample is evaluated adaptively, with no prior hypothe-
sis, except K, which is chosen so that the best results are obtained
with the training data. From the Bayes theorem, the subsequent
probability of a test sample is obtained using (1), as seen in (2):
Kk
p(Ck |Z) =
K
where Kk is the number of data points from the class of order k
[29]. This approach does not require a training phase. However, to
classify a test sample, all training data points should be saved and
the distances between the test samples and all the training sam-
ples should be calculated and classied. Therefore, the k-Nearest
Neighbors are selected to make the nal decision.
The similarity, or proximity, is dened in terms of the distance
between the bodies. Accordingly, various distances measures, such
251
as the Euclidean distance method, may be used to implement this
classier. Euclidean distance is given by (3):

 n

d(X, Y ) =  (xi yi )2 (3)
i=1
where X and Y represent the set of n samples to calculate the
Euclidean distance. In recent studies related to the recognition of
electrical machinery faults, some studies address the use of the KNN
method, as shown in [8,9].
2.2. RIPPER
The RIPPER method implements a learning algorithm based
on propositional rules, proposed by Cohen [30] as an optimized
version of IREP (Incremental Reduced Error Pruning) [31], which
incorporates the simplication of trees by error reduction using
the divide and conquer technique. This algorithm has a set of
rules which are tested one at a time. After a rule is found, all exam-
ples covered by it are deleted. This process is repeated until no
samples will be correctly classied or until the rule found by the
algorithm has an acceptable error. Although the fundamental prin-
ciples remain the same, the RIPPER algorithm improves on IREP in
several respects, and it can also accommodate multi-class problems
[30].
The RIPPER algorithm is organized into two parts: (i) the heuris-
tic method, which is conducted on the greedy principle, where
established rules are constructed using the information gain and
the optimization is then performed through the pruning process
to improve the rules accuracy; and (ii) the classication process,
which is the condition to apply the following rules to ensure the
effectiveness of the pruning as v . This process is repeated until
there is no improvement of conditionals v , as shown by (4) and
(5):
p + (N n)
v= (4)
P+N
where P and N is the total number of positive and negative samples,
respectively, p and n represent the positive and negative pruning
samples, respectively, observed at the time,
pn
v = (5)
p+n
(1) and where v denes the improved pruning measure.
The decision is made over the need to include the revised,
replacement, or original rule through the MDL (Minimal Descrip-
tion Length) heuristic. The measure of the algorithm gain is dened
by (6):
 p p0

p log 2 log2 (6)
p+n p0 + n0
given that R is a set of original rules and R is a set of rules after
adding candidates features, where p and n are the number of prPos
and prNeg samples covered by rule R , respectively, and where p0
and n0 are the numbers of prPos and prNeg samples covered by rule
R, respectively.
(2) In summary, the RIPPER method principle uses the model-
building process, which is based on rules and divided into two
parts: (i) based on the greedy principle, the heuristic method uses
gain information to form a collection of initial rules, from which
a collection of rules is pruned by an optimized process in order to
improve the accuracy of the collection of rules; and (ii) the sam-
ples classication is performed, whereby the rules must adhere to
categorization that meets the characteristics of a given rule, which
expresses the conjunction of several typical items. The output rip-
per comprises a categorized model of acquired training data.
252 R.H. Cunha Palcios et al. / Electric Power Systems Research 127 (2015) 249258

Some studies related to the use of the RIPPER method for fault
identication in rotating machinery appear in the research pub-
lished in [32,33].
2.3. C4.5 Decision Tree
The recognition methods based on decision trees patterns seek
to divide a training set until each subset obtained from this par-
titioning contains instances of a single class. The composition of a
tree is given by the following characteristics: (i) a root node formed
from all the data; (ii) a hierarchical set of internal nodes, called par-
titions, which are responsible for decision making using logical test,
which denes the next descendant node; and (iii) terminal nodes
called leaves. A binary decision tree divides each intermediate node
into two descendant nodes: the left node, when the logical test is
satised; and the right node, when the logical test is not satised.
The division of each internal node of the tree is dened by sta-
tistical criteria regarding the training data. In the regression tree
method the decision-making in each leaf contains a constant or
linear equation for a given set of data.
The ID3 algorithm [34] was one of the rst decision tree
algorithms-based inference systems and learning concepts, and
2.4. Naive Bayes
Bayesian networks use probability theory as a means to solve
problems, and are also used for reasoning and decision making with
uncertainty, based on the Bayes rule [36], which may be dened as
follows: (i) considering Ai attributes, where i = 1, 2, 3, . . ., n and
which have values j , where j = 1, 2, 3, . . ., m; (ii) assuming C as
class and E = (1 , 2 , . . ., m ) as instances of no classied tests. E
will be classied at class C with maximum posterior probability.
The Bayes rule for this classication is dened by (13).
P(C|E) = arg max P(C)P(E|C) (13)
C
Bayesian networks can represent uncertain attributes, however,
it has been proven that the ideal learning for a Bayesian Network is
through the non-deterministic polynomial method [37]. According
to [38], Bayesian networks provide a representation for machine
learning. The Naive Bayes classier is part of the Bayesian classiers
and a recognized technique. It assumes that all attributes of the
same class are independent. Based on this assumption, the Bayesian
rule is modied, as demonstrated in (14):

n

was followed a few years later by the C4.5 algorithm [35]. Various P(C|E) = arg max P(C) P(Ai |C) (14)
C
advantages of the C4.5 over ID3 allow users to: avoid data over- i=1
tting; determine how deeply to grow a decision tree; reduce the where Ai represents a vector of random variables that denotes the
pruning error of the tree; introduce the post-pruning rule; han- values of observed attributes.
dle continuous attributes; choose a measure of selection of proper Some works [6,7] associated with identifying faults in induction
attributes; handle training data with missing attribute values; and motors adopt the Naive Bayes method on the classication patterns
improve computational efciency. tests.
The C4.5 algorithm uses the normalized gain information to
select the most signicant feature to classify the training set.
2.5. Support Vector Machine
Assuming that the training set X has m classes (C1 , C2 , . . ., Cm ), the
entropy information of the node is calculated as follows in (7):
Support Vector Machines are learning machine systems trained

m
f (C , X)
 f (C , X)  with a mathematical algorithm optimization which implements
i i
I(X) = log2 (7) a derivative threshold of statistical learning theory. This learn-
|X| |X|
i=1 ing strategy was developed by Vladimir Vapnik [39] and can
achieve high performance in practical applications. The SVM train-
where |X| is the number of cases in X and f(Ci , X) is the number of ing involves optimization of a convex quadratic function. This
cases Ci pertaining to X. To select the features Z, it is assumed that optimization is a mathematical problem involving a few free
the training set is classied into n subsets (S1 , S2 , . . ., sn), allowing parameters required for user settings: there is no explicit depend-
the expected information entropy to be calculated according to (8): ence on determining the size of the input space of the problem.
Decision-making in classication problems can be done by

n
|Sj |
Ie (X, Z) = I(Sj ) (8) means of functions that use hyperplanes to divide the feature space
|X| into regions. The main SVM function is to map the input data into a
j=1
feature space by employing a nonlinear map. In feature space, the
where Sj is the number of cases. The gain G of the characteristic Z linear decision function is constructed.
can be calculated according to (9): According to [40], a set of training samples (xi , yi ) ; i = 1 N are
considered, where N is the total number of samples. The hyperplane
G(X, Z) = I(X) Ie (X, Z) (9) f(x) = 0, which separates provided data, can be obtained as a solution
to the optimization problem, as outlined in (15):
then, the information divided SI is given by (10). The gain rate (GT)
is then normalized according to (11) and the C4.5 method selects
1  N
features with maximum GT(X) of a node (Zn ) according to (12). min ||||2 + C i (15)
2
 |Sj |
n
|Sj |
i=1

SI(X, Z) = log2 (10)

|X| |X|
j=1 yi (T xi + b) 1 i
subject to
G(X, Z) i 0, i = 1, 2, . . ., N
GT (X, Z) = (11)
SI(X, Z)
where C is a constant that represents the threshold of error.
Zn = max(GT (X, Z)) (12) Rewriting the above optimization problem in terms of Lagrange
multipliers, as per (16):
In situations which require the identication of faults in induc-
tion motors, several studies in the literature apply methods based 
N
1
N

on decision trees, as noted in articles [16,17], which were published max W () = i yi yj i j (xi xj ) (16)
2
recently. i=1 i,j=1
 R.H. Cunha Palcios et al. / Electric Power Systems Research 127 (2015) 249258 253

Table 1

0 i C Characteristics of the induction motors used in the experiments.

subject to N Parameters Motor 1 Motor 2 Motor 3

i yi = 0, i = 1, 2, . . ., N
Manufacturer WEG WEG WEG
i=1 Power 1 hp 1 hp 2 hp
Poles 4 4 4
The Sequential Minimal Optimization (SMO) proposed by Platt Yield Standard High High

[41] is an algorithm to solve the Quadratic Programming (QP)

problem which arises during SVM training. This implementation
replaces all missing values globally, transforms attributes from
nominal to binary, and normalizes all attributes by default. The
multi-class problems are solved by using classication in pairs,
through the coupling-in-pairs method [42].
Problems of pattern classication in induction motors are often
solved by SVM, as can be seen in the works of [4,1015].
2.6. Articial Neural Networks
Articial Neural Networks are exible methods that can be
applied to various types of engineering problems. They are charac-
terized as computational models with particular properties, whose
main features are: (i) an inherent ability for parallel operation; (ii)
no need for a a priori knowledge of possible mathematical models
that describe the behavior of a particular application; (iii) inspired
systems in the human brain; and (iv) an ability to learn from expe-
rience.
This work applies the Multilayer Perceptron (MLP) neural net-
work, whose main characteristic is the presence of one or more
hidden layers, which are responsible for the extraction and stor-
age of knowledge associated with a specic application. Layered
networks are used in various applications, such as function approx-
imation, pattern recognition, optimization systems, and process
control [43].
The training or learning process of an Articial Neural Network
consists of adjusting the synaptic weights (weight matrix W) so
that the application of a set of inputs can produce a set of desired
outputs. The supervised training of the ANN used in this work is
based on the backpropagation learning algorithm [43]. The func-
tions square error (E(k)) and mean square error (EM ) are used as
performance criteria and to stop the training process.
The squared error provides the instantaneous value of the sum of
the squared errors of all neurons in the output layer of the network
in relation to the kth training pattern, as shown in (17):
1
N2
E(k) = (dj (k) y2j (k))
2
(17)
2 The reconstruction of stator faults consists of simulating an insu-
j=1
where dj (k) is the value of the desired output neuron j relative to
the kth input pattern.
The mean square error is obtained from the sum of the squared
errors for all input patterns used in the training set of the network,
according to (18):
1
p
3. Methodology
During its lifetime a three-phase induction motor may oper-
ate under varied loads and environmental conditions, resulting in
natural wear and tear of parts. However, early-stage checking for
faults in a motor provides better maintenance scheduling, lower
costs, and may prevent production faults which can otherwise lead
to losses [3]. In the context of this work, stator, rotor, and bearings
faults are analyzed by using pattern classication methods.
To approximate the tests performed in the laboratory with the
industrial environment, a laboratory structure must be created
which emulates practical situations. Table 1 presents the param-
eters of the motors used in this study.
The motors used in the experiments are manufactured by WEG:
motors 2 and 3 are high performance and motor 1 is standard.
These factors directly affect the dynamic response of the machines,
as each is built with different materials, such as conductors and
core stator windings. To better simulate the reality of the indus-
trial environment, bearings, rotor, and stator faults are created in
the laboratory.
The reproduction of bearings faults is based on the distributed
and localized conditions, which emulate the faulty operating con-
ditions at the industrial scale. Faults related to excessive bearing
wear, grooves on the outer and inner tracks, electrical discharge
and faults in the balls are introduced. For each fault, 36 tests with
1 hp motors (motors 1 and 2 in Table 1) were performed, for a total
of 144 tests. For the 2 hp motor (motor 3), a total of 120 tests were
conducted for faults in balls (60 tests) and outer race (60). Each
test represents a motor operation condition in accordance with its
power, based on the load variation and voltage unbalance, as shown
in Table 2.
To reproduce faults related to rotor broken bars, the bars which
comprise the rotor cage are sheared. This fault is articially created
by drilling into the rotor bars. This procedure results in a set of
rotors for testing. The faults are made for 1 to 4 bars by placing
the faults in sequence and also on the opposite side of the rotor
circumference. A total of 240 tests with faults in the rotor for the
1 hp motor are conducted.
lation failure on the stator winding. Once unprotected, the turning
may cause short-circuiting between the phases, phase and carcass,
or between the turns of a single coil. Accordingly, different levels of
short-circuits are tested for motor 1 (at 1%, 3%, 5% and 7%), motor
2 (at 1%, 3%, 5%, 10%, 15% and 20%), and motor 3 (at 1%, 3%, 5%, 10%,
15% and 20%). In total, 302 tests are conducted with the 2 hp motor
and 504 tests with the 1 hp motor.
To complete this set of samples, in order to characterize the

EM = E(k) (18) motors under normal operating conditions, the currents signals of
p motors 1, 2 and 3 are collected under normal conditions. Tests are
k=1

where the parameter p species the number of training patterns Table 2

Limits of voltages and load variation for data acquisition: 1 and 2 hp motors.
or the amount of input vectors. The objective of the learning pro-
cess, using the backpropagation algorithm, consists of adjusting the %Va %Vb %Vc 1 hp (Nm) 2 hp (Nm)
weights matrix of the network in order to minimize the function 100 100 100 0.56.0 0.59.0
EM . The process of adjusting the weights can be found in [43]. 90 100 100 0.56.0 0.59.0
The use of ANN to detect faults in electrical machines has been 100 110 100 0.56.0 0.59.0
the subject of considerable recent research [5,1822,44,23]. 90 100 110 0.56.0 0.59.0
254 R.H. Cunha Palcios et al. / Electric Power Systems Research 127 (2015) 249258

Fig. 1. Experimental test bench.

Fig. 2. Input treatment data diagram.

also performed considering the load variation and imbalance in the
motors power supply conditions, as shown in Table 2. A total of 36
tests are conducted with motor 1, 36 tests with motor 2, and 60 tests
bearing faults, rotor faults, and stator faults. Data are randomly pre-
with motor 3, for an overall total of 126 tests with the most common
sented to the intelligent classication process, whereas the data
types of faults found in industrial environments, considering load
preparation algorithm ensures that the samples used cover the
variation and unbalanced supply voltage.
entire operating domain observed in experimental tests. Experi-
mental data are divided into two groups: training and validation.
3.1. Generation and treatment of experimental data The tests performed in this study have different numbers of sam-
ples. Every test employs the cross-validation method, with 10-folds
Experimental data are obtained directly from the test motors by for training and validation.
using the test bench shown in Fig. 1. The tests with motors which
produce signals with bearings, rotor, and stator faults also include 4. Experimental results
the application of varying load and voltage unbalance conditions,
as shown in Table 2. The methodological aspects used in this work appear in Fig. 2,
In all tests the data is acquired by a National Instruments USB- which features a block diagram of the various steps involved in
6221 data acquisition board, and with Matlab software operating at the fault classication and their comparison with different intel-
a sampling rate of 6k samples/s. The stator current signal is suitable ligent methods. The data obtained through experimental tests
for the acquisition system, which employs a Hall sensor for signal are divided into training data and validation. The data used dur-
condition in the range of 10 V. The torque is measured by a Kistler ing the validation procedure are not the same as those used
torquimeter. in the training and organization of the input data structures
The speed is measured by a dc tachogenerator which is applied described in Section 3. Six intelligent pattern classication meth-
to the signal conditioning board 10 V. The acquisition rate is deter- ods were tested for every fault analyzed. In order to verify the
mined by the data acquisition card and the number of different
samples, considering the specic type of fault and motor power. In
Fig. 2 presents a block diagram relating to the input data processing
of the intelligent algorithms used in this work. Since the goal is to
identify the various faults in induction motors, the data set with
the currents signs of each phase (Ia , Ib and Ic ) is selected from the
motors in steady state. In order to obtain a standard sample for each
test, a half wave period of the phase currents in the time domain is
selected.
Each half-cycle of current wave extracts ten points that must
be submitted to the proposed methods by proceeding to the sig-
nal differentiation, as shown in Fig. 3. This method is based on
the proposal presented in [45]. The ten points which are extracted
result in the creation of input vectors with 30 points as inputs from
all intelligent methods evaluated in this work. Accordingly, only
the absolute amplitude values of the signal in the half-cycles are
considered. In Fig. 3 presents the methodology used in this work.
This data must be normalized by maximum current value, prior to
submitting samples for intelligent algorithms.
After data processing and depending on the fault type, the values
00 or 01, used for the classication of a single or multi-classication
fault, are given the values 00, 01, 10 or 11, which respectively rep-
resent signals from motors under the following conditions: normal, Fig. 3. Input data modeling: current segmentation.
 R.H. Cunha Palcios et al. / Electric Power Systems Research 127 (2015) 249258 255

Table 3
Training and testing results for the algorithms tested using data from a 1 hp motor (motors 1 and 2) for all faults with four classications something with 30 input current
data.

Type Stator faults

Classiers NB C4.5 SVM KNN RIP MLP

Inputs (Ia , Ib , Ic ) 30 30 30 30 30 30
Accuracy 93.7% 98.9% 98.3% 99.9% 99.2% 99.9%
Building time 0.03 s 0.56 s 0.54 s 0.0001 s 1.12 s 29 s
Testing time 0.04 s 0.01 s 0.02 s 0.56 s 0.01 s 0.09 s
Kappa statistic 0.78 0.95 0.92 0.99 0.96 0.99
Mean absolute error 0.061 0.012 0.016 0.001 0.009 0.001
Relative absolute error 28.5% 5.7% 7.6% 0.4% 4.1% 0.8%
Training and validation samples 576 576 576 576 576 576

generalization of the models, three separate three-phase induc-
tion motors were used in various test combinations, as shown in
Table 1.
The pattern classiers presented in Section 2 are employed in
this work through WEKA (Waikato Environment for Knowledge
Analysis) [46]. This study seeks to investigate intelligent methods of
pattern classication performance, based on the typical problems
of accuracy and computing time.
For the implementation of the strategies proposed in this work,
specic settings for training and validation of each classier method
are performed. In the Naive Bayes method, no decisive test setup
is performed. The basic settings of the SVM/SMO method deter-
mine the rounding error rate at 1.0E12 , the core function of the
SVM/SMO function being the Polynomial Kernel, while the toler-
ance parameter error receives a value of 0001.
For the ANN congurations, the learning rate is dened as 0.3,
the term momentum receives 0.2, and the maximum number of
epochs for network training is 500. Fifteen neurons (empirically
dened) are in the single hidden layer. For the hidden layer and
the output layer, the hyperbolic tangent activation function is
used.
The conguration of the KNN classication method is dened
as the number of neighbors used (one only), while the search algo-
rithm used is based on the calculation of the Euclidean distance
method.
For the RIPPER classier method, the parameter Seed used for
randomizing the data is dened as 1, and the use of the pruning
method is enabled. The parameter Folds determines the amount
of data used for pruning, and is dened as 3. The number of opti-
mization runs is set at 2. The minimum overall weight of the
instances in a rule is characterized by the value of 2.
The conguration of the C4.5 method used in this study does
not employ binary divisions in nominal attributes to construct the
trees. The seed is used to randomize the data with the value 1. The
pruning condence factor is set at 0.25. The minimum number of
instances per leaf tree receives a value of 2, and the amount of data
used for pruning error reduction is set as 3. This approach also uses
the pruning method.
4.1. Experimental results with training, and validations with 1 hp
motors
Table 3 presents the stator test results for the standard 1 hp
motor plus the motors with greater efciency (motors 1 and 2).
As specied in Table 1, this generated 504 experimental tests of
stator faults, to which were added 72 tests on the same machines
under normal working conditions i.e. with no faults. This totals 576
tests sampled for training and validation of the classiers methods.
The standard motor is tested with short-circuits of 1%, 3%, 5% and
7%. For the high performance motor, the test considers the follow-
ing short-circuit conditions: 1%, 3%, 5%, 10%, 15% and 20%. Data is
prepared for training and validation which comprises information
for the entire application domain. The construction models and the
implementation of validation tests are measured in seconds. The
KNN method presents the fastest performance for model-building
(0.0001 s). For validation, C4.5, along with RIPPER, take 0.1 s for
the execution. However, KNN and ANN/MLP achieve the highest
accuracy rates, at 99.9%.
Table 4 shows the rotor faults for the standard 1 hp motor, as
specied in Table 1. Data are prepared for training and validation
covering the entire application eld, with a separate training and
validation database. KNN is the pattern classication method which
posts the fastest time to build the model, with 0.01 s. For the vali-
dation tests, the C4.5, SVM/SMO, and RIPPER methods take 0.1 s to
process the operation. In terms of accuracy, ANN/MLP has a rate of
99.9%, KNN has 99.7%, while C4.5 is at 99.2%.
The test results of bearing faults for motor 1 are presented in
Table 5. Data are prepared for faults resulting from excessive wear
on the ball bearings and outer race. Based on the validation tests
of the sampled data, KNN is the fastest method for model build-
ing, posting 0.01 s. In the task of tests running, the C4.5, SVM/SMO,
RIPPER, and ANN/MLP methods take 0.01 s. For accuracy, the KNN
and ANN/MLP methods obtain the best rates, at 99.7% and 99.9%,
respectively.
The results of the fault combination of bearings, rotor, and sta-
tor appear in Table 6. This combinatorial procedure is designed to
check how robust the intelligent methods are, and measure their

Table 4
Training and testing results for the algorithms using real data from a 1 hp motor (motor 1), for all faults with four classications something with 30 input current data.

Type Rotor faults

Classiers NB C4.5 SVM KNN RIP MLP

Inputs (Ia , Ib , Ic ) 30 30 30 30 30 30
Accuracy 87.0% 99.2% 97.4% 99.7% 98.7% 99.9%
Building time 0.03 s 0.21 s 0.24 s 0.01 s 0.76 s 25.75 s
Testing time 0.02 s 0.01 s 0.01 s 0.48 s 0.01 s 0.02 s
Kappa statistic 0.50 0.96 0.88 0.98 0.94 0.99
Mean absolute error 0.132 0.008 0.025 0.002 0.015 0.002
Relative absolute error 58.8% 3.5% 11.1% 1.1% 6.8% 1.1%
Training and validation samples 312 312 312 312 312 312
256 R.H. Cunha Palcios et al. / Electric Power Systems Research 127 (2015) 249258

Table 5
Training and testing results for the algorithms tested using real data from a 1 hp motor (motor 1) for all faults with four classications something with 30 input current data.

Type Bearing faults

Classiers NB C4.5 SVM KNN RIP MLP

Inputs (Ia , Ib , Ic ) 30 30 30 30 30 30
Accuracy 75.9% 97.6% 85.1% 99.9% 95.4% 99.6%
Building time 0.02 s 0.37 s 0.52 s 0.01 s 1.02 s 21.34 s
Testing time 0.02 s 0.01 s 0.01 s 0.18 s 0.01 s 0.01 s
Kappa statistic 0.26 0.90 0 0.90 0.84 0.98
Mean absolute error 0.26 0.02 0.14 0.001 0.05 0.007
Relative absolute error 98.1% 10.9% 55.1% 0.5% 19% 2.7%
Training and validation samples 216 216 216 216 216 216

Table 6
Training and testing results for the algorithms using data from a 1 hp motor (motors 1 and 2) for all faults with four classications something with 30 input current data.

Type Stator, rotor and bearing faults

Classiers NB C4.5 SVM KNN RIP MLP

Inputs (Ia , Ib , Ic ) 30 30 30 30 30 30
Accuracy 51.7% 86.3% 74.1% 92.5% 84.4% 93.1%
Time to build 0.07 s 2.42 s 6.46 s 0.01 s 23.88 s 80.04 s
Time to test 0.12 s 0.02 s 0.01 s 2.37 s 0.01 s 0.04 s
Kappa statistic 0.33 0.81 0.60 0.94 0.77 0.92
Mean absolute error 0.24 0.07 0.27 0.03 0.10 0.04
Relative absolute error 71.3% 21.6% 79.9% 10.8% 29.2% 11.6%
Samples training and validation 960 960 960 960 960 960

multi-class capabilities through the inclusion of unbalanced volta-
ges and load variation for the 1 hp motor, using data from motors
1 and 2. Results obtained in these tests show that KNN is the
fastest at model building, with 0.01 s. For validation, the RIPPER and
SVM/SMO methods achieve the index of 0.1 s. For accuracy, while
the ANN/MLP and KNN methods achieve the best results, with 93.1%
and 92.5%, respectively.
An analysis of the general aspects for stator, rotor, and bearing
fault classication tests for the 1 hp motors shows results above
90% for the methods veried in this study.
4.2. Experimental results with training with 1 hp motors and
validation with data from a 2 hp motor
In order to investigate the generalization of the methods studied
in this work, the authors propose the systematic usage of the mod-
els, trained with faults and also under normal conditions for 1 hp
motors (motors 1 and 2), and to validate samples of 2 hp motors
under fault conditions, as both motors present different dynamics,
even with different load variation. In Table 7, the models are trained
with various 1 hp motor data, considering several fault types and
validation with samples of a 2 hp motor with faults in the ball bear-
ings. As a result, the C4.5, RIPPER, and ANN/MLP methods stand out
in relation to testing time, at 0.01 s. The SVM/SMO and NB methods
obtain 100% and 97.9% accuracy, respectively. The Kappa statistic
index of the SVM/SMO method returns a value of 1, revealing the
reliability of the test.
The validation results of bearing outer race fault for the 2 hp
motor, trained with data related to the 1 hp motor, are presented
in Table 8. Testing time for the NB, C4.5, SVM/SMO, RIPPER, and ANN
(MLP) methods ranging from 0.03 to 2.96 s. The SVM/SMO method
again achieved superior accuracy rate.
Based on the same models generated from the fault data in the
1 hp motors, Table 9 presents the validation data related to the sam-
ples gathered from the 2 hp motor with a stator fault (short-circuit
at 1%). Since the level of functional impairment of the motor is mini-
mal, the accuracy indicates that both pattern classication methods
present low results, although the KNN method obtains 92.3%, and
the shortest building time.
Table 10 present validation results concerning a set of samples
from the 2 hp motor (motor 3) with a stator fault (short-circuit at
10%). The KNN and ANN/MLP methods give the best results, with
accuracy of 93.4% and 93.5%, such as processing times of 1.14 and
0.05 s, respectively.
The validation of the models generated by methods trained with
samples gathered from various fault conditions articially gener-
ated in the laboratory for 1 hp motors, and validation of data from
some recurring faults collected from the 2 hp motor, demonstrate
how the pattern classication methods are robust, can generalize,
and help determine the most suitable solution in real applications.

Table 7
Training data for the 1 hp motor (motors 1 and 2) with rotor, stator, and bearing faults, plus a motor with no faults, as well as validation of data for the 2 hp motor (motor 3)
with a ball bearings fault. This set of input data considers 30 values of currents per phases A, B and C.

Type Ball bearing faults

Classiers NB C4.5 SVM KNN RIP MLP

Inputs (Ia , Ib , Ic ) 30 30 30 30 30 30
Accuracy 97.9% 56.7% 100% 66.1% 88.8% 67.1%
Building time 0.12 s 1.22 s 1.53 s 0.01 s 6.58 s 40.02 s
Testing time 0.12 s 0.01 s 0.02 s 0.64 s 0.01 s 0.01 s
Kappa statistic 0.26 0 1 0 0.43 0
Mean absolute error 0.01 0.21 0.25 0.16 0.06 0.15
Relative absolute error 2.7% 58.4% 66.9% 45.3% 16.9% 40.2%
Samples training 960 960 960 960 960 960
Samples validation 60 60 60 60 60 60
 R.H. Cunha Palcios et al. / Electric Power Systems Research 127 (2015) 249258 257

Table 8
Training data for rotor, stator, and bearing faults, plus 1 hp motors with no faults (motors 1 and 2), as well as validation of data for the 2 hp motor (motor 3) with a bearing
outer race fault. This set of input data considers 30 values of currents per phases A, B and C.

Type Bearing fault outer race

Classiers NB C4.5 SVM KNN RIP MLP

Inputs (Ia , Ib , Ic ) 30 30 30 30 30 30
Accuracy 97.8% 63.2% 98.8% 57.9% 53.6% 84.1%
Building time 0.16 s 2.67 s 6.29 s 0.02 s 27.82 s 76.5 s
Testing time 0.2 s 0.05 s 0.14 s 2.96 s 0.03 s 0.06 s
Kappa statistic 0.49 0 0.78 0 0 0.42
Mean absolute error 0.01 0.1 0.2 0.2 0.2 0.1
Relative absolute error 2.8% 49.5% 67.3% 56.5% 62.8% 35.8%
Samples training 960 960 960 960 960 960
Samples validation 60 60 60 60 60 60

Table 9
Training data for rotor, stator, and bearing faults, plus 1 hp motors with no faults (motors 1 and 2), as well as validation of data for the 2 hp motor (motor 3) with the stator
short-circuited at 1%. This set of input data considers 30 values of currents per phases A, B and C.

Type Stator short-circuit of 1%

Classiers NB C4.5 SVM KNN RIP MLP

Inputs (Ia , Ib , Ic ) 30 30 30 30 30 30
Accuracy 0.18% 67.6% 0.5% 92.3% 8.2% 32.1%
Building time 0.09 s 2.47 s 6.49 s 0.01 s 27.47 s 78.98 s
Testing time 0.11 s 0.04 s 0.03 s 1.59 s 0.03 s 0.04 s
Kappa statistic 0 0.28 0 0.72 0 0
Mean absolute error 0.4 0.1 0.4 0.03 0.4 0.3
Relative absolute error 155.5% 50.2% 139.3% 11.9% 139.3% 106.5%
Samples training 960 960 960 960 960 960
Samples validation 60 60 60 60 60 60

Table 10
Training data for rotor, stator, and bearing faults, plus 1 hp motors with no faults (motors 1 and 2), as well as validation of data for the 2 hp motor (motor 3) with the stator
short-circuited at 10%. This set of input data considers 30 values of currents per phases A, B and C.

Type Stator short-circuit of 10%

Classiers NB C4.5 SVM KNN RIP MLP

Inputs (Ia , Ib , Ic ) 30 30 30 30 30 30
Accuracy 69.4% 78.9% 77.9% 93.4% 73.9% 93.5%
Building time 0.1 s 2.48 s 6.65 s 0.01 s 26.16 s 93.11 s
Testing time 0.05 s 0.01 s 0.01 s 1.14 s 0.01 s 0.05 s
Kappa statistic 0 0.35 0.36 0.81 0.13 0.73
Mean absolute error 0.15 0.1 0.28 0.03 0.15 0.03
Relative absolute error 48.6% 32.8% 89% 10.3% 48.5% 10.1%
Samples training 960 960 960 960 960 960
Samples validation 60 60 60 60 60 60

In this comparative study between intelligent methods to identify
short-circuit faults in the stator of broken bars in the rotor and bear-
ings of 1 hp motors, the KNN and ANN/MLP methods deliver the best
results and act as multi-classiers of these faults. However, when
the samples are trained with multiple faults in the 1 hp motors and
presented to the bearings fault samples in the 2 hp motor, the SVM
method shows the best results, with accuracy above 98%. In this
same situation, for samples with faults related to short-circuits in
the stator (1% and 10%) of the 2 hp motor, the KNN method has the
best classication rates, reaching results above 92%.
5. Conclusion
The identication of the correct fault pattern in induction
motors, considering load variations and unbalanced voltages
through the use of single current data, seeks to provide: (a) safety,
via decision-making for the process; (b) reduced maintenance
costs; (c) increased availability in the case of a real application;
and (d) the possibility of an embedded hardware solution. This
paper proposes a comparative study between traditional meth-
ods for intelligent pattern recognition in three-phase induction
motors with sinusoidal supply. Analysis of the results obtained by
the methods evaluated shows that these algorithms have proved to
be applicable for correct classication of faults in bearings, rotors,
and stators by using current data of induction motors in the time
domain, as their accuracy exceeds 90%.
Specically, for the classication of 1 hp motors with stator
faults, the KNN and ANN/MLP methods showed better results,
reaching 100% accuracy. As for the tests with data from 1 hp motor
with broken rotor bars faults, both ANN/MLP and KNN methods
also reaches accuracy rates over to 99.7%. As for the classi-
cation of bearings faults, the KNN method reaches an accuracy
of 99.9%. Also, ANN/MLP and C4.5 methods present promising
results. For the tests considering faults multi-classication, the
ANN/MLP and KNN methods exceeds 92.5% accuracy. For the
robustness tests with defective bearings, considering training with
data from 1 hp motor and validation with samples from 2 hp motor,
the SVM/SMO method showed the best performance in terms of
accuracy and processing time. Also, considering results obtain-
ing with data validated with 2 hp motor, with stator faults, the
KNN method shows overall accuracy of 92.3% and 93.4%, respec-
tively in both robustness tests performed. Evaluated classiers
methods can be embedded in hardware along with pre-processing
method.
258 R.H. Cunha Palcios et al. / Electric Power Systems Research 127 (2015) 249258
The experimental validation process was carried out on differ-
ent motors connected to the power supply under conditions of load
variation and voltage unbalance, which was designed to emulate
real industrial environment situations. However, there are situa-
tions where a hybrid system may be necessary to generalize and
cover the entire application eld. A correct and early diagnosis of
incipient faults, using the strategies tested in this study, can mini-
mize process damage. Accordingly, it is possible to increase process
availability and reduce maintenance costs.
Acknowledgements
The authors gratefully acknowledge the contributions of CNPq
(process #552269/2011-5), FAPESP (process #2011/17610-0),
Araucria Foundation/CAPES (CP 13/2014), University of So Paulo
and Federal Technological University of Paran for their nancial
support towards the development of this research.
References
[1] M. Hajian, J. Soltani, G. Markadeh, S. Hosseinnia, Adaptive nonlinear direct
torque control of sensorless IM drives with efciency optimization, IEEE Trans.
Ind. Electron. 57 (3) (2010) 975985.
[2] T.H. dos Santos, A. Goedtel, S.A.O. da Silva, M. Suetake, Scalar control of an
induction motor using a neural sensorless technique, Electr. Power Syst. Res.
108 (0) (2014) 322330.
[3] A. Bellini, F. Filippetti, C. Tassoni, G.A. Capolino, Advances in diagnostic tech-
niques for induction machines, IEEE Trans. Ind. Electron. 55 (12) (2008)
41094126.
[4] P. Konar, P. Chattopadhyay, Bearing fault detection of induction motor using
wavelet and Support Vector Machines (SVMs), Appl. Soft Comput. 11 (6) (2011)
42034211.
[5] H. Ertunc, H. Ocak, C. Aliustaoglu, ANN- and ANFIS-based multi-staged decision
algorithm for the detection and diagnosis of bearing faults, Neural Comput.
Appl. 22 (1) (2013) 435446.
[6] M. Hajiaghajani, H. Toliyat, I.M.S. Panahi, Advanced fault diagnosis of a DC
motor, IEEE Trans. Energy Convers. 19 (1) (2004) 6065.
[7] J. Wang, S. Liu, R.X. Gao, R. Yan, Current envelope analysis for defect iden-
tication and diagnosis in induction motors, J. Manuf. Syst. 31 (4) (2012)
380387.
[8] O. Ondel, E. Boutleux, G. Clerc, A method to detect broken bars in induction
machine using pattern recognition techniques, IEEE Trans. Ind. Appl. 42 (4)
(2006) 916923.
[9] B. Ebrahimi, M. Javan Roshtkhari, J. Faiz, S. Khatami, Advanced eccen-
tricity fault recognition in permanent magnet synchronous motors using
stator current signature analysis, IEEE Trans. Ind. Electron. 61 (4) (2014)
20412052.
[10] M. Barzegaran, A. Mazloomzadeh, O. Mohammed, Fault diagnosis of the asyn-
chronous machines through magnetic signature analysis using nite-element
method and neural networks, IEEE Trans. Energy Convers. 28 (4) (2013)
10641071.
[11] X. Li, A. Zheng, X. Zhang, C. Li, L. Zhang, Rolling element bearing fault detec-
tion using support vector machine with improved ant colony optimization,
Measurement 46 (8) (2013) 27262734.
[12] E. Esfahani, S. Wang, V. Sundararajan, Multisensor wireless system for eccen-
tricity and bearing fault detection in induction motors, IEEE/ASME Trans.
Mechatron. 19 (3) (2014) 818826.
[13] A. Moosavian, H. Ahmadi, B. Sakhaei, R. Labba, Support vector machine and
K-nearest neighbour for unbalanced fault detection, J. Qual. Maint. Eng. 20 (1)
(2014) 6575.
[14] S. Das, P. Purkait, C. Koley, S. Chakravorti, Performance of a load-immune classi-
er for robust identication of minor faults in induction motor stator winding,
IEEE Trans. Dielectr. Electr. Insul. 21 (1) (2014) 3344.
[15] J. Seshadrinath, B. Singh, B. Panigrahi, Investigation of vibration signatures for
multiple fault diagnosis in variable frequency drives using complex wavelets,
IEEE Trans. Power Electron. 29 (2) (2014) 936945.
[16] H.-W. Peng, P.-J. Chiang, Control of mechatronics systems: ball bearing fault
diagnosis using machine learning techniques, in: Control Conference (ASCC),
2011 8th Asian, 2011, pp. 175180.
[17] I. Aydin, M. Karakose, E. Akin, An approach for automated fault diagnosis based
on a fuzzy decision tree and boundary analysis of a reconstructed phase space,
ISA Trans. 53 (2) (2014) 220229.
[18] J. Bossio, C. Angelo, G. Bossio, Self-organizing map approach for classication
of mechanical and rotor faults on induction motors, Neural Comput. Appl. 23
(1) (2013) 4151.
[19] M. Seera, C.P. Lim, S. Nahavandi, C.K. Loo, Condition monitoring of induction
motors: a review and an application of an ensemble of hybrid intelligent mod-
els, Expert Syst. Appl. 41 (10) (2014) 48914903.
[20] E. Germen, M. Basaran, M. Fidan, Sound based induction motor fault diagnosis
using Kohonen self-organizing map, Mech. Syst. Signal Process. 46 (1) (2014)
4558.
[21] M. Seera, C. Lim, D. Ishak, H. Singh, Application of the fuzzy min-max neural
network to fault detection and diagnosis of induction motors, Neural Comput.
Appl. 23 (1) (2013) 191200.
[22] M. Seera, C.P. Lim, D. Ishak, H. Singh, Ofine and online fault detection and
diagnosis of induction motors using a hybrid soft computing model, Appl. Soft
Comput. 13 (12) (2013) 44934507.
[23] J. Zarei, M.A. Tajeddini, H.R. Karimi, Vibration analysis for bearing fault detec-
tion and classication using an intelligent lter, Mechatronics 24 (2) (2014)
151157.
[24] S. Moosavi, A. Djerdir, Y. Ait-Amirat, D. Khaburi, ANN based fault diagnosis
of permanent magnet synchronous motor under stator winding shorted turn,
Electr. Power Syst. Res. 125 (0) (2015) 6782.
[25] O. Duque-Perez, L.-A. Garcia-Escudero, D. Morinigo-Sotelo, P.-E. Gardel, M.
Perez-Alonso, Analysis of fault signatures for the diagnosis of induction motors
fed by voltage source inverters using ANOVA and additive models, Electr. Power
Syst. Res. 121 (0) (2015) 113.
[26] A.G. Garcia-Ramirez, L.A. Morales-Hernandez, R.A. Osornio-Rios, J.P. Benitez-
Rangel, A. Garcia-Perez, R. de Jesus Romero-Troncoso, Fault detection in
induction motors and the impact on the kinematic chain through thermo-
graphic analysis, Electr. Power Syst. Res. 114 (0) (2014) 19.
[27] D. Aha, D. Kibler, Instance-based learning algorithms, Mach. Learn. 6 (1991)
3766.
[28] T. Cover, Estimation by the nearest neighbor rule, IEEE Trans. Inf. Theory 14 (1)
(1968) 5055.
[29] C.M. Bishop, Pattern Recognition and Machine Learning (Information Science
and Statistics), Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2006.
[30] W.W. Cohen, Fast effective rule induction, in: Twelfth International Conference
on Machine Learning, Morgan Kaufmann, 1995, pp. 115123.
[31] J. Frnkranz, G. Widmer, Incremental reduced error pruning, in: W.W. Cohen,
H. Hirsh (Eds.), Proceedings of the 11th International Conference on Machine
Learning (ML-94), Morgan Kaufmann, New Brunswick, NJ, 1994, pp. 7077.
[32] S. Jianghong, X. Xiaoli, Large rotating machinery fault diagnosis and knowledge
rules acquiring based on improved RIPPER, in: Second International Conference
on Intelligent Computation Technology and Automation, 2009. ICICTA 09, vol.
2, 2009, pp. 549552.
[33] W. Hongjun, X. xiaoli, H. Qiushi, Study of intelligent fault diagnosis system
based on data mining technology, in: 2010 International Forum on Information
Technology and Applications (IFITA), vol. 2, 2010, pp. 329332.
[34] J. Quinlan, Induction of decision trees, Mach. Learn. (1) (1986) 81106.
[35] R. Quinlan, C4.5: Programs for Machine Learning, Morgan Kaufmann Publishers,
San Mateo, CA, 1993.
[36] F. Jensen, Introduction to Bayesian Networks, Hugin Expert, Denmark, 1993.
[37] L. Jiang, Y. Guo, Learning lazy naive Bayesian classiers for ranking, in: 17th
IEEE International Conference on Tools with Articial Intelligence, 2005, pp.
411416.
[38] G.H. John, P. Langley, Estimating continuous distributions in Bayesian classi-
ers, in: Eleventh Conference on Uncertainty in Articial Intelligence, Morgan
Kaufmann, San Mateo, 1995, pp. 338345.
[39] V.N. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag New
York, Inc., New York, NY, USA, 1995.
[40] P. Kankar, S.C. Sharma, S. Harsha, Fault diagnosis of ball bearings using contin-
uous wavelet transform, Appl. Soft Comput. 11 (2) (2011) 23002312.
[41] J. Platt, Fast training of support vector machines using sequential minimal
optimization, in: B. Schoelkopf, C. Burges, A. Smola (Eds.), Advances in Kernel
Methods Support Vector Learning, MIT Press, 1998.
[42] T. Hastie, R. Tibshirani, Classication by pairwise coupling, in: Proceedings of
the 1997 Conference on Advances in Neural Information Processing Systems
10, MIT Press, Cambridge, MA, USA, 1998, pp. 507513.
[43] S. Haykin, Neural Networks: A Comprehensive Foundation, 2nd ed., Prentice
Hall PTR, Upper Saddle River, NJ, USA, 1998.
[44] V.T. Tran, F. AlThobiani, A. Ball, B.-K. Choi, An application to transient current
signal based induction motor fault diagnosis of Fourier-Bessel expansion and
simplied fuzzy artmap, Expert Syst. Appl. 40 (13) (2013) 53725384.
[45] C.F. do Nascimento, A.A. de Oliveira Jr., A. Goedtel, P.J.A. Serni, Harmonic iden-
tication using parallel neural networks in single-phase systems, Appl. Soft
Comput. 11 (2) (2011) 21782185.
[46] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, I.H. Witten, The
WEKA data mining software: an update, SIGKDD Explor. Newslett. 11 (1) (2009)
1018.