Electric Power Systems Research 143 (2017) 347–356

Contents lists available at ScienceDirect

Electric Power Systems Research

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

Stator fault analysis of three-phase induction motors using

information measures and artificial neural networks
Gustavo Henrique Bazan a,b , Paulo Rogério Scalassara a,∗ , Wagner Endo a ,
Alessandro Goedtel a , Wagner Fontes Godoy a , Rodrigo Henrique Cunha Palácios a
a
Federal University of Technology – Paraná, Department of Electrical Engineering, Av. Alberto Carazzai, 1640, Centro, 86300-000 Cornélio Procópio, PR,
Brazil
b
Federal Institute of Paraná, Department of Electromechanical, Av. José Felipe Tequinha, 1400, Jardim das Nações, 87703-536 Paranavaí, PR, Brazil

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

Article history: The three-phase induction motors are considered one of the most important elements of the industrial
Received 21 May 2016 process. However, in this environment, these machines are subject to electrical and mechanical faults,
Received in revised form which may cause significant financial losses. Thus, the purpose of this paper is to present a pattern
23 September 2016
recognition method for the detection of stator windings short circuits based on measures of mutual
Accepted 29 September 2016
information between the phase current signals. In order to validate the proposed patterns, feature vectors
Available online 26 October 2016
obtained from normal and faulty motors are applied to two topologies of artificial neural networks.
The classification results presented accuracies over 93% even when the motors were subject to several
Keywords:
Three-phase induction motor
conditions of load torque and power supply voltage unbalance.
Artificial neural networks © 2016 Elsevier B.V. All rights reserved.
Information measures
Stator fault

1. Introduction
The three-phase induction motor (TIM) is one of the most com-
mon machines used to transform electrical energy into mechanical
energy in industrial processes. This is mainly due to the low acqui-
sition and maintenance costs, reliability, robustness to aggressive
environments, as well as being easily adapted to several load con-
ditions [1–3]. Despite these favorable characteristics, TIMs can
present electrical or mechanical faults that may be caused by
prolonged activity times, harsh operating conditions, voltage or
current unbalance, among other factors. Even incipient faults can
affect the motor performance, which may result in financial losses
[4,5]. Therefore, the search for adequate diagnostic methods is a
recurrent research topic in the past few years, aiming to detect the
faults in an early stage and avoid unwanted production stops [6].
Among the electrical failures, up to 36% of the cases are due to
stator winding faults [7]. Specifically regarding this kind of fault, in
[8], Discrete Wavelet Transform (DWT) is used to obtain patterns
from high frequency components that are classified using Artifi-
cial Neural Networks (ANN). Also, in [9], statistical measures of the
stator current, such as mean, variance, skewness, and kurtosis, are
used as inputs to a multilayer perceptron (MLP) ANN. Similarly, in
∗ Corresponding author.
E-mail address: prscalassara@utfpr.edu.br (P.R. Scalassara).
[10], these measures are also used for rotor eccentricity faults along
with stator winding inter-turn short circuit. In [11], the authors
showed that Park vector modules combined with Hilbert Trans-
form contain relevant information for detecting short circuit faults
in TIM stator windings. In addition, in [12], wavelet transform was
used for feature extraction of current signals for the identification of
fault severity, even with the machine subject to different operating
conditions.
A more recent approach of pattern recognition for motor fault
detection is the use of information theory (IT) measures such as
entropy. One example is presented in [13], in which TIM stator
currents are analyzed using DWT and entropy estimations for fault
diagnosis. Also, in [14], multiscale entropy estimations from cur-
rent and vibration signals are used as patterns for stator fault
detection. Similar entropy measures obtained from the stator cur-
rent are presented in [15], where they are applied to an ANN for
fault prediction. Entropy is also employed in [16], along with sup-
port vector machines (SVMs) and ANNs, for the detection of TIM
faulty bearings and broken rotor bars. Finally, in [17], bearings faults
are also the object of analysis, but using relative entropy and DWT
of the TIM current signals.
Another IT measure that has recently being applied in fault
diagnosis is mutual information (MI). It is a similarity measure
used in feature extraction methods that has applications in many
areas, such as in speech intelligibility prediction [18]. For motor
fault detection, in [19], MI measurements of vibration signals

http://dx.doi.org/10.1016/j.epsr.2016.09.031
0378-7796/© 2016 Elsevier B.V. All rights reserved.
348 G.H. Bazan et al. / Electric Power Systems Research 143 (2017) 347–356
in time-frequency domain are used to evaluate the condition of
helicopter bearings and shafts. Similarly, in [20], MI is used to
assess the correlation of diesel engine signals in order to obtain
the most informative ones for fault diagnosis. One final example
is [21], where the authors employed MI in the feature extraction
step to select the most relevant and less redundant measurements
from TIM current signals for eccentricity fault identification.
Considering this recent application of IT measures, the aim of
this work is to present an alternative approach to the diagnosis
of TIM stator windings short circuit. This method is based on the
analysis of MI values obtained from delayed line currents of two
TIM phases. Using this tool, it is possible to evaluate the common
information of the signals, as well as verify the changes of MI values
due to different TIM operating conditions.
The efficiency of the proposed feature extraction method is
assessed using two ANN topologies, multilayer perceptron (MLP)
and radial basis function (RBF), which are proved to be adequate
for motor fault detection as shown in [9,10,22,23]. For that, we
performed 814 experimental tests with two TIMs of 0.74 kW and
1.48 kW in steady state and under several operating conditions
that are common in an industrial environment, such as varia-
tions of load torque and power supply voltage unbalance. It is
worth pointing out that the technique has a characteristic of
being low cost, therefore the current signals are acquired nonin-
vasively, using Hall-effect sensors, and no frequency domain tools
are employed.
This paper is structured as follows. Section 2 presents aspects
related with TIM stator winding faults. In Section 3, the main
concepts of MI are addressed. Section 4 presents aspects of the pro-
posed methodology and experimental setup. The results are shown
in Section 5, and finally, the conclusions in Section 6.
2. Stator faults
In this section, aspects related to TIM stator winding faults are
presented, which are one of the main problems found in elec-
tric motors [4,7,24,25]. Also shown are brief descriptions of other
researches addressing the identification of these faults at an early
stage, aiming to avoid unscheduled downtime and reduce mainte-
nance costs.
The most common defects associated with TIM stator are phase
to ground, phase to phase, and coil to coil short circuits [26]. When
there is a short circuit involving just a few coil turns, the stator
insulation starts to deteriorate; this could rapidly develop into a
more serious fault and affect the motor operation [4,7,27].
In order to monitor the TIM operating conditions and identify
these faults, one must use a proper signal processing technique
for the extraction and selection of relevant features of the signals.
The motor current signature analysis is a commonly used method
because it is simple, noninvasive, and can be automated [28]. This
approach is focused on the current spectral analyses and has been
widely used for TIM fault diagnosis [29–31].
In the literature of fault detection, there are other examples
of signal processing techniques based on spectral decomposi-
tion using short-time Fourier transform and Wavelet transform
[8,12,32–34]. In other papers, applications using these transforms
with the Hilbert transform are presented [11,35]. Other authors
propose methods for TIM fault diagnosis based on the analysis of
vibrations [5,36], magnetic signals [5,37,38], active and reactive
power [7], temperature [39], among other variables.
One disadvantage of most of these methods is the requirement
for high-priced sensors, which increase the cost of the diagnostic
system. In order to avoid that, some authors are looking for other
noninvasive methods, in which measurements are provided by low
cost sensors, and using time domain strategies. In [22], an alterna-
tive approach to the traditional intelligent systems methods was
considered to diagnose stator faults. The TIMs were driven by three
different types of frequency inverters and the currents amplitudes,
in time domain, were used to classify stator fault severity. Also, in
[23], the authors proposed an evaluation of several pattern classifi-
cation methods, namely naive Bayes, k-nearest neighbors, support
vector machines, ANNs, and decision trees, for multi-fault classi-
fication (rotor, stator, and bearings faults) using the TIM currents
amplitudes in time domain.
It is noteworthy that intelligent systems methods are widely
used for fault identification of electrical machines, mainly due to
their ability to classify features without requiring complex math-
ematical models. The most usual methods are ANNs [22,40–42],
SVMs [31,43], fuzzy logic [44,45], decision trees [46], and hybrid
systems [30,47]. Thus, based on this, the features obtained with MI
measures of the TIM current signals are classified using ANNs in
this paper.
3. Mutual information
MI is a measure of how much information a random variable
(RV) has over another one. In other words, it is the uncertainty
reduction of a RV brought by information of the other [48]. This
quantity describes the similarity between time series collected
simultaneously from a system under study [49,50].
Thus, by estimating the MI between the TIM current signals, it
is possible to quantify the association between them, allowing an
efficient monitoring of the machine operating conditions. Consid-
ering two RV X and Y, the MI can be estimated using (1), where x
and y are realizations of the RVs, p(x) and p(y) are the probability
density functions (PDFs) of X and Y, respectively, and p(x, y) is the
joint PDF of the them.
 
p(x, y)
I(X, Y ) = p(x, y)log2 dxdy (1)
Y X
p(x)p(y)
In practice, the PDFs are approximated by the signals normalized
histograms and the expected values are estimated by time averages.
Since the winding faults tend to present asymmetrical effects
in the TIM line current signals and MI is a similarity measure, this
quantity should reflect such variations. In this work, we analyzed
the delayed MI of two TIM phase current signals. This quantity is
dependent on a time shift of  samples that is applied to one of the
signals before estimation using (1). Delayed MI has been used in
dynamic structural analysis of complex systems, due to its capacity
of quantifying the dependence between RVs, seeking the similarity
between them relative to the time delay  [51–53].
In order to obtain a reasonable analysis of the TIM current vari-
ations caused by the stator faults, several MI values are estimated
with shifted versions of the phase B current signal. As an example,
Fig. 1(a) presents a part of a three-phase current acquired during 3 s
on a normal TIM. The MI values obtained using the phase A current
signal and 150 delayed versions of the phase B current signal are
shown in Fig. 1(b).
In the next section, a detailed description of the methodology
employed in this work is presented.
4. Proposed approach
As mentioned in the previous sections, the purpose of this work
is to identify stator short circuit faults by using MI values of the
TIM current signals. Fig. 2 presents a diagram of the proposed fault
detection method explained in this section. In order to validate
this approach, experimental tests were performed. Next, details are
presented about the current signals acquisition, and feature analy-
sis and classification using ANN.
 G.H. Bazan et al. / Electric Power Systems Research 143 (2017) 347–356 349

Fig. 1. (a) Part of a three-phase current acquired on a normal TIM showing phases A (solid line), B (dashed line), and C (dotted line). (b) Estimated MI values obtained using
phase A and 150 delayed versions of phase B (using the whole 3 s signals).

4.1. Signal acquisition Table 1

Combinations of experiments using the two TIMs and variations of power supply
voltage unbalance and load torque for each short circuit fault level.
The current signals were acquired from TIMs with normal and
faulty operating conditions. The faults were obtained using short %Va %Vb %Vc 0.74 kW TIM 1.48 kW TIM
circuit levels of 1%, 3%, 5%, and 10% of the stator windings. These Torque (Nm) # Exp. Torque (Nm) # Exp.
faults consist in a controlled emulation of an insulation breakage,
100 100 100 0.7–4.9 22 0.9–9.3 22
thus, it is possible to artificially create short circuit faults between 98 100 100 0.7–4.9 22 0.9–9.3 22
turns of the same coil, between coils of the same phase, and also 96 100 100 0.7–4.9 22 0.9–9.3 22
between coils of different phases. The two TIMs used in this work 100 102 100 0.7–4.9 22 –
had all their stator coils rewinded in order to allow the creation of 100 104 100 0.7–4.9 22 –

the short circuits. Both machines have stators with six coils, 306
turns per phase for the 0.74 kW TIM and 780 turns for the 1.48 kW
TIM. For this latter, the minimum possible short circuit level was
3%.
Fig. 3 presents the experimental workbench used for signal
acquisition. This system was designed to monitor TIM electrical
and mechanical characteristics, namely voltage, current, vibration,
torque and speed. In addition, it allows emulation of several oper-
ating conditions such as variations of load torque and power supply
unbalance. More details are presented in [17,22,23].
The database consists of 814 experiments, 550 performed on the
0.74 kW TIM and 264 on the 1.48 kW one, both operating in steady
state under various conditions commonly found in the indus-
trial production process as mentioned before. The machines were
connected directly to the power network and tested under condi-
tions ranging from 20% to 120% of their nominal load (0.7–4.9 Nm,
in intervals of 0.2 Nm, for the 0.74 kW TIM, and 0.9–9.3 Nm, in inter-
vals of 0.4 Nm, for the 1.48 kW TIM). The power supply voltages for
Total 110 66
the 0.74 kW TIM were varied as 0%, −2%, and −4% of phase A, and
0%, +2%, and +4% of phase B. For the 1.48 kW TIM, it was 0%, −2%,
and −4% voltage variations of phase A.
The combinations of the load torque and power supply volt-
age unbalance (including the balanced condition) resulted, for each
short circuit fault level, in 110 experiments for the 0.74 kW TIM
and 66 for the 1.48 kW TIM, as shown in Table 1. The sampling fre-
quency was 15.5 kHz, amplitude quantization of 16 bits, and 5 s of
acquisition time.
4.2. Feature analysis
The feature extraction method consists of estimating the MI of
the TIM phase A current signal and delayed versions of the phase B
signal, considering delays from 0 to 150 samples. From the analysis
350 G.H. Bazan et al. / Electric Power Systems Research 143 (2017) 347–356

Fig. 2. Diagram of the proposed fault detection method.

of the MI versus delay curves of the acquired signals for all the
operating conditions described, one can notice that each of these
conditions presents a characteristic signature. In order to illustrate
this, Figs. 4 and 5 show the curves obtained for the 0.74 kW TIM,
with no short circuit, but subject to variations of load torque and
power supply voltage unbalance, respectively.
From the analysis of Fig. 4, one can observe that an increase in
load level leads to a slight shift right and peak increase of the MI
characteristic curve in relation to the reference curve, that is, the
one with minimum torque and no voltage unbalance. Regarding
Fig. 5, it is noticeable that variations of phase A voltage also results
in a slight shift right but with peak decrease compared with the
reference curve. But, the same variations of phase B voltage causes
a shift left and also a peak decrease of the curves.
The same analysis can be extended to TIMs with short circuits in
the stator windings. Figs. 6 and 7 illustrate these situations for the
0.74 kW TIM operating with 10% of stator short circuit of phase A,
variations of load torque and power supply voltage unbalance are
respectively considered in each figure. The behavior of the curves
for the faulty TIM is similar to what is observed for the normal
operating condition. Load torque increase caused a larger shift right
of the curves, as observed in Fig. 6. Also, considering Fig. 7, the same

Fig. 3. Experimental workbench used for TIM signal acquisition.

 G.H. Bazan et al. / Electric Power Systems Research 143 (2017) 347–356 351

Fig. 4. MI versus delay curves obtained for the 0.74 kW TIM with no short circuit, balanced power supply voltage, and variations of load torque.

Fig. 5. MI versus delay curves obtained for the 0.74 kW TIM with no short circuit, 0.7 Nm of load torque, and variations of power supply voltage.

Fig. 6. MI versus delay curves obtained for the 0.74 kW TIM with 10% of stator short circuit of phase A, balanced power supply voltage, and variations of load torque.
352 G.H. Bazan et al. / Electric Power Systems Research 143 (2017) 347–356

Fig. 7. MI versus delay curves obtained for the 0.74 kW TIM with 10% of stator short circuit of phase A, 0.7 Nm of load torque, and variations of power supply voltage.

Table 2
Details of the ANN topologies.

Network architecture MLP1 MLP2 MLP3 RBF1 RBF2 RBF3

Number of layers 2 2 2 2 2 2
Neurons on the hidden layer 10 20 50 3 5 7
Neurons on the output layer 2 2 2 2 2 2
Training algorithm BP BP BP SO SO SO
Activation function (hidden layer) Sig Sig Sig G G G
Activation function (output layer) Linear Linear Linear Linear Linear Linear
Learning rate 0.1 0.1 0.1 0.01 0.01 0.01
Momentum 0.9 0.9 0.9 – – –
Epochs 526 688 856 – – –

BP: backpropagation, SO: self-organization, sig: sigmoid, G: Gaussian.

kind of curve shifts is observed when there are voltage variations,
but with an increase of MI values for unbalance of phase B voltage.
Fig. 8 presents a comparison of the MI curves obtained for the
0.74 kW TIM under normal and faulty operating conditions with
0.7 Nm load torque and balanced power supply voltage. The sta-
tor winding short circuit levels are 1%, 3%, 5%, and 10%. The faults
caused a shift left of the MI curves proportional to the short circuit
levels, a decrease of MI values, and an enlargement of the curve.
The same behavior shown in Figs. 4–8 for the 0.74 kW TIM was
observed for the 1.48 kW TIM, but omitted here for simplicity.
4.3. Feature classification using ANN
For each of the 814 experimental tests, a feature vector of 150
MI values was created and used as the input to two ANN topologies:
MLP and RBF networks, which are briefly described next.
The ANNs are mathematical models based on the nervous sys-
tems of living creatures. They are defined as a set of processing
units, the artificial neurons, and have the capacity for acquiring
and maintaining knowledge from information. According to [54],
the essential ANN characteristics are the adaptation through expe-
rience, learning capability, generalization ability, data organization,
fault tolerance, and distributed storage.
Several ANN topologies are built using one or many neuron
layers, examples are the MLP and RBF networks that are used in
this work. The former consists of at least one hidden layer and one
output layer. This topology uses the backpropagation algorithm,
which has two stages: feedforward and backward. The first is a
supervised stage where only the current synaptical weights and
thresholds are considered. These values are adjusted in the second
stage. Differently, the RBF topology consists of an input layer, only
one middle layer, and an output layer. Besides that, the neurons
of the middle layer have Gaussian activation functions, instead of
the logistical or hyperbolic tangent functions used for the MLP.
The RBF training has two stages, the first is associated with the
adjustment of the synaptical weights of the middle layer using
an unsupervised learning method that depends only on the input
data. The second stage is related to the adjustment of the output
layer weights using the generalized delta learning technique [54].
The ANN training and testing were performed in the software
WEKA [55], using cross-validation with the samples randomly
divided in ten groups. Using this procedure, the network is evalu-
ated ten times, in each time, one group is selected as validation and
the rest is used for training. At each iteration, the misclassification
rate of the test set is calculated, and, at the end, the performance is
estimated by the classifier average errors.
The details of the ANN topologies are shown in Table 2. In
the next section, the results of the ANN validation stage are pre-
sented.
5. Results and discussion
The proposed ANN architectures were trained according to the
details discussed in the previous section. Tables 3 and 4 present
the classification errors and accuracies achieved by the MLP and
RBF networks, respectively for the feature vectors obtained from
the current signals of the 0.74 kW TIM. The kappa index is also
shown, which is a measure of consistency of the network results.
According to [22], kappa values higher than 0.8 represent the ideal
concordance of the obtained success rates.
 G.H. Bazan et al. / Electric Power Systems Research 143 (2017) 347–356 353

Fig. 8. MI versus delay curves obtained for the 0.74 kW TIM with 0.7 Nm of load torque, balanced power supply voltage, and variations of short circuit levels of phase A.

Table 3 Table 5
Classification results of the MLP network for the 0.74 kW TIM features. Confusion matrices of the MLP topologies for the 0.74 kW TIM.

Type MLP1 MLP2 MLP3 MLP1 MLP2 MLP3

Classification error 2/550 3/550 3/550 Normal Faulty Normal Faulty Normal Faulty
Classification accuracy (%) 99.64 99.45 99.45
Kappa index 0.989 0.983 0.983 Normal 108 2 107 3 107 3
Mean absolute error 0.0096 0.0095 0.0082 Faulty 0 440 0 440 0 440
Root mean squared error 0.0596 0.0726 0.0661
Relative absolute error 2.999 2.956 2.547
Root relative squared error 14.909 18.158 16.517
Recall 0.996 0.995 0.995 Similarly, for the RBF networks, the reached accuracies were
F-Measure 0.996 0.995 0.995 over 93%, regardless of the number of neurons in the interme-
ROC 1 0.999 1 diate layer. The network with 7 neurons (RBF3) obtained results
Time taken to build model (s) 9.39 18.83 44.41
slightly better than the others, with 95.82% of classification accu-
racy, as shown in Table 4. Despite that the RBF networks presented
low computational costs, their accuracy are relatively lower than
For the 0.74 kW TIM, both ANN topologies achieved accuracies
the MLP networks. Furthermore, other statistical indices, such as
over 93%, but the MLP networks had a better performance than the
the relative absolute error and root relative squared error, which
RBF ones. The MLP network with 10 neurons in the hidden layer
present results over 15% and 46%, respectively, confirmed that, for
(MLP1) reached the highest classification rate of 99.64%. This result
the 0.74 kW TIM, the RBF networks are less suitable in comparison
is slightly higher then those of the other MLP networks with 20 and
with the MLP ones.
50 neurons, which were 99.45%. But the greatest advantage of MLP1
The Kappa indices above 0.8 obtained by the two ANN topolo-
is that it has a lower computational weight as shown in Table 3 for
gies confirm their potential for classification of the TIM condition:
the time taken to build the model.
normal or faulty (stator short circuit). Tables 5 and 6 present the
Other statistical indices are also presented in the tables, such
confusion matrices of the MLP and RBF topologies results, respec-
as mean absolute error, root mean squared error, relative abso-
tively.
lute error, root relative squared error, recall, F-measure, and region
Based on these tables, for the MLP networks, only normal sam-
of convergence (ROC). For the MLP network with 10 neurons, the
ples were misclassified (false positives). This kind of error leads to
values of these indices were 0.0096, 0.0596, 2.999, 14.909, 0.996,
programmed interruption of the industrial process for an unnec-
0.996, and 1, respectively, thus confirming that it is the most ade-
essary maintenance. On the other hand, the false negative would
quate topology for the 0.74 kW TIM.
probably result in an unexpected stop due to the TIM collapse,
which could have a greater financial loss depending on the pro-
Table 4 cess costs. Despite the fact that the RBF network presented false
Classification results of the RBF network for the 0.74 kW TIM features. negative classifications, the accuracy rate and the Kappa index indi-
Type RBF1 RBF2 RBF3 cate that this topology had similar results to the MLP networks, but
with lower processing time.
Classification error 34/550 28/550 23/550
Classification accuracy (%) 93.82 94.91 95.82
Kappa index 0.803 0.843 0.863
Mean absolute error 0.0878 0.0638 0.0493 Table 6
Root mean squared error 0.2258 0.2004 0.1869 Confusion matrices of the RBF topologies for the 0.74 kW TIM.
Relative absolute error 27.378 19.879 15.363
Root relative squared error 56.448 50.109 46.72 RBF1 RBF2 RBF3
Recall 0.938 0.949 0.958
Normal Faulty Normal Faulty Normal Faulty
F-measure 0.938 0.949 0.958
ROC 0.945 0.974 0.967 Normal 90 20 98 12 98 12
Time taken to build model (s) 0.18 0.38 0.31 Faulty 14 426 16 424 5 435
354 G.H. Bazan et al. / Electric Power Systems Research 143 (2017) 347–356

Table 7 Table 9
Classification results of the MLP network for the 1.48 kW TIM features. Confusion matrices of the MLP topologies for the 1.48 kW TIM.

Type MLP1 MLP2 MLP3 MLP1 MLP2 MLP3

Classification error 1/264 1/264 1/264 Normal Faulty Normal Faulty Normal Faulty
Classification accuracy (%) 99.62 99.62 99.62
Kappa index 0.989 0.989 0.989 Normal 65 1 65 1 65 1
Mean absolute error 0.0043 0.0037 0.0038 Faulty 0 198 0 198 0 198
Root mean squared error 0.0429 0.0409 0.0493
Relative absolute error 1.136 0.993 1.012
Root relative squared error 9.908 9.442 11.374 Table 10
Recall 0.996 0.996 0.996 Confusion matrices of the RBF topologies for the 1.48 kW TIM.
F-measure 0.996 0.996 0.996
RBF1 RBF2 RBF3
ROC 1 1 1
Time taken to build model (s) 4.52 9.08 22.10 Normal Faulty Normal Faulty Normal Faulty

Normal 64 2 65 1 64 2
Table 8 Faulty 5 193 0 198 0 198
Classification results of the RBF network for the 1.48 kW TIM features.

Type RBF1 RBF2 RBF3

with accuracies of 98.48% for the normal operation tests. For the
Classification error 7/264 1/264 2/264
RBF networks, only RBF1 (3 neurons) presented misclassification
Classification accuracy (%) 97.35 99.62 99.24
Kappa index 0.930 0.989 0.979 for faulty conditions, obtaining accuracy of 97.47%. As for the classi-
Mean absolute error 0.0405 0.0055 0.0093 fication of the normal operation tests, accuracies over 96.97% were
Root mean squared error 0.1537 0.0624 0.0889 obtained.
Relative absolute error 10.757 1.47 2.464
Therefore, based on the results for both TIMs, the MLP networks
Root relative squared error 35.496 14.406 20.517
Recall 0.973 0.996 0.992
using the proposed features are considered successful for the clas-
F-measure 0.974 0.996 0.992 sification of stator short circuit faults of several levels. It is efficient
ROC 0.978 0.991 0.979 even in the presence of adversities commonly faced in industrial
Time taken to build model (s) 0.19 0.13 0.15 environment, such as load torque variations and power supply volt-
age unbalance.

For the 1.48 kW TIM, 264 experimental tests were performed, 66

tests for normal operating conditions and 198 tests for stator short
circuit conditions. Tables 7 and 8 present the classification results
concerning this machine for MLP and RBF networks, respectively.
Similarly, for the 1.48 kW TIM features, high performance of the
classification system was observed. The tests using this machine
achieved accuracies over 97%, the best values were 99.62% accuracy
for the MLP networks, regardless of the number of neurons in the
hidden layer, and also for the RBF network with 5 neurons (RBF2).
Although the classification accuracies are the same for both
topologies, it is worth to note that the root relative squared
error values are significantly different. MLP1 and MLP2 networks
obtained 9.908 and 9.442, respectively, versus 14.406 by the RBF2
network. This indicates that the MLP networks are the most ade-
quate for detecting stator short circuit faults of the 1.48 kW TIM.
The other statistical indices presented in Tables 7 and 8, such as
relative absolute error, recall, F-measure, and ROC, also support the
proposed diagnostic system using the two ANN topologies for the
1.48 kW TIM under different operating conditions of power sup-
ply voltage unbalance and load torque. For these tests, the Kappa
indices were over 0.93.
Tables 9 and 10 present the confusion matrices of the MLP and
RBF topologies results, respectively. All the MLP networks obtained
the same classification rates for both normal and faulty conditions,
5.1. Comparison with previous works
Based on previous works concerning stator faults, this paper
presents some important aspects that should be highlighted. While
several researchers, such as in [5,7,9,10,12], studied the signals in
the frequency domain using statistical measures, envelope analysis,
or preprocessing by DWT, this work considers time domain analysis
with MI values of currents from the TIM phases A and B.
Another important aspect of this work, that is similar to [5,12],
but usually not considered in other studies, refers to the emulation
of power supply voltage unbalance during signal acquisition. This
is important in order to verify the method robustness when the TIM
is subject to adverse operating conditions. In addition, the variation
of the TIM load toque was also employed in references [7,9,10].
Considering the classification accuracy, in [5,7,12], the authors
present graphical analyses of the obtained accuracies with no fur-
ther statistical discussion. In [9,10], accuracies range between 96%
and 99%, load torque variations were considered in the tests, but
not power supply voltage unbalance. For our work, the classifica-
tion accuracies were greater than 93%, considering TIMs under a
wide range of load conditions and voltage unbalance, besides the
use of several classification statistical indices. Table 11 summarizes
this comparison.

Table 11
Comparison summary of some studies reporting stator fault diagnosis.

Literature Variable load Unbalance Method Classifier Accuracy

[5] No Yes Envelope analysis SHFC NS

[7] Yes No FFT IAPSA/IRPSA NS
[9,10] Yes No Statistical measures ANN 96–99%
[12] No Yes DWT RMFI/RAT NS
This work Yes Yes Delayed MI ANN 93–99%

SHFC – slot harmonic frequency components, NS – not specified.

IAPSA – instantaneous active power signature analysis.
IRPSA – instantaneous reactive power signature analysis.
RMFI – relative value of maximum fault index, RAT – relative value of adaptive threshold.
 G.H. Bazan et al. / Electric Power Systems Research 143 (2017) 347–356
6. Conclusion
This paper describes a methodology for the detection of TIM sta-
tor winding faults using ANN with feature vectors obtained from
information measures. Specifically, it uses MI of the current signals
of two TIM phases, considering several load levels, including over-
load (120% of nominal load), and up to 4% power supply voltage
unbalance.
Two ANN topologies, with three configurations each, were used
to classify the feature vectors, namely MLP networks with 10, 20,
and 50 neurons in the hidden layer and RBF with 3, 5, and 7 neu-
rons in this layer. These vectors were created with MI values of the
current signal of TIM phase A and 150 shifted versions of the phase
B signal. Both network topologies presented accuracies over 93%
and ideal kappa indices. Besides that, the MLP networks obtained
better results than the RBF ones, presenting only false positive clas-
sifications.
In conclusion, based on these tests, the proposed features are
considered efficient for the detection of stator winding short circuit,
with 1% error margin for the MLP networks and 6% margin for the
RBF ones, even considering load torque variations and power sup-
ply voltage unbalance, common conditions in an industrial process.
In addition, for the implementation of the classification method
in an embedded system for future works, the best ANN topology
is the MLP with 10 neurons in the hidden layer, which obtained
classification accuracies over 99% for the experimental tests.
Acknowledgements
The authors would like to thank the Federal University of
Technology – Paraná, the Coordination for the Improvement of
Higher Level Personnel (CAPES) for the scholarships, the Araucária
Foundation for the Support of the Scientific and Technological
Development of the State of Paraná (processes no. 338/2012
and 06/56093-3), and the National Council for Technological and
Scientific Development (CNPq) (processes no. 474290/2008-5,
473576/2011-2, 552269/2011-5, and 201902/2015-0).
References
[1] M.B.K. Bouzid, G. Champenois, New expressions of symmetrical components
of the induction motor under stator faults, IEEE Trans. Ind. Electron. 60 (9)
(2012) 4093–4102.
[2] 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 (2015) 1–13.
[3] R.H.C. Palácios, I.N. da Silva, A. Goedtel, W.F. Godoy, A novel multi-agent
approach to identify faults in line connected three-phase induction motors,
Appl. Soft Comput. 45 (2016) 1–10.
[4] A. Bellini, F. Filippetti, C. Tassoni, G.-A. Capolino, Advances in diagnostic
techniques for induction machines, IEEE Trans. Ind. Electron. 5 (12) (2008)
4109–4126.
[5] P.C.M. Lamim Filho, R. Pederiva, J.N. Brito, Detection of stator winding faults
in induction machines using flux and vibration analysis, Mech. Syst. Signal
Process. 42 (1-2) (2014) 377–387.
[6] H. Henao, G.-A. Capolino, M. Fernandez-Cabanas, F. Filippetti, C. Bruzzese, E.
Strangas, R. Pusca, J. Estima, M. Riera-Guasp, S. Hedayati-Kia, Trends in fault
diagnosis for electrical machines: a review of diagnostic techniques, IEEE Ind.
Electron. Mag. 8 (2) (2014) 31–42.
[7] M. Drif, A.J.M. Cardoso, Stator fault diagnostics in squirrel cage three-phase
induction motor drives using the instantaneous active and reactive power
signature analyses, IEEE Trans. Ind. Inf. 10 (2) (2014) 1348–1360.
[8] D.A. Asfani, A.K. Muhammad, Syafaruddin, M.H. Purnomo, T. Hiyama,
Temporary short circuit detection in induction motor winding using
combination of wavelet transform and neural network, Expert Syst. Appl. 39
(5) (2012) 5367–5375.
[9] V.N. Ghate, S.V. Dudul, Optimal MLP neural network classifier for fault
detection of three phase induction motor, Expert Syst. Appl. 37 (4) (2010)
3468–3481.
[10] V.N. Ghate, S.V. Dudul, Cascade neural-network-based fault classifier for
three-phase induction motor, IEEE Trans. Ind. Electron. 58 (5) (2011)
1555–1563.
355
[11] M. Sahraoui, A. Ghoggal, S. Guedidi, S.E. Zouzou, Detection of inter-turn
short-circuit in induction motors using Park-Hilbert method, Int. J. Syst.
Assur. Eng. Manag. 5 (3) (2014) 337–351.
[12] N.R. Devi, D.V.S.S. Siva Sarma, P.V. Ramana Rao, Detection of stator incipient
faults and identification of faulty phase in three-phase induction motor –
simulation and experimental verification, IET Electr. Power Appl. 9 (8) (2015)
540–548.
[13] S.-H. Lee, S. Kim, J.M. Kim, C. Choi, J. Kim, S. Lee, Y. Oh, Extraction of induction
motor fault characteristics in frequency domain and fuzzy entropy, in:
Proceedings of the 2005 IEEE International Conference on Electric Machines
and Drives, IEEE, San Antonio, TX, USA, 2005, pp. 35–40.
[14] A. Verma, S. Sarangi, M.H. Kolekar, Stator winding fault prediction of
induction motors using multiscale entropy and grey fuzzy optimization
methods, Comput. Electr. Eng. 40 (7) (2014) 2246–2258.
[15] P.S. Bhowmik, M. Prakash, S. Pradhan, A novel neuro-classifier using
multiscale permutation entropy for motor fault diagnosis, in: Proceedings of
the 2014 IEEE Conference on Control Applications, IEEE, Juan Les Antibes,
France, 2014, pp. 370–375.
[16] M. Hernandez-Vargas, E. Cabal-Yepez, A. Garcia-Perez, Real-time SVD-based
detection of multiple combined faults in induction motors, Comput. Electr.
Eng. 40 (7) (2014) 2193–2203.
[17] H.L. Schmitt, P.R. Scalassara, A. Goedtel, W. Endo, Detecting bearing faults in
line-connected induction motors using information theory measures and
neural networks, J. Control Autom. Electr. Syst. 26 (5) (2015) 535–544.
[18] J. Jensen, C.H. Taal, Speech intelligibility prediction based on mutual
information, IEEE Trans. Audio Speech Lang. Process. 22 (2) (2014) 430–440.
[19] D. Coats, K. Cho, Y.-J. Shin, N. Goodman, V. Blechertas, A.-M.E. Bayoumi,
Advanced time-frequency mutual information measures for condition-based
maintenance of helicopter drivetrains, IEEE Trans. Instrum. Meas. 60 (8)
(2011) 2984–2994.
[20] A. Joshi, P. Deignan, P. Meckl, G. King, K. Jennings, Information theoretic fault
detection, in: Proceedings of the 2005 American Control Conference, vol. 3,
IEEE, Portland, OR, USA, 2005, pp. 1642–1647.
[21] J. Faiz, B.M. Ebrahimi, B. Akin, H.A. Toliyat, Dynamic analysis of mixed
eccentricity signatures at various operating points and scrutiny of related
indices for induction motors, IET Electr. Power Appl. 4 (1) (2010) 1–16.
[22] W.F. Godoy, I.N. Silva, A. Goedtel, R.H.C. Palácios, Evaluation of stator winding
faults severity in inverter-fed induction motors, Appl. Soft Comput. 32 (2015)
420–431.
[23] R.H.C. Palácios, I.N. Silva, A. Goedtel, W.F. Godoy, A comprehensive evaluation
of intelligent classifiers for fault identification in three-phase induction
motors, Electr. Power Syst. Res. 127 (2015) 249–258.
[24] A.H. Bonnett, C. Yung, Increased efficiency versus increased reliability, IEEE
Ind. Appl. Mag. 14 (1) (2008) 29–36.
[25] S.S. Moosavi, A. Djerdir, Y. Ait-Amirat, D.A. Khaburi, ANN based fault diagnosis
of permanent magnet synchronous motor under stator winding shorted turn,
Electr. Power Syst. Res. 125 (2015) 67–82.
[26] M. Eftekhari, M. Moallem, S. Sadri, M.-F. Hsieh, Online detection of induction
motor’s stator winding short-circuit faults, IEEE Syst. J. 8 (4) (2014)
1272–1282.
[27] A. Gandhi, T. Corrigan, L. Parsa, Recent advances in modeling and online
detection of stator interturn faults in electrical motors, IEEE Trans. Ind.
Electron. 58 (5) (2011) 1564–1575.
[28] A.G. Garcia-Ramirez, L.A. Morales-Hernandez, R.A. Osornio-Rios, J.P.
Benitez-Rangel, A. Garcia-Perez, R.J. Romero-Troncoso, Fault detection in
induction motors and the impact on the kinematic chain through
thermographic analysis, Electr. Power Syst. Res. 114 (2014) 1–9.
[29] R. Sharifi, M. Ebrahimi, Detection of stator winding faults in induction motors
using three-phase current monitoring, ISA Trans. 50 (1) (2011) 14–20.
[30] M. Seera, C.H. Lim, D. Ishak, H. Singh, Fault detection and diagnosis of
induction motors using motor current signature analysis and a hybrid
FMM-CART model, IEEE Trans. Neural Netw. Learn. Syst. 23 (1) (2012)
97–108.
[31] B.M. Ebrahimi, M.J. Roshtkhari, J. Faiz, S.V. Khatam, Advanced eccentricity
fault recognition in permanent magnet synchronous motors using stator
current signature analysis, IEEE Trans. Ind. Electron. 61 (4) (2014) 2041–2052.
[32] J. Pons-Llinares, J.A. Antonino-Daviu, M. Riera-Guasp, S.B. Lee, T. Kang, C. Yang,
Advanced induction motor rotor fault diagnosis via continuous and discrete
time-frequency tools, IEEE Trans. Ind. Electron. 62 (3) (2015) 1791–1802.
[33] J. Antonino-Daviu, S. Aviyente, E.G. Strangas, M. Riera-Guasp, Scale invariant
feature extraction algorithm for the automatic diagnosis of rotor asymmetries
in induction motors, IEEE Trans. Ind. Inf. 9 (1) (2013) 100–108.
[34] P. Shi, Z. Chen, Y. Vagapov, Z. Zouaoui, A new diagnosis of broken rotor bar
fault extent in three phase squirrel cage induction motor, Mech. Syst. Signal
Process. 42 (1–2) (2014) 388–403.
[35] P. Konar, P. Chattopadhyay, Multi-class fault diagnosis of induction motor
using Hilbert and Wavelet transform, Appl. Soft Comput. 30 (2015) 341–352.
[36] J. Seshadrinath, B. Singh, B.K. Panigrahi, Investigation of vibration signatures
for multiple fault diagnosis in variable frequency drives using complex
wavelets, IEEE Trans. Power Electron. 29 (2) (2014) 936–945.
[37] M. Barzegaran, A. Mazloomzadeh, O. Mohammed, Fault diagnosis of the
asynchronous machines through magnetic signature analysis using
finite-element method and neural networks, IEEE Trans. Energy Convers. 28
(4) (2013) 1064–1071.
[38] L. Frosini, A. Borin, L. Girometta, G. Venchi, A novel approach to detect short
circuits in low voltage induction motor by stray flux measurement, in:
356 G.H. Bazan et al. / Electric Power Systems Research 143 (2017) 347–356
Proceedings of the 2012 International Conference on Electrical Machines,
IEEE, Marseille, France, 2012, pp. 1538–1544.
[39] Z. Gao, T.G. Habetler, R.G. Harley, R.S. Colby, A sensorless adaptive stator
winding temperature estimator for mains-fed induction machines with
continuous-operation periodic duty cycles, IEEE Trans. Ind. Appl. 44 (5)
(2008) 1533–1542.
[40] M.-Y. Chow, R.N. Sharpe, J.C. Hung, On the application and design of artificial
neural networks for motor fault detection – Part I, IEEE Trans. Ind. Electron. 40
(2) (1993) 181–188.
[41] M.-Y. Chow, R.N. Sharpe, J.C. Hung, On the application and design of artificial
neural networks for motor fault detection – Part II, IEEE Trans. Ind. Electron.
40 (2) (1993) 189–196.
[42] S.R. Kolla, S.D. Altman, Artificial neural network based fault identification
scheme implementation for a three-phase induction motor, ISA Trans. 46 (2)
(2007) 261–266.
[43] P. Konar, P. Chattopadhyay, Bearing fault detection of induction motor using
wavelet and support vector machines (SVMs), Appl. Soft Comput. 11 (6)
(2011) 4203–4211.
[44] F. Zidani, M.E.H. Benbouzid, D. Diallo, M.S. Nait-Said, Induction motor stator
faults diagnosis by a current Concordia pattern-based fuzzy decision system,
IEEE Trans. Energy Convers. 18 (4) (2003) 469–475.
[45] P.V.J. Rodríguez, A. Arkkio, Detection of stator winding fault in induction
motor using fuzzy logic, Appl. Soft Comput. 8 (2) (2008) 1112–1120.
[46] 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, Expert Syst. Appl.
36 (2) (2009) 1840–1849.
[47] M. Seera, C.P. Lim, D. Ishak, H. Singh, Offline, Online fault detection and
diagnosis of induction motors using a hybrid soft computing model, Appl. Soft
Comput. 13 (12) (2013) 4493–4507.
[48] T.M. Cover, J.A. Thomas, Elements of Information Theory, 2nd ed., Wiley, New
Jersey, 2006.
[49] A. Dionisio, R. Menezes, D.A. Mendes, Mutual information: a measure of
dependency for nonlinear time series, Physica A: Stat. Mech. Appl. 344 (1–2)
(2004) 326–329.
[50] R. Steuer, J. Kurths, C.O. Daub, J. Weise, J. Selbig, The mutual information:
detecting and evaluating dependencies between variables, Bioinformatics 18
(Suppl. 2) (2002) S231–S240.
[51] J.M. Nichols, Examining structural dynamics using information flow, Probab.
Eng. Mech. 21 (4) (2006) 420–433.
[52] J.F. Alonso, M.A. Mañanas, D. Hoyer, Z.L. Topor, E.N. Bruce, Evaluation of
respiratory muscles activity by means of cross mutual information function at
different levels of ventilatory effort, IEEE Trans. Biomed. Eng. 54 (9) (2007)
1573–1582.
[53] U. Melia, M. Guaita, M. Vallverdú, C. Embid, I. Vilaseca, M. Salamero, J.
Santamaria, Mutual information measures applied to EEG signals for
sleepiness characterization, Med. Eng. Phys. 37 (3) (2015) 297–308.
[54] S. Haykin, Neural Networks and Learning Machines, 3rd ed., Pearson, New
Jersey, 2008.
[55] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, I.H. Witten, The
WEKA data mining software: an update, ACM SIGKDD Explor. Newsl. 11 (1)
(2009) 10–18.