Information Theoretical Measurements From Induction Motors Under Several Load and Voltage Conditions For Bearing Faults Classification
Information Theoretical Measurements From Induction Motors Under Several Load and Voltage Conditions For Bearing Faults Classification
Information Theoretical Measurements From Induction Motors Under Several Load and Voltage Conditions For Bearing Faults Classification
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
Abstract—Three-phase induction motors are widely employed electrical machines. In addition, such motors easily adapt to
in industrial production processes, due to their low costs of different coupled load conditions and are robust to be used
acquisition and maintenance as well as their adaptation in front in harsh environments [1]–[4]. However, these motors are
of different coupled load conditions and robustness for use in susceptible to several types of failures, which can be divided
harsh environments. Methods that detect an incipient fault when into electrical and mechanical. Among the mechanical failures,
these motors are in operation, even before significant faults may
occur, have received special attention from the industry. Thus, a
several researchers in recent works highlight the bearing faults,
proper diagnosis can lead to a reduction of process losses, service which represent between 40% to 52% of all faults [1], [3]–[5].
time out, and equipment replacement, also ensure consistent and These faults can affect the operating condition of the machine,
reliable operation of industrial systems. Thus, this paper presents cause unexpected time out, and have catastrophic effects. In
a comprehensive study concerning the detection of bearing faults general, the bearing faults are associated with contamination,
in induction motors using information theoretical measurements corrosion, inappropriate lubrication, and installation problems,
and intelligent tools. Various motor operation conditions were such as misalignment, and overload [5].
considered, such as variations in coupled load torque and two Because of the several kinds of failures in three-phase induc-
kinds of power supply: line connected or inverter fed. This tion motors (TIMs), many researches are proposing method-
work proposes the extraction of fault characteristics in time
domain, based on mutual information measurements between
ologies to monitor the conditions of these machines aiming at
two phase current signals. The method was evaluated under detecting the initial faults related to them. This offers important
various fault severity levels and the performance of three different advantages to an industrial productive process, such as the
pattern recognition techniques were compared: artificial neural prevention of costly shutdowns, the reduction of maintenance
networks, specifically multilayer perceptron, k-nearest neighbors, costs, and results in a more reliable operation [1]–[6].
and support vector machines. The experimental results presented The majority of the methods used for condition monitoring
in this study validate the robustness and effectiveness of the and fault detection of induction motors are classified based
proposed methodology. on the analyzed physical measurement, being divided into the
Keywords—Three-Phase Induction Motors; Bearing Fault Detec- following categories: vibration monitoring, temperature mon-
tion; Mutual Information; Pattern Recognition itoring, oil or debris analysis, acoustic emission monitoring,
and current, voltage, power or magnetic signal monitoring
I. I NTRODUCTION [2]. However, some of these methodologies require additional
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
they are capable of detecting bearing faults regardless of the type of fault without demanding complex mathematical mod-
choice and positioning of the adopted sensor [7]. Therefore, in els, which makes their implementation simpler and accessible.
this paper, the analysis of current signals is proposed, which Moreover, along with features that capture the fault signatures,
are collected by means of Hall effect sensors. intelligent systems can classify and show incipient motor
In order to improve the performance of traditional fault de- faults, thus contributing to an efficient industrial process main-
tection methods of induction motors, recent works use several tenance and reducing costs. Among these systems, one can
data processing techniques [8]. They use tools that extract and mention artificial neural networks (ANN) [20], [21], support
select the most relevant characteristics of the signals. Some vector machines (SVM) [7], [22], [23], k-nearest neighbors
authors work with methods of feature extraction in frequency (k-NN) [22], and hybrid systems [24].
domain that perform spectral decomposition of signals, such In spite of that, when TIMs are supplied by frequency invert-
as Fourier transform [1]–[6], Wavelet transform [2], [4], [9], ers, the fault detection methods become more complex. Mainly
and Hilbert transform [10], [11]. because of changes of supply frequency, noise increase, and
Some of those techniques based on the frequency domain are amplitude variations of the magnetic field harmonics. The latter
being associated with other methods in order to result in a more is influenced by supply voltage and dependent of the inverter
precise diagnosis of the motors. In [2], statistical parameters control technique and switching frequencies. Nowadays, the
extracted from least squares estimators are used to improve researchers’ attention is focused on fault diagnostic tools that
the identification of bearing and rotor faults in induction can identify an inadequate TIM operation independently of the
motors. Other authors, such as [3], use the root-multiple signal kind of power supply [25].
classification (MUSIC) method to help in detecting TIM faults. Thus, the main contribution of this work consists in present-
Recently, in [4], a new approach to detect bearing faults was ing an alternative methodology for the diagnosis of bearing
proposed, it uses the continuous wavelet transform combined faults in TIMs based on MI measures of current signals
with criteria based on energy to extract and identify failures. obtained in time domain, regardless of the machine power
In [5], algorithms based on the spectral kurtosis, such as supply, either obtained directly from the supply network or
fast kurtogram and Wavelet kurtogram, are used to detect the by means of frequency inverters. Furthermore, an evaluation
severity of TIM external bearing faults. On the other hand, [12] of the pattern classifiers is performed in order to obtain the best
and [13] used finite elements techniques, as well as statistical one to identify the fault characteristics extracted in the stage
analysis, for the development of methods to diagnose faults in of data processing. The considered classifiers are multilayer
the bearings of electrical machines. perceptron (MLP) ANN, k-NN, and SVM.
In addition to the kind of signal under analysis, it is Additionally, in order to validate the proposed model, 3,544
important to consider the manner in which these signals were samples were collected from motors operating under perma-
acquired and the domain in which the analysis is performed. nent regime from sinusoidal power sources directly fed by
Techniques to reduce data dimensionality are also deemed supply networks and also by frequency inverters. From these
important in the field of machine learning, data mining, and experimental tests, 1,100 were performed on line connected
pattern recognition. Preprocessing high-dimension data with TIMs, including tests with variations of power supply voltage
appropriate tools can lead to less computational processing unbalance and coupled load torque (10%≤Tn ≤120%, where
costs and also to a better interpretation of the features, which Tn is the nominal torque). Also, 2,444 samples were collected
results in performance improvement [14]. from TIMs fed by three different frequency inverters, with
The extraction methods are aimed at creating new character- variations of frequency from 20% to 100% of the nominal
istics by transforming or combining an original data set [15]. value. The coupled load variations were the same used for
Feature selection consists in the process of selecting a subset direct supply in two different motors: 0.74 and 1.48 kW, named
of characteristics that have the most relevant information of Motor 1 and Motor 2, respectively. The severity levels analyzed
the original set, so that the classes can be distinguished from in both cases are based on 15 to 90 minutes of bearings wear
each other, aiming at data quality and classifier performance from the two TIMs.
[15], [16]. This paper is divided into five sections. Section II describes
A modern approach to pattern recognition of TIM faults is the main concepts of the used tools. Section III shows the
based on Information Theory (IT), which uses a set of tools materials and methods. The experimental results are presented
that statistically measure the association between the variables and discussed in Section IV and, finally, Section V brings a
involved in the problem [17], [18]. According to [19], the conclusion and final remarks.
Mutual Information (MI) is a measure of association of signals,
in addition to being able to identify the linear and non-linear II. I NFORMATION M EASURES AND C LASSIFIERS
dependencies among time series. As the TIM is a complex In this section, the main tools used for the proposed eval-
non-linear system, the approach proposed in this paper uses uation system of bearing fault detection of line connected or
the MI as a measure of association of the phase currents in frequency inverter fed TIMs are presented. Also, the general
order to extract significant information that enable an efficient aspects of the employed feature classifiers are discussed.
diagnosis of bearing faults.
Similar to the signal processing tools, intelligent systems A. Information Theory Measurements
have become relevant in the identification of faults in electrical Several IT measurements are applied to studies and prob-
machines. These techniques are efficient in determining the lems related to the field of pattern recognition, aiming at
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
identifying measures of association or connectivity between capacity for parallel operation, no requirement for complicated
signals and relevant variables, in addition to their applicability mathematical models, the fact that they are systems inspired by
in classification methods [17], [18]. This has become possible the human brain, and their ability to learn based on experience
with the introduction of two concepts by Shannon: entropy [29], [30]. One of the most used ANN topologies is the MLP,
and MI [26]. The first is defined as a measure of uncertainty that is a network that consists of at least one hidden neural
of isolated or combined random variables and MI consists layer, between the input and output layers.
in the measure of the stochastic dependence between these This topology uses the backpropagation algorithm, which
variables and quantifies the common information and measure has two stages: forward and backward processes. The first one
of association between them. So, in this section, the main is a supervised stage in which the input vectors of a sample
concepts related to this quantity are described. of the training set are propagated layer by layer until their
1) Mutual Information: MI is a quantity that indicates the respective outputs are produced in order to obtain the output
reduction of the level of uncertainty associated with a ran- of the network. In this step, only the current values of the
dom variable due to information brought by another variable synaptic weights and neuron thresholds are considered. As it
[26]. This measurement describes the association among time is a supervised learning process, the outputs generated by the
series collected simultaneously [19], [27], [28]. Therefore, network outputs are compared with their respective desired
by applying this tool in signals of an electrical machine, outputs. Therefore, in the second stage of the backpropagation
their association can be quantified, which, therefore, enables algorithm, the synaptic weights and neuron thresholds are
the efficient monitoring of the machine conditions. Eq. 1 adapted based on the error between the current output and
determines this measure: the desired one. By applying these two stages successively,
XX p(x, y) these values are automatically altered in each iteration of the
I(X, Y ) = p(x, y) log2 (1) training process, resulting in a gradual decrease of the sum of
x∈χ y∈γ
p(x) · p(y)
the errors produced by the outputs of the networks regarding
where X and Y are random variables; p(x) and p(y) are the their respective desired values [29], [30].
marginal probability density functions (pdfs) of X and Y , Because of the ample application of this type of classifier
respectively; and p(x, y) is the joint pdf between these two in problems related to faults in electrical machines, such as in
variables. [20], [21], the ANNs are here employed to classify the patterns
In this work, the analyzed time signals are the currents of of the bearing faults in TIMs.
phases A and B of the motor, which are 120 degrees out of
phase with each other. Therefore, it is necessary to use another C. k-Nearest Neighbors
approach known as shifted mutual information (SMI). The k-NN classifier was developed from the algorithm
2) Shifted Mutual Information: SMI is a measure that allows of the nearest neighbor (NN) [31]. It is a non-parametric
the quantification of the dependence or association of time se- methodology based on a modification of the original algorithm
ries based on a shift τ of one of them [19], [27], [28]. The shift in the training and validation steps, as the algorithm employs
τ can be viewed as the delay of the information transmitted the k nearest neighbors, unlike NN, which uses only the nearest
between both signals [19]. In this work, the shift τ is important neighbor. The k-NN methodology classifies a sample according
in order to analyze signals that are not synchronized. to its distance to certain pre-labeled training samples, that is,
According to [27], if the random variables, X and Y , are if most of the samples near the unknown sample belongs to
considered as being stationary signals, their joint probability a certain class, this class will be attributed to this sample.
will depend only on the displacement τ . Therefore, the shifted Usually, the Euclidean distance method is adopted.
mutual information is defined by Eq. 2: Moreover, it is possible to state that the k-NN is a classifier
that has only one free parameter, the number of k-neighbors,
X X p(xn , yn−τ ) controlled by the user to obtain a better classification. In
I(X, Y τ ) = p(xn , yn−τ ) log2
xn ∈χ yn−τ ∈γ
p(x n ) · p(yn−τ ) recent studies related to the recognition of faults in electrical
(2) machines, some authors mention the use of the k-NN method,
where p(xn ) and p(yn−τ ) are the marginal pdfs of X e Y τ as shown in [22].
respectively and p(xn ,yn−τ ) is their joint pdf dependent of the
shift τ . D. Support Vector Machines
After this brief presentation of MI concepts, related to the Support vector machines, developed by Vladimir Vapnik
feature extraction stage, the general aspects of the pattern [32], are learning systems trained with an optimization of a
classifiers are described. In this work, MLP ANN, k-NN, and mathematical algorithm that implements a limit derived from
SVM are used because they are techniques widely applied statistical learning theory. This process consists in training
in recent researches in the fields of data mining, pattern a classifier which learns a mapping through the presented
recognition, and detection of electrical machines faults [20]– samples, so that the machine is capable of classifying a
[24]. previously unseen sample by following the same probabil-
ity distribution of the training samples. Therefore, decision-
B. Artificial Neural Networks making in classification problems can be performed by means
ANNs are computational models that are ubiquitous in of functions that use hyperplanes to divide the feature space
engineering problems due to features such as their intrinsic of the samples into regions.
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
SVM has been proven to be a promising tool to classify conditions previously described were considered for each level
faults in electrical machines as shown in [22] and [23]. With of excessive wear in the bearings.
this in mind, this technique was also chosen in this work as a For the tests performed with the motor supplied by the
classifier of TIM conditions. frequency inverters, the 2,444 samples were divided as follows:
for Motor 1, 1,258 samples were collected with coupled load
varying from 0.5 N.m to 5.0 N.m, in intervals of 0.5 N.m,
III. P ROPOSED C ONDITION M ONITORING S YSTEM
and the supply frequency varied from 20% to 100% of the
In order to consider the several mentioned TIM operation nominal value, with intervals of 10%. In the case of Motor
conditions, an experimental database was created by acquiring 2, 1,186 samples were acquired with coupled load varying
current signals of two TIMs operating without faults and with from 1.0 N.m to 9.0 N.m, in intervals of 1.0 N.m, with the
bearing faults. These faults were emulated as conditions of same supply frequency variations of Motor 1. It is important
excessive use, lack of lubrication, and excessive coupled load. to mention that the motors were fed by three frequency
After cleaning the bearing, the lubrication was replaced by inverters of different models, as follows: Schneider Electric
an abrasive slurry used to wear the pieces, and the motor was ATV12HU15M2 (ATV), Siemens Sinamics G110 (G110), and
operated for 15, 30, 60, and 90 minutes for bearing degradation Siemens MicroMaster 440 (MM440). As in the case of the
resulting in four levels of severity of the abrasive wear. After line connected TIMs, the conditions previously described were
this process, the bearings were cleaned and lubricated again, considered for each level of excessive wear in the bearings.
so as to acquire data appropriately. Fig. 1 shows the procedure For this study, data were acquired using the experimental
of bearing degradation used in this work. workbench shown in Fig. 2, which was designed to monitor
electrical and mechanical quantities, such as voltage, current,
1 2 3 vibration, torque, and speed. Moreover, it is possible to emulate
several TIM operation conditions, such as variations of power
supply and coupled load torque.
DC Generator
Load
4 5 Voltage Vibration
DC Generator
Variator Sensor
Signal
Conditioning
Torque Meter
Fig. 1. Procedure of bearing degradation. (1) Process of removing the Display
bearings from the motor shaft. (2) Bearings prepared for fault emulation. (3) DAQ 6221
Use of abrasive slurry to emulate excessive use and lack of lubrication. (4) DC Supply
Use of paint thinner to clean the abrasive slurry. (5) Bearings lubricated and
prepared for use.
Fig. 2. Experimental workbench.
In this context, the experimental database consists of 3,444
tests performed with Motor 1 (0.74 kW) and Motor 2 (1.48 The variation of motor power supply is performed by
kW) operating under stationary regime, with power supply fed independent phase voltage regulators. The imposition of shaft
by the electrical grid or by means of frequency inverters, in load on the TIM occurs by coupling a direct current machine.
addition to several operation conditions, all of which common The workbench also has a dual-range Kistler torque meter, that
to an industrial environment. acts as a built-in speed sensor and is responsible for reading the
For the trials performed using line connected motors, 1,100 torque and speed data. The current signals used in this work are
samples were collected, which were divided as follows: for collected individually through Hall effect sensors, which are
Motor 1, 550 samples were acquired with coupled load torque responsible for conditioning these signals before transferring
varying from 0.5 N.m to 5.0 N.m, in intervals of 0.5 N.m; them to the analogical inputs of a National Instruments data
the voltage unbalance in phase A, from 0 to +10%, with acquisition device, DAQ USB-6221, which is connected to a
intervals of +2%; in phase B, from 0 to +10%, with intervals microcomputer.
of +2%, and in phase C, from 0 to -10%, with intervals of -2%. Based on the use of an algorithm for data preprocessing
Regarding Motor 2, 550 samples were acquired with coupled and acquisition, the previously described signals are processed
load varying from 1.0 N.m to 9.0 N.m, in intervals of 1.0 N.m, and their results are evaluated in MATLAB 2012b. This step
with the same asymmetrical voltage unbalance of Motor 1. The is aimed at extracting relevant information of the data, so as
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
displays the best behavior for the total test set of Motor 1. classes. The kappa indexes of 0.94, 0.88, and 0.94, for the
One of the possible reasons for the incorrect classification MLP, k-NN, and SVM classifiers, respectively, confirm the
of the samples without faults is that the abrasive wear at results obtained for this Motor 2 operation condition. As the
15 and 30 minutes are initial faults, which can induce the voltage unbalance increases, the classification rates decrease,
diagnosis system to erroneously identify this type of sample according to the results shown for Motor 1.
as a machine operating without faults. In opposite situations, Once more, the classification accuracies can be highlighted
this also occurs. In addition, considering that the bearing faults for the MLP classifier that achieved, even with a voltage
tend to have a symmetrical behavior in all phases of the motor, unbalance of 4%, classification rates of 96% of the global
this type of fault can be erroneously classified, with the motor sample set, 88% of the healthy samples, and 98% of the faulty
operating without faults and with voltage unbalance. samples, as shown in Table V. It is important to verify that the
In a similar manner to the tests performed on Motor 1, tests kappa index for this test was 0.87, which confirms the good
were conducted using the data set of Motor 2, with coupled results of the MLP classifier. This indicates that the MLP is
load varying from 10% to 120% of the nominal value and the most adequate for MI feature classification.
power supply voltage unbalance up to 10% in phases A, B, The results shown in this section reveal that the proposed
and C. These tests are aimed at verifying whether the data diagnosis system is a promising tool to identify bearing faults
shows a similar behavior to that observed for Motor 1. Tables of line connected TIMs. In the next section, new tests are
V, VI, and VII show the classification results for this data set presented with the TIMs fed by frequency inverters, in order
when using the MLP, k-NN, and SVM classifiers, respectively. to verify whether the methodology is robust when facing a
It is important to mention that Motor 2 is also subject to the different type of power supply.
abrasive wear in the bearings for 15, 30, 60, and 90 minutes,
in addition to the variations in the coupled load and voltage B. Results for Motors Fed by Frequency Inverters
unbalance of the phases, which are common situations in an This section shows the results of experimental tests with the
industrial environment. motors fed by frequency inverters. Such tests are performed
in order to verify whether the method used in this work is
TABLE V. R ESULTS CONSIDERING VARIATIONS OF THE COUPLED
LOAD OF 10% TO 120% OF THE NOMINAL VALUE - M OTOR 2 - MLP robust regardless of the type of machine power supply, either
CLASSIFIER line connected or inverter fed. In a similar manner to Section
MLP
IV-A, the MI values obtained in the previous stage of feature
Voltage unbalance ≤0.5% ≤2% ≤4% ≤6% ≤8% ≤10% extraction are used as input data for the MLP network, k-NN,
Total accuracy (%) 98.0 98.7 96.0 92.9 93.1 92.0 and SVM classifiers. In order to verify the behavior of the TIM
Healthy accuracy (%) 100 93.3 88.0 78.6 77.8 74.5
Faulty accuracy (%) 97.5 100 98.0 96.4 96.9 96.4
facing various operation conditions, these tests were conducted
Kappa index 0.94 0.96 0.87 0.77 0.78 0.74 with variations in the supply frequency from 20% to 100% of
Building model (s) 5.5 16.5 27.0 40.8 53.4 64.6 the nominal value, as well as variations in the coupled load,
as described in Section III. It is important to highlight that
TABLE VI. R ESULTS CONSIDERING VARIATIONS OF THE COUPLED the results are separated by frequency inverters, as each one
LOAD OF 10% TO 120% OF THE NOMINAL VALUE - M OTOR 2 - k-NN has a specific switching feature, according to [39]. Therefore,
CLASSIFIER
Table VIII shows the classification results for the data set of
k-NN Motor 1, when this motor is subject to the abrasive wear in
Voltage unbalance ≤0.5% ≤2% ≤4% ≤6% ≤8% ≤10%
Total accuracy (%) 96.0 87.3 86.0 86.3 86.9 84.2
the bearings for 15, 30, 60, and 90 minutes.
Healthy accuracy (%) 100 50.0 56.0 57.1 54.4 49.1 Based on Table VIII, when the motor is fed by the ATV
Faulty accuracy (%) 95.0 96.7 93.5 93.6 95.0 93.0 frequency inverter, the best results are achieved when the MLP
Kappa index 0.88 0.54 0.53 0.54 0.55 0.46
Building model (s) 0.01 0.01 0.01 0.01 0.01 0.01 classifier is used. In this case, correct classifications occur for
97.3% of the global sample set, 87.8% of the healthy sample
set, and 99.7% of the faulty samples. In addition, the kappa
TABLE VII. R ESULTS CONSIDERING VARIATIONS OF THE COUPLED
LOAD OF 10% TO 120% OF THE NOMINAL VALUE - M OTOR 2 - SVM index of 0.91 confirms these results. When the k-NN classifier
CLASSIFIER is used, similar results are achieved, while the SVM classifier
SVM
shows an inferior accuracy compared with the others.
Voltage unbalance ≤0.5% ≤2% ≤4% ≤6% ≤8% ≤10% For the data set obtained when the motor is fed by the
Total accuracy (%) 98.0 94.7 88.0 86.3 86.9 86.2 G110 inverter, the MLP and SVM classifiers presented global
Healthy accuracy (%) 100 73.3 46.0 35.7 40.0 38.2
Faulty accuracy (%) 97.5 100 98.5 98.9 98.6 98.2
classification rates of 98.1%; however, the SVM classifier
Kappa index 0.94 0.81 0.54 0.45 0.49 0.46 reached a superior rate in the classification of the healthy
Building model (s) 0.03 0.01 0.02 0.04 0.05 0.08 samples. The kappa indexes of 0.94 indicate a total agreement
with the results obtained when using both classifiers, either
As shown in Tables V, VI, and VII, for a voltage unbalance the MLP or the SVM. The k-NN classifier obtained inferior
up to 0.5%, the three pattern classifiers, MLP network, k- classification rates, especially for the healthy samples, reaching
NN and SVM, show satisfactory outcomes, reaching 98%, an accuracy of 74% for this sample set.
96%, and 98% classification accuracies for the total set of When using the MM440 frequency inverter with Motor 1,
samples, respectively. This shows that under normal operation the three classifiers obtained satisfactory accuracies, all of them
situations, the proposed system can effectively separate the above 98%, in addition to kappa indexes greater than 0.95,
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
TABLE VIII. R ESULTS CONSIDERING SUPPLY FREQUENCY VARIATION FROM 20% TO 100% OF THE NOMINAL VALUE - M OTOR 1
ATV G110 MM440
Classifier MLP k-NN SVM MLP k-NN SVM MLP k-NN SVM
Total accuracy (%) 97.3 96.7 92.4 98.1 93.2 98.1 99.5 98.4 99.5
Healthy accuracy (%) 87.8 85.6 64.4 91.8 74.0 94.5 98.9 94.3 97.7
Faulty accuracy (%) 99.7 99.4 99.4 99.7 98.0 99.0 99.7 99.4 99.4
Kappa index 0.91 0.89 0.73 0.94 0.77 0.94 0.99 0.95 0.99
Building model (s) 17.7 0.01 0.17 15.1 0.01 0.06 16.2 0.01 0.03
TABLE IX. R ESULTS CONSIDERING SUPPLY FREQUENCY VARIATION FROM 20% TO 100% OF THE NOMINAL VALUE - M OTOR 2
ATV G110 MM440
Classifier MLP k-NN SVM MLP k-NN SVM MLP k-NN SVM
Total accuracy (%) 98.1 93.7 92.3 99.7 98.2 99.7 99.2 99.7 98.2
Healthy accuracy (%) 90.4 72.3 61.4 100 94.7 100 100 98.7 93.7
Faulty accuracy (%) 100 99.1 100 99.7 99.0 99.7 99.0 100 99.4
Kappa index 0.94 0.78 0.72 0.99 0.94 0.99 0.98 0.99 0.94
Building model (s) 13.0 0.01 0.26 13.6 0.01 0.04 12.4 0.01 0.03
which confirms the good separation between the classes with and IV-B, the results showed that the MLP classifier is the
and without faults. more adequate to be used along with the proposed feature
Therefore, considering the results shown for all three invert- extraction method, therefore being selected for these new tests.
ers, when working with a wide frequency range, 20% to 100% Also, the tests were separated according to the type of power
of the nominal value, the most appropriate classifier is the MLP supply: line connected motors or fed by frequency inverters,
with which the highest classification rates were achieved. in consonance to Sections IV-A and IV-B.
As shown for Motor 1, tests were performed with Motor 2 Table X presents the classification results for line connected
fed by frequency inverters. For this data set, supply frequency Motors 1 and 2. As described previously, the motors were
and coupled load were varied in similar levels to those used subjected to bearings abrasive wear for 15, 30, 60, and 90
for Motor 1. Table IX shows the classification results for the minutes, along with variations of coupled load from 10% to
total data set of Motor 2 with abrasive wear of the bearings 120% of the motor nominal value and voltage unbalance of
for 15, 30, 60, and 90 minutes, in addition to considering the phases A, B, and C up to 10%.
three different types of frequency inverters. TABLE X. R ESULTS FOR I NTERCHANGING M OTOR DATA - M OTOR 1
When using the ATV frequency inverter, there was a better FOR T RAINING AND M OTOR 2 FOR VALIDATION - L INE C ONNECTED
separation between the healthy and faulty classes when the M OTORS
MLP classifier was employed, as described in Table IX. The Classifier MLP
classification accuracies were 98.1% of the total samples, in Voltage unbalance ≤0.5% ≤4% ≤10%
addition to 90.4% and 100% of the healthy and faulty samples, Total accuracy (%) 98.0 94.0 92.9
Healthy accuracy (%) 90.0 78.0 79.1
respectively. Therefore, for this data set, the MLP is the most Faulty accuracy (%) 100 98.0 96.4
appropriate classifier. Kappa index 0.94 0.80 0.77
Building model (s) 2.79 14.2 21.3
On the other hand, when the G110 and MM440 frequency
inverters used to fed Motor 2, regardless of the classifier, either As the voltage unbalance levels increase, the classification
MLP, k-NN, or SVM, global classification rates above 98.2% accuracies tend to decrease, as was shown in Section IV-A. It
are achieved, as well as kappa indexes greater than 0.94, which is noteworthy that considering voltage unbalance up to 0.5%,
confirms the results. Therefore, promising results have been what is considered normal in an industrial environment, the
achieved, regardless of the classifier. proposed method achieved accuracies of 98.0%, 90%, and
The results shown in Sections IV-A and IV-B demonstrate 100% respectively for the total sample set, healthy, and faulty
that the proposed system has a promising performance for samples. The kappa index was 0.94, what indicates a total
detection of bearings faults regardless of power supply and agreement of the results. Also, even with voltage unbalance of
coupled load. Additionally, aiming at increasing the reliability 10%, what is considered an extremely adverse condition for
and generalization ability of the proposed method, the data the motors, the accuracies are higher than 92%. These results
set of the two motors were interchanged for the training and demonstrate the generalization ability of the proposed method
validation stages of the classifiers. In the next section, the for bearing fault detection of line connected TIMs.
results of these tests are presented. Next, the performance of the proposed system is verified
in the case of motors fed by frequency inverters. Table XI
C. Results for Interchanging Motor Data in the Classification presents these results, where the same levels of frequency
In this section, classification tests are presented where data variations, coupled load torque, and abrasive wear severities
sets from motors with different power ratings are interchanged, of Section IV-B were used. Also, the results were separated
that is, the classifiers training stage is performed with the data according to the type of frequency inverters as previously
from one motor and the validation uses data from the other described.
motor. Specifically, the data from Motor 1 is used for training Again, based on Table XI, the proposed method was able
and from Motor 2 for validation. Moreover, in Sections IV-A to detect bearing faults even for data from a motor that was
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
to load variations and to power quality issues, such as power [4] S. Singh and N. Kumar, “Detection of bearing faults in mechanical
supply voltage unbalance, that are common in an industrial systems using stator current monitoring,” IEEE Trans. Ind. Informat.,
vol. 13, no. 3, pp. 1341–1349, Jun. 2017.
environment.
[5] V. C. M. N. Leite, J. G. B. da Silva, G. F. C. Veloso, L. E. B. da Silva,
V. C ONCLUSION G. Lambert-Torres, E. L. Bonaldi, and L. E. d. L. de Oliveira, “Detection
This study presented an alternative approach for the fault of localized bearing faults in induction machines by spectral kurtosis
diagnosis of abrasive wear in the bearings of TIMs, regard- and envelope analysis of stator current,” IEEE Trans. Ind. Electron.,
vol. 62, no. 3, pp. 1855–1865, March 2015.
less of the type of power supply (line connected or inverter
[6] P. A. Delgado-Arredondo, D. Morinigo-Sotelo, R. A. Osornio-Rios,
fed), which is based on mutual information and intelligent J. G. Avina-Cervantes, H. Rostro-Gonzalez, and R. J. Romero-Troncoso,
systems. The experimental tests were performed using a data “Methodology for fault detection in induction motors via sound and
set consisting of several current signals from TIMs subject to vibration signals,” Mech. Syst. Signal Pr., vol. 83, pp. 568 – 589, 2017.
different operation conditions, such as variation of the coupled [7] S. E. Pandarakone, Y. Mizuno, and H. Nakamura, “Distinct fault analy-
load, voltage unbalance between the phases, variation of the sis of induction motor bearing using frequency spectrum determination
supply frequency, and also abrasive wear of the bearings, all of and support vector machine,” IEEE Trans. Ind. Appl., vol. 53, no. 3,
pp. 3049–3056, May 2017.
which are possible situations to be encountered in an industrial
environment. [8] A. Bellini, F. Filippetti, C. Tassoni, and G. Capolino, “Advances
in diagnostic techniques for induction machines,” IEEE Trans. Ind.
In the feature extraction step, the MI was used as a measure Electron., vol. 55, no. 12, pp. 4109–4126, Dec. 2008.
of association between the current signals of phases A and [9] N. R. Devi, D. Siva Sarma, and P. Ramana Rao, “Detection of
B, in order to extract fault signatures and, therefore, diagnose stator incipient faults and identification of faulty phase in three-phase
the real situation of the machine. Then, an evaluation of the induction motor - simulation and experimental verification,” IET Electr.
pattern classifier that suited this type of data was performed. Power App., vol. 9, no. 8, pp. 540–548, 2015.
The following classifiers were considered: MLP, k-NN, and [10] P. Konar and P. Chattopadhyay, “Multi-class fault diagnosis of induction
SVM. motor using Hilbert and wavelet transform,” Appl. Soft Comp., vol. 30,
pp. 341–352, May 2015.
The results have shown that the MLP network is the
most suitable classifier for our proposed approach using MI [11] F. B. Batista, P. C. M. L. Filho, R. Pederiva, and V. A. D. Silva,
“An empirical demodulation for electrical fault detection in induction
features for detection of TIM bearing faults. This was obtained motors,” IEEE Trans. Instrum. Meas., vol. 65, no. 3, pp. 559–569, Mar.
considering several severities of abrasive wear in the bearings, 2016.
regardless of operation conditions and type of power supply: [12] F. Immovilli, C. Bianchini, M. Cocconcelli, A. Bellini, and R. Rubini,
line connected or fed by frequency inverters. “Bearing fault model for induction motor with externally induced
The next steps of this work aim at improving the proposed vibration,” IEEE Trans. Ind. Electron., vol. 60, no. 8, pp. 3408–3418,
methodology for the identification of other types of induction Aug 2013.
motor faults. Also, future works intend to optimize the com- [13] X. Jin, M. Zhao, T. W. S. Chow, and M. Pecht, “Motor bearing fault
diagnosis using trace ratio linear discriminant analysis,” IEEE Trans.
putational tools, such as finding the ideal number of neurons Ind. Electron., vol. 61, no. 5, pp. 2441–2451, May 2014.
of the MLP intermediate layers in order to obtain the most
[14] M. Zhao, X. Jin, Z. Zhang, and B. Li, “Fault diagnosis of rolling
efficient system for bearing faults diagnosis. The search for element bearings via discriminative subspace learning: Visualization and
high performance classifiers in failure detection applications classification,” Expert Syst. Appl., vol. 41, no. 7, pp. 3391–3401, 2014.
is an active field of research, where the trade-off of accuracy, [15] P. A. Estévez, M. Tesmer, C. A. Perez, and J. M. Zurada, “Normal-
speed, and robustness plays a key role. Detailed information on ized mutual information feature selection,” IEEE Trans. Neural Netw.,
the comparison of classifiers for TIM fault analysis is presented vol. 20, no. 2, pp. 189–201, 2009.
in [43]. [16] G. Herman, B. Zhang, Y. Wang, G. Ye, and F. Chen, “Mutual
Still, another goal is the real-time implementation in embed- information-based method for selecting informative feature sets,” Pat-
tern Recognit., vol. 46, no. 12, pp. 3315–3327, 2013.
ded hardware of the proposed fault detection system. However,
due to the high computational effort of the MI estimation, [17] R. J. Romero-Troncoso, R. Saucedo-Gallaga, E. Cabal-Yepez,
A. Garcia-Perez, R. A. Osornio-Rios, R. Alvarez-Salas, H. Miranda-
which presents intricate statistical calculations, it is important Vidales, and N. Huber, “FPGA-based online detection of multiple
to consider the limitations imposed by the embedded system combined faults in induction motors through information entropy and
memory. Thus, the feasibility of this real-time fault diagnosis fuzzy inference,” IEEE Trans. Ind. Electron., vol. 58, no. 11, pp. 5263–
system is also the focus of future studies. 5270, Nov. 2011.
[18] J. Faiz, B. Ebrahimi, B. Akin, and H. Toliyat, “Dynamic analysis of
R EFERENCES mixed eccentricity signatures at various operating points and scrutiny
[1] L. Frosini, C. Harlişca, and L. Szabó, “Induction machine bearing of related indices for induction motors,” IET Electr. Power App., vol. 4,
fault detection by means of statistical processing of the stray flux no. 1, pp. 1–16, 2010.
measurement,” IEEE Trans. Ind. Electron., vol. 62, no. 3, pp. 1846– [19] U. Melia, M. Guaita, M. Vallverdú, C. Embid, I. Vilaseca, M. Salamero,
1854, March 2015. and J. Santamaria, “Mutual information measures applied to EEG
[2] Y. Trachi, E. Elbouchikhi, V. Choqueuse, and M. E. H. Benbouzid, signals for sleepiness characterization,” Med. Eng. Phys., vol. 37, no. 3,
“Induction machines fault detection based on subspace spectral estima- pp. 297–308, 2015.
tion,” IEEE Trans. Ind. Electron., vol. 63, no. 9, pp. 5641–5651, Sept [20] M. Barzegaran, A. Mazloomzadeh, and O. A. Mohammed, “Fault
2016. diagnosis of the asynchronous machines through magnetic signature
[3] A. H. Boudinar, N. Benouzza, A. Bendiabdellah, and M. E. A. Khodja, analysis using finite-element method and neural networks,” IEEE Trans.
“Induction motor bearing fault analysis using a root-music method,” Energy Convers., vol. 28, no. 4, pp. 1064–1071, Dec. 2013.
IEEE Trans. Ind. Appl., vol. 52, no. 5, pp. 3851–3860, Sept 2016. [21] M. D. Prieto, G. Cirrincione, A. G. Espinosa, J. A. Ortega, and
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2939678, IEEE
Transactions on Industrial Informatics
H. Henao, “Bearing fault detection by a novel condition-monitoring [41] F. Dalvand, A. Kalantar, and M. S. Safizadeh, “A novel bearing con-
scheme based on statistical-time features and neural networks,” IEEE dition monitoring method in induction motors based on instantaneous
Trans. Ind. Electron., vol. 60, no. 8, pp. 3398–3407, 2013. frequency of motor voltage,” IEEE Trans. Ind. Electron., vol. 63, no. 1,
[22] J. Seshadrinath, B. Singh, and B. K. Panigrahi, “Investigation of pp. 364–376, Jan 2016.
vibration signatures for multiple fault diagnosis in variable frequency [42] M. J. Picazo-Ródenas, J. Antonino-Daviu, V. Climente-Alarcon,
drives using complex wavelets,” IEEE Trans. Power Electron., vol. 29, R. Royo-Pastor, and A. Mota-Villar, “Combination of noninvasive
no. 2, pp. 936–945, Feb. 2014. approaches for general assessment of induction motors,” IEEE Trans.
[23] Y. Yagami, C. Araki, Y. Mizuno, and H. Nakamura, “Turn-to-turn Ind. Appl., vol. 51, no. 3, pp. 2172–2180, May 2015.
insulation failure diagnosis of stator winding of low voltage induction [43] R. Liu, B. Yang, E. Zio, and X. Chen, “Artificial intelligence for fault
motor with the aid of support vector machine,” IEEE Trans. Dielectr. diagnosis of rotating machinery: A review,” Mech. Syst. Signal Pr., vol.
Electr. Insul., vol. 22, no. 6, pp. 3099–3106, 2015. 108, pp. 33–47, 2018.
[24] M. Seera and C. P. Lim, “Online motor fault detection and diagnosis
using a hybrid FMM-CART model,” IEEE Trans. Neural Netw., vol. 25, Gustavo Henrique Bazan was born in Brazil,
no. 4, pp. 806–812, 2014. in 1989. He received the B.S. and M.Sc. degree
[25] L. A. Garcı́a-Escudero, O. Duque-Perez, M. Fernandez-Temprano, and in Electrical Engineering from the Federal
D. Morinigo-Sotelo, “Robust detection of incipient faults in vsi-fed Technological University of Paraná, Brazil, in
induction motors using quality control charts,” IEEE Transactions on 2013 and 2016, respectively. Since 2016, he is
Industry Applications, vol. 53, no. 3, pp. 3076–3085, May 2017. an Assistant Professor with the Federal Institute
[26] T. Cover and J. Thomas, Elements of Information Theory, 2nd ed. New of Paraná, Jacarezinho, Brazil. Currently, he is
York: Wiley-Interscience., 2006. pursuing Ph.D. degree in the Federal Technological
University of Paraná. His research interests are
[27] J. Nichols, “Examining structural dynamics using information flow,”
within the fields of signal processing, pattern
Probabilist. Eng. Mech., vol. 21, no. 4, pp. 420–433, 2006.
recognition, electrical machinery, and expert
[28] J. F. Alonso, M. A. Mananas, D. Hoyer, Z. L. Topor, and E. N. Bruce, systems.
“Evaluation of respiratory muscles activity by means of cross mutual
information function at different levels of ventilatory effort,” IEEE
Trans. Bio-Med. Eng., vol. 54, no. 9, pp. 1573–1582, Sep. 2007. Paulo Rogério Scalassara was born in Brazil, in
[29] S. Haykin, Neural Networks and Learning Machines, 3rd ed. New 1980. He received his B.S. and M.S. degrees in
Jersey: Pearson, 2008. Electrical Engineering from the State University of
[30] I. N. Silva, D. H. Spatti, R. A. Flauzino, L. H. B. Liboni, and S. F. R. Londrina, Brazil, in 2003 and 2005, respectively.
Alves, Artificial Neural Networks: A Practical Course, 2nd ed. New He obtained his Ph.D. from the University of São
York, USA: Springer, 2017. Paulo in 2009. Since 2010, he has been with the
Department of Electrical Engineering of the Federal
[31] C. Bishop, Pattern Recognition and Machine Learning (Information University of Technology – Paraná, in Cornélio
Science and Statistics). Secaucus, NJ, USA: Springer-Verlag New Procópio, as an Associate Professor. In 2016, he
York, Inc., 2006. was a Postdoctoral Researcher at the Computational
[32] J. Platt, Fast training of support vector machines using sequential NeuroEngineering Laboratory of the University
minimal optimization, in: B. Schoelkopf, C. Burges, A. Smola (Eds.), of Florida, Gainesville, USA. His interests are in information theoretical
Advances in Kernel Methods – Support Vector Learning. Boston, MA, learning, signal processing, machine learning, and embedded systems.
USA: MIT Press, 1998.
[33] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, and I. H.
Witten, “The WEKA data mining software: An update,” ACM SIGKDD Wagner Endo was born in Brazil, in 1982. He
Explor. Newsl., vol. 11, no. 1, pp. 10–18, 2009. received the B.S. degree in Electrical Engineering
in 2004, the M.Sc. degree in Electrical Engineering
[34] P. Kankar, S. C. Sharma, and S. Harsha, “Fault diagnosis of ball bearings in 2006 both from the State University of Londrina,
using machine learning methods,” Expert Syst. Appl., vol. 38, no. 3, pp. Brazil and the Ph.D. degree in Electrical Engineering
1876 – 1886, 2011. from the University of São Paulo (USP), Brazil, in
[35] D. Pandya, S. Upadhyay, and S. Harsha, “Fault diagnosis of rolling 2014. Currently, he is an Associate Professor with
element bearing with intrinsic mode function of acoustic emission data the Federal Technological University of Paraná,
using apf-knn,” Expert Syst. Appl., vol. 40, no. 10, pp. 4137 – 4145, Cornélio Procópio, Brazil. His research interests
2013. are within the fields of signal processing and
[36] P. Konar, J. Sil, and P. Chattopadhyay, “Knowledge extraction using automation of systems.
data mining for multi-class fault diagnosis of induction motor,” Neuro-
computing, vol. 166, pp. 14 – 25, 2015.
Alessandro Goedtel was born in Brazil, in 1972. He
[37] E. Togban and D. Ziou, “Classification using hierarchical mixture of dis- received the B.S. degree in Electrical Engineering
criminative learners: How to achieve high scores with few resources?” from the Federal University of Rio Grande do
Expert Syst. Appl., vol. 96, pp. 14 – 24, 2018. Sul, Brazil, in 1997, the M.Sc. degree in Industrial
[38] J. R. Landis and G. G. Koch, “The measurement of observer agreement Engineering from the São Paulo State University
for categorical data,” Biometrics, vol. 33, no. 1, pp. 159–174, 1977. (UNESP), Brazil, in 2003, and the Ph.D. degree in
[39] I. Martin-Diaz, D. Morinigo-Sotelo, O. Duque-Perez, P. A. Arredondo- Electrical Engineering from the University of São
Delgado, D. Camarena-Martinez, and R. J. Romero-Troncoso, “Analysis Paulo (USP), Brazil, in 2007. Currently, he is an
of various inverters feeding induction motors with incipient rotor fault Associate Professor with the Federal Technological
using high-resolution spectral analysis,” Electr. Pow. Syst. Res., vol. 152, University of Paraná, Cornélio Procópio, Brazil. His
no. Supplement C, pp. 18 – 26, 2017. research interests are within the fields of electrical
machinery, intelligent systems, and power electronics.
[40] M. G. S. and V. Hegde, “Application of mems accelerometer for
detection and diagnosis of multiple faults in the roller element bearings
of three phase induction motor,” IEEE Sensors Journal, vol. 16, no. 1,
pp. 145–152, Jan 2016.
1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.