Gradual Fault Condition Detection in the Outer Race

of Induction Motor Hybrid Bearings Based on Stray

Flux and LDA-FFNN Approaches
J. Cureño-Osornio, J.J. Saucedo-Dorantes, D. A. Elvira-Ortiz, A. Y. Jaen-Cuellar, I.
Zamudio-Ramirez, J.A. Antonino-Daviu and R.A. Osornio-Rios

Abstract –. The electric motors are yet important elements at if are not early detected, can affect their performance and,
industry since they represent the major energy consumption even worse, cause catastrophic damage. In this sense,
around the world. Any improvement in its operation could yield according to [4] the main problems in electric motors are, for
significant energy and costs savings. Among the elements that instance, short-circuit faults in windings, wrong connection of
compose an induction motor the rolling bearings are one of the the stator winding, broken rotor bars, dynamic and/or static
most analyzed parts because they suffer directly the motor loads
eccentricities, rotor field winding breakage, rolling bearing
and speeds changes. There exist different materials for the rolling
bearings, and the lifespan of this elements depends on an faults, etc. From these faults, those related with rolling
adequate maintenance. Therefore, this work presents an bearings and broken rotor bars use to appear with higher
approach for detecting the gradual fault condition in the outer frequency than the other problems. Additionally, it has been
race of an induction motor hybrid bearings based on the stray determined, through several studies, that between the 40% and
flux signals acquired around the motor case. These signals are the 50% of the faults in induction motors are directly related
processed by extracting statistical and non-statistical features for with rolling bearings [5].
generating a high dimensional matrix, which can be reduced by The faults that usually affect the rolling bearings can be
Linear Discriminant Analysis (LDA) optimizing the produced by short-circuits in the stator windings, flowing of
identifications of the classes detected. Then a simple topology of
electric current through the bearing (electric erosion), broken
a Feed Forward Neural Network (FFNN) is in charge for making
the fault conditions classification. The obtained results are rotor bars, fracturing or fissuring of bearing balls, wearing in
compared with other signals like current and vibration to show the outer race, excessive loads, excessive speeds, incorrect
the reliability of the stray flux in the fault analysis field. lubrication, external contaminants in the lubricant, and the
Index Terms—Induction motors; Fault diagnosis; Hybrid materials used in them [6 - 7]. In addition, there exist different
bearing; LDA, FFNN; Stray flux. materials from which the rolling bearings in induction motors
are made of, and they can be classified in three main groups:
I. INTRODUCTION metallic, ceramic, and hybrid. The metal rolling bearings
Nowadays, it is very palpable the importance of induction usually have major presence in the motors at industry because
motors at industry, for that reason, reducing to a minimal the of their low cost, however, their components get fatigued very
affectations produced by faults has been, and still is, a priority fast. Thus, a necessity of looking for strong materials with
topic of investigations. The electric energy is transformed by long lifespan to replace them has emerged. By its part, the
motors into mechanical energy and is exploited as motion ceramic rolling bearings are employed in very demanding
interchange in the processes; thus, any improvement in the processes where a high resistance to wear is required, and this
efficiency of the motors will be directly reflected as energy is possible thanks to their thermal, mechanical, and
savings and, hence, as production costs savings [1]. Therefore, tribological properties [8 - 10]. Through modelling, it has been
several investigations have addressed the study of fault demonstrated that hybrid bearings, composed by rolling
conditions in induction motors such as mechanical, thermal, elements of silicon nitride (Si3N4) and steel rings, have better
electrical, and environmental, generated by the combination performance under exigent conditions, like low lubrication or
of inherent factors to the motor design and its operation. contaminated environments; reason for which their use is
Additionally, there are very important components inside the every time most common. Although hybrid rolling bearings
motor that need to be considered in the study of fault have shown exceptional properties, they are still prone to fail.
conditions such as terminal boxes, fans, support caps, and However, there are not enough works that focus to study these
bearings, that typically present anomalies [2 - 3]. According kinds of bearings. Some works tackle the detection of faults in
to this, into the induction motors appear different faults that, hybrid bearings employing vibration signals, that are acquired
through an accelerometer, and process them by the fast Fourier

Vera s/n, 46022, Valencia, SPAIN (e-mails: [email protected],

This work was supported by the Spanish ‘Ministerio de Ciencia [email protected]).
Innovación y Universidades’ and FEDER program in the framework of the J. Cureno-Osornio, J. J. Saucedo-Dorantes, D. A. Elvira-Ortiz, A. Y. Jaen-
‘Proyectos de I+D de Generación de Conocimiento del Programa Estatal de Cuellar, and R. A. Osornio-Rios, are with the HSPdigital CA-Mecatronica
Generación de Conocimiento y Fortalecimiento Científico y Tecnológico del Engineering Faculty, Autonomous University of Queretaro, San Juan del Rio
Sistema de I+D+i, Subprograma Estatal de Generación de Conocimiento’ 76806, Mexico. (e-mail: [email protected],
(ref: PGC2018-095747-B-I00) and by the scholarships 652815 by the [email protected], [email protected], [email protected],
CONACyT. [email protected])
I. Zamudio-Ramirez and J. A. Antonino-Daviu are with the Instituto .
Tecnologico de la Energia, Universitat Politècnica de València, Camino de
transform (FFT) to compare their efficiency versus metallic characteristics and for carrying out the data processing. They
bearings. Other works study the properties of faults in hybrid are listed in Table 1 from (1) to (20) [18].
bearings by introducing silicon particles in the lubricant, and
by using a capacitive sensor are detected the damages Table 1. Statistical and non-statistical indicators
triggered by such faults [11 - 12]. An important characteristic considered for processing.
in the studies of faults in rolling bearings is that, generally, the
1
Indicator Mathematical expression
= ∙
magnitudes of the variables used for detecting the anomalies Mean (1)
are very low, which complicates to define gradualism, from
incipient to severe level. Thanks to the evolution of sensors Maximum value = max (2)
and processing techniques, the stray flux has become, once
1
= ∙
again, an important signal for detecting faults in induction
Root mean square (RMS) (3)
motors [13]. For example, the stray flux has been used for
analyzing the faults caused by inter-turn short circuits by 1
applying the FFT, this way the related frequencies to the fault Square root mean (SRM) = ∙ | | (4)

1
are found. [14]. Also, in [15] the rotor faults in motors are
analyzed through the stray flux signature and the finite Standard deviation = ∙ − (5)

1
element method. It is evident that stray flux still represents a
potential solution for detecting fault conditions in induction Variance = ∙ − (6)
motors.
=
1
∙∑ | |
From the state of the art, few works tackle the analysis of RMS Shape factor (7)
faults in hybrid rolling bearings and their graduality.
=
1
Therefore, this work proposes a methodology based on stray
∙∑ | |
SRM Shape factor (8)
flux for detecting fault conditions graduality in the outer race
=
of induction motor hybrid bearings. For the study, there are Crest factor (9)
considered three conditions to know: healthy state, 3 mm hole,
and 5 mm hole in the outer race of the bearings. Also, the Latitude factor ! = (10)

=
experimental tests consider a variable frequency drive (VFD)
1
∙∑ | |
feeding the induction motor with four different operation Impulse factor (11)

∑" # $
frequencies. The methodology starts by capturing the stray
flux signals from the motor through a proprietary 3-axial Skewness = (12)
∑" $
sensor board. These signals firstly pass through a filtering
=
#
stage, to later calculate statistical and non-statistical indicators Kurtosis (13)
from time windows. Finally, by fusing the LDA and the ∑" $
=
#
FFNN, the classification of fault conditions graduality in the Fifth moment (14)
∑" $
=
#
bearings is achieved. As additional contribution, the
Sixth moment (15)
comparisons of the results using stray flux versus the results
log (/*
=
obtained with the same approach by using current signals, and
log +/*
Katz’s fractal dimension
(16)
posteriorly by using vibration signals are carried out. The (KFD)
.
= (, -
comparisons indicate that stray flux signals processed through Higuchis’s fractal
(17)
dimension (HFD) ,
∑8 7 5 6
the proposed methodology provides better results than
= 10 ∙ log 4 9
;
transform energy /012
Discrete wavelets
∑8 7 + 6
currents or vibrations. (18)
9 :
:
II. THEORETICAL BACKGROUND Wavelet entropy <12 =− ln >? (19)
8 ,
#
A. Statistical and non-statistical indicators Shannon entropy ! =− > log "> $ (20)
!
When analyzing data, it is very practical to use some kinds
of meters for describing the data set. These meters, known as
Where x is the input signal; N the total samples in the time
indicators (or features), can help for providing distinguished
characteristics that allow to define similitude, or differences, window; xi the i-th observation for i=1,2,3,…,N; L the
with other sets [16]. This way, the statistical and non- Euclidian distance between successive samples of time
domain; a the average of the Euclidian distances; d the
maximum Euclidian distance between of the samples; 5 the
statistical indicators are powerful, and very useful tools in the
stray flux measure; +: the wavelet signal containing the
analysis of faults, since they can provide specific information
about the signal’s behavior related to the operation condition,
majority of the fault frequency component; p(xi) the
healthy or faulty. Through the years a wide variety of
statistical functions in the time-domain like mean, root mean probability of the sample xi.
squared, variance, standard deviation, among others, are often B. Linear Discriminant Analysis (LDA).
used to identify the difference between signals; but also, more The machine learning techniques, such as the case of the
complex indicators such as asymmetries and the kurtosis of a LDA, have allowed the pattern recognition in several
signal can be employed [17]. Although there are many and applications, since they can find out traits for comparing sets
diverse, for this proposed approach the statistical and non- among them, to establish similitude or differences. Indeed,
statistical indicators were considered because of the signal
most of the time the LDA is usually defined as a supervised output neurons Σ|G first the weighted sum is carried out (22),
algorithm for features extracting analysis, but this tool also for later apply the SoftMax function (23) [24].
works as a linear classifier. The LDA is used for maximizing
Σ|G = H8 = I*Jℎ L ,8 ∙ N
the existing relationship of the variance with the distinct
(21)
groups, or classes, as well as the variance into the same class.
P
Σ|G = = L8,O ∙ H8
The goal is to achieve the greatest separation among the
O (22)
8
features of certain sets for every class; being the final result of
QO = RSTIE*
this technique the projection of reduced data sets formed by
O (23)
dimensional vectors into a smaller subspace through linear
mapping [19]. A detailed description of the implementation of
Where Ii, Jj, Tt, and Ot are the input, hidden, output, and
LDA is described in [20] showing the main steps as follows:
post-softMax neurons for i=1,2,..,N, j=1,2..,M, and t=1,2..,K;
Assuming the inputs as @ ∈ R:xB , and the outputs as C ∈
N, M, and K are the number of input, hidden and output
:D.
neurons; finally, Wi,j and Wj,t are the weights of the input and
R . Where nxd, is the matrix dimensionality of the inputs, hidden neuros, respectively.
and nxk is the matrix dimensionality of the outputs.
III. METHODOLOGY AND EXPERIMENTATION
• Step 1: Construct two matrices of dispersion according
to the inputs (among classes and inside the class). The Fig. 2 presents the general block diagram of the
proposed scheme for detecting gradual fault conditions in the
• Step 2: Compute the Eigenvalues and the Eigenvectors
outer race of an induction motor hybrid bearings by using stray
from the dispersion matrices.
flux generated around the motor case. The methodology
• Step 3: Make the classification of the Eigenvectors in
considers three main blocks, to know: the electromechanical
• Step 4: Formulate the matrix EBD. by means of the
descending order.
process, the data acquisition system, and the LDA-FFNN
methodology.
Eigenvalues.
• Step 5: Transform the matrix of inputs X into the matrix
of outputs Y through the matrix M.

C. Feed Forward Neural Network (FFNN)

The artificial neural networks (ANN) are data processing
topologies based on neuron structures like those of human
beings, achieving making decisions considering a previous
analysis. The classifications made by the ANN are depending
on the topology adopted [21]. There exist different forms for
modelling an ANN that rely on the final purpose and the
application requirements, however, many of the researchers
choose to use the FFNN because of its simplicity [22]. The
training of this network is usually through the back
propagation (BP) algorithm, which consist of comparing the
output values with the desired values to actualize the output
layer, as well as the weights and the activation functions [23].
Therefore, the structure of the FFNN implemented in this
work consist of an input layer with two neurons, one hidden
layer with two neurons, and one output layer with three
neurons, as observed in Fig. 1. Fig. 2. General block diagram of the proposed methodology.

The proposed strategy begins with the electromechanical

process block, which consists of a simple kinematic chain. In
the chain a 1HP induction motor, model WEG, is connected
to an automotive alternator, as the imposed load; both are
connected through a transmission band placed on the pulleys
mounted in the motor and load shafts. Inside the induction
motor, there were induced artificial faults, equivalent to the
damage in the outer race, in the hybrid rolling bearings by
drilling them with holes of 3 mm and 5 mm diameters. This
way, there were three operating conditions for the study of
graduality: the healthy state (HLT), 3mm, and 5mm.
Fig. 1. Diagram of the FFNN topology defined for the proposed approach.
Moreover, a VFD feeds the motor with four different
The activation function Σ|G is defined for the neurons of
frequencies (5 Hz, 15 Hz, 50 Hz, 60 Hz) to have different
motor speeds and variability of the signals for further
the hidden layer as the hyperbolic tangent (21), and for the
processing. Additionally, these frequencies allow observing considered the 4 different frequencies of the VFD (5 Hz, 15
the fault’s behavior in frequencies near to the commercial and Hz, 50 Hz, 60 Hz). In summary, there exists a total of 24
nominal operation values (50 Hz and 60 HZ), and at low windows x 14 signals x 4 frequencies = 1344 samples per fault
frequencies into the operational range of the motor. Since the condition (HLT, 3 mm, and 5 mm).
VFD induces into the acquisition system non-desired high These samples integrate the high dimensional matrix of
frequencies, a filtering of the signals is performed in hardware features which is then reduced through the LDA technique,
by using active low-pass filters with a cut-off frequency which yields a reduced 2-dimensional representation of the
defined at 1500 Hz. data. The LDA optimizes the separation between the classes
The next block is the data acquisition system that is identified, therefore a simple structure of the FFNN like that
composed by the following proprietary boards: data of Fig. 1 and implemented through (21) to (23) provides the
acquisition system based on FPGA, the 3-axial stray flux classification of the fault conditions.
sensor, the current clamps, and the vibration sensor. The stray
flux sensor board integrates three individual hall-effect IV. RESULTS AND DISCUSSION
sensors from ALLEGROTM microsystems brand model In this section the obtained results of the proposed approach
A1325. A low pass filtering stage is applied to the stray flux
based on stray flux for detecting graduality in fault conditions
signals for eliminating the environmental noise. The system
in the outer race of an induction motor hybrid bearings are
acquires simultaneously the signals coming from the motor
presented. The results are described in two parts: results of the
corresponding to the stray flux from directions axial, radial,
and combined; the vibration signals in axis Ax, Ay and Az; and methodology through stray flux signals and comparative of
finally, the current signals from the three phases. The the stray flux results versus the results of vibrations signals
sampling frequencies in the data acquisition system are 3 kHz and ac current signal.
for the stray flux and for vibration signals, and 6 kHz for the A. Results of the methodology through stray flux signals
ac current signal, respectively. These values were defined
For the sake of simplicity and illustrative purposes, a plot
according to the restrictions in the sampling frequencies of the
of a single stray flux signal used for processing is presented.
sensors. In this work the proposed approach is developed
Thus, the plot of Fig. 4(a) depicts the time response of a stray
specifically for stray flux signals, therefore, the ac current
flux signal, and the Fig. 4(b) presents a zoom of the signal
signals and the vibration signals are used for comparative
showing six periods in detail. This figure is randomly selected
purposes.
from the acquired data set, in this case the stray flux signal is
The last block of the proposed strategy is the LDA-FFNN
from directions axial+radial.
methodology. In this block 5 sub steps are appreciated: 1)
signals segmentation, 2) calculation of statistical and non-
statistical indicators, 3) definition of the high-dimensional
matrix of features, 4) features reduction through LDA, and 5)
classification of fault conditions through FFNN. The Fig. 3
illustrates the experimental design for the data processing of
the five sub steps in detail. (a)
Amplitude (V)

(b)

Fig. 3. Block diagram of the experimental design.

From figure, the data are segmented in windowed signals of

1s time. Next, per window there are calculated the 20 Figure 4. Time response plots, in (a) stray flux signal from direction
statistical and non-statistical indicators according to (1) to axial+radial and (b) zoom of the signal showing six periods.
(20). Since the time windows overlap in 50%, a total of 24
windows are defined per signal acquired. Here, although the The procedure described in the block LDA-FFNN
sampling frequencies of measured signals are 3 kHz and 6kHz methodology over the stray flux signals from the induction
the extraction of features is not affected, since the number of motor yields as results the plots observed in Fig. 5. In one
windows is the same for all the signals. Also, per stray flux hand, the Fig. 5(a) depicts the generated decision regions over
direction (axial, radial, and combined) there must be the 2-dimensional representation obtained by means of LDA,
considered 14 acquired signals. Besides, there must be that posteriorly feeds the FFNN classifier. On the other hand,
as consequence, in Fig. 5(b) is observed the confusion matrix B. Comparative of the stray flux results versus the results
with the respective effectiveness percentages for classifying of vibrations signals and ac current signal
the three conditions studied on the hybrid bearings: HTL, 3 The proposed approach was applied over other measured
mm, and 5 mm. signals, such as vibrations and ac current, for comparing the
effectiveness of the obtained results in the graduality detection
of fault conditions. Therefore, the plots of Fig. 6(a) show the
decision regions obtained through of LDA from the combined
signals acquired in the axes Ax, Ay, and Az of the vibration
sensor. However, it is observed a moderate overlapping
among the three conditions detected in the hybrid bearings.
(a)

(a)

(b)
Confusion Matrix

384 0 16 96.0%
31.7% 0.0% 1.3% 4.0%
Output class

1 387 12 96.8%
0.1% 31.9% 1.0% 3.2%
Figure 5. Classification results of the proposed methodology obtained for
19 17 376 91.3% (b)
the stray flux signals considering direction of combined flux, and axial
1.6% 1.4% 31.0% 8.7%
flux; (a) Resulting decision regions modelled by the proposed FFNN
classifier over the 2-dimensional space of the LDA; (b) Effectiveness
95.0% 95.8% 93.1% 94.6%
percentage of the proposed FFNN-based method (confusion matrix).
5.0% 4.2% 6.9% 5.4%

As it can be observed, the Fig 5(a) shows a clear

delimitation of the decision regions generated by the LDA Target Class
technique when the high dimensional matrix of statistical and Figure 6. Classification results of the proposed methodology obtained for
vibration signals combined of the axes Ax, Ay and Az, (a) Resulting
non-statistical features is reduced. The LDA has the advantage decision regions modelled by the proposed NN-based classifier over the
not only for reducing the number of features required for the 2-dimensional space; (b) Effectiveness percentage of the proposed FFNN-
2-dimensional representation of detected classes, but also for based method (confusion matrix).
optimizing the separation among them by eliminating
redundant information found in the features. This is very The previous results are verified through the confusion
advantageous because just a simple structure of the FFNN is matrix showed in Fig. 6(b). From this matrix, it can be
required, at the classification stage of the fault conditions, to observed that by using the combined signals of the three axes
provide excellent results. The abovementioned is validated of the vibration sensor the categorization of the three fault
through the confusion matrix of Fig. 5(b) that shows how a conditions is achieved with an overall performance of 96.4%.
100% of effectiveness is achieved for each condition in the Although the global accuracy reached is pretty good, it
categorization. This accuracy in the classification stage is also remains under the accuracy of 100% reached by using the
achieved thanks to the effectiveness of the LDA in selecting stray flux signals. In detail, the worst behavior in the
the significant features related to the fault conditions, avoiding classifications is for the 5 mm fault condition that is confused
non-valuable and redundant information. Additionally, it is with the conditions HLT and 3 mm in a 6.9%.
clear how a well separation between the classes was achieved
during reduction of matrix of features facilitating the At last, but not least, the Fig. 7(a) shows the results obtained
categorization task of the FFNN without overfitting. It worth for detecting graduality of the three conditions in the outer
noticing that the proposed methodology works for analyzing, race of an induction motor hybrid bearings, but in this case by
simultaneously, all the variations of frequency defined for the using the signals measured from the ac current sensor. Like
experimental trials (5 Hz, 15 Hz, 50 Hz, and 60 Hz) and not the motor requires a three-phase connection for its normal
for a particular frequency by separate, as other approaches do. operation, then the signals to be analyzed are acquired from
whatever the three lines connected to the motor. In this case
the electric line L2 is considered for the study. Thus, from coming from directions axial, radial, and combined allows
figure is observed a severe overlapping among the analyzed fusing the signals in order to obtain better results just than
classes. These data are verified through the confusion matrix simply use a single sensor. It is clear through this proposed
of Fig. 7(b) that indicates a general performance in the methodology that faults in the outer race of hybrid bearings
classification of the fault conditions of 62.8%. From all the can be gradually detected according to the severity of the
considered signals, the ac current presents the worst condition. It is important because the work demonstrates that
performance results since the 3mm and 5mm conditions are anomalies, damage, wearing, etc., present in hybrid bearings
misclassified in a 47.8% and 41.1%, respectively. into an induction motor can be observed clearly through the
stray flux generated around the motor case. In future work, it
would be interesting to explore the potential of using the stray
flux signals for detecting other kind of fault conditions, or for
detecting multiple faults in the same system.

VI. REFERENCES
[1] E. Martinez, M Torrent, (2009). Determinacion del rendimiento den
(a) motores de induccion trifasicos segun la norma IEC 60034-2-1, Tecnica
industrial 282.
[2] E. Harper, (2005). Transformers and induction motors. Editorial Limusa
Maquinas electricas, Mexico:Limusa ISBN:968-18-5993-6
[3] H. Bonnett, C. Soukup (2001). Cause and analysis of stator and rotor
failures in 3 phase squirrel cage induction motors. Division of Emerson
Eleectric, IEEE
[4] Díaz, D., Díaz, R., Amaya, M. C., & Palacios, J. (2010). Análisis del
cortocircuito entre espiras de un motor de inducción tipo jaula de ardilla
mediante la aplicación del método de elementos finitos (mef) Inter-turn
shortcircuit analysis in an induction machine by finit elements method.
In Rev. Fac. Ing. Univ. Antioquia N.° (Vol. 51).
[5] Lebaroud, A., & Clerc, G. (2008). Classification of Induction Machine
Faults by Optimal Time-Frequency Representations. IEEE
TRANSACTIONS ON INDUSTRIAL ELECTRONICS, 55(12), 4290–
(b) 4298. https://doi.org/10.1109/TIE.2008.2004666ï
[6] Raison, B., & Rostaing, G. (n.d.). Investigations of Algorithms for
Bearing Fault Detection in Induction Drives Seals Thrust bearing Oil
leakage Other.
[7] Zulma Yadira, M.-H., Carlos, P.-T., Julio, G.-S., & Maximiliano, V.-P.
(n.d.). Bearings in a Synchronous Machine by Processing Vibro-
Acoustic Signals Using Power Spectral Density. Novel Methodology of
Fault Diagnosis On.
Figure 7. Classification results of the proposed methodology obtained for
L2 ac current signals, (a) Resulting decision regions modelled by the [8] Manual SKF, (2019). Enero 10000 ES PUB BU/P1 17000/1 ES
proposed NN-based classifier over the 2-dimensional space; (b) [9] Lewinschal, L., “Bearings for high speed operation”, SKF Evolution,
Effectiveness percentage of the proposed FFNN-based method (confusion N.° 2 1994, págs. 22-26. http://evolution.skf.com/bearings-for-high-
matrix). speed-operations/
[10] Riley, F. L. (2000). Silicon Nitride and Related Materials.. Departament
of material, School of Process Enviromental Riley, F. L. (n.d.). Silicon
From previous results, it is demonstrated that stray flux Nitride and Related Materials.
signals provide better performance that vibration signals, even [11] OhtaH.KobayashiK.(1996) VIBRATIONS OF HYBRID CERAMIC
combined from the three axes, and the ac current signals. For BALL BEARINGS, Elsevier. Volumen 192, número 2 ,páginas 481-493
all the cases processed through the same methodology. This is [12] Dempsey, P. J., Certo, J. M., & Morales, W. (2005). Current status of
an indication of the potentiality for exploiting the stray flux hybrid bearing damage detection. Tribology Transactions, 48(3), 370–
376. https://doi.org/10.1080/05698190591008568
physical magnitudes in the field of monitoring and diagnosis
[13] Kedjar, B., Merkhouf, A. and Al-Haddad, K. Large Synchronous
systems. Machines Diagnosis Based on Air-Gap and Stray Fluxes - An Overview,
2020 International Conference on Electrical Machines (ICEM), 2020,
V. CONCLUSIONS pp. 1384-1389, https://doi.org/10.1109/ICEM49940.2020.9270868
[14] Ehya, H., Nysveen, A. and Nilssen, R. Pattern Recognition of Inter-Turn
In conclusion, the proposed approach based on stray flux Short Circuit Fault in Wound Field Synchronous Generator via Stray
signals for detecting graduality of fault conditions in the outer Flux Monitoring, 2020 International Conference on Electrical
race of induction motor hybrid bearings demonstrates its Machines (ICEM), 2020, pp. 2631-2636,
https://doi.org/10.1109/ICEM49940.2020.9270986
robustness and efficiency, by achieving an overall
[15] Panagiotou, P. A., Arvanitakis, I., Lophitis, N. and Gyftakis, K. N. FEM
performance in the fault conditions classification of 100%. Study of Induction Machines Suffering from Rotor Electrical Faults
The proposed methodology does not suffer from overfitting Using Stray Flux Signature Analysis, 2018 XIII International
because a clear and well-defined separation between the Conference on Electrical Machines (ICEM), 2018, pp. 1861-1867,
https://doi.org/10.1109/ICELMACH.2018.8506707
classes identified is observed in the decision regions generated
[16] Huerta-Rosales, J. R., Granados-Lieberman, D., Garcia-Perez, A.,
through the LDA, which reduces the arduousness of the Camarena-Martinez, D., Amezquita-Sanchez, J. P., & Valtierra-
classification task. The use of a proprietary board that Rodriguez, M. (2021). Short-circuited turn fault diagnosis in
integrates triaxial hall-effect sensor for detecting the stray flux transformers by using vibration signals, statistical time features, and
 support vector machines on fpga. Sensors, 21(11). National Researcher level 1 with the Mexican Council of Science and
https://doi.org/10.3390/s21113598 Technology, CONACYT, and is currently a Head Professor with the UAQ,
[17] Caesarendra, W., & Tjahjowidodo, T. (2017). A review of feature Queretaro, Mexico. His research interests include digital signal processing on
extraction methods in vibration-based condition monitoring and its FPGAs for applications in engineering, condition monitoring and fault
application for degradation trend estimation of low-speed slew bearing. diagnosis in electromechanical systems, fault detection algorithms, artificial
In Machines (Vol. 5, Issue 4). MDPI AG. intelligence and signal processing methods.
https://doi.org/10.3390/machines5040021
[18] Jaen-Cuellar, A. Y., Osornio-Ríos, R. A., Trejo-Hernández, M., David Alejandro Elvira-Ortiz (S’17, M’20) received his M.S. and Ph. D.
Zamudio-Ramírez, I., Díaz-Saldaña, G., Pacheco-Guerrero, J. P., & degrees in Mechatronics, both from the Autonomous University of Queretaro
Antonino-Daviu, J. A. (2021). System for tool-wear condition in 2015 and 2019, respectively. He works at the Autonomous University of
monitoring in cnc machines under variations of cutting parameter based Queretaro as a full-time professor, and he is accredited by the Mexican
on fusion stray flux-current processing. Sensors, 21(24). Council of Science and Technology (CONACYT) as National Researcher
https://doi.org/10.3390/s21248431 level 1. His current research interests include hardware signal processing,
[19] Haddad, R. Z., & Strangas, E. G. (2016). On the Accuracy of Fault fault detection and diagnostics of induction machines, power quality, and
Detection and Separation in Permanent Magnet Synchronous Machines smart grids.
Using MCSA/MVSA and LDA. In IEEE Transactions on Energy
Conversion (Vol. 31, Issue 3, pp. 924–934). Institute of Electrical and Arturo Yosimar Jaen-Cuellar received the M.Sc. degree in
Electronics Engineers Inc. https://doi.org/10.1109/TEC.2016.2558183 Instrumentation and Automatic Control in 2011, and the Ph.D. degree in
[20] Anowar, F., Sadaoui, S., & Selim, B. (2021). Conceptual and empirical Engineering from the Autonomous University of Queretaro, Queretaro,
comparison of dimensionality reduction algorithms (PCA, KPCA, LDA, Mexico, in 2014. He is a National Researcher level 1 with the Mexican
MDS, SVD, LLE, ISOMAP, LE, ICA, t-SNE). In Computer Science Council of Science and Technology, CONACYT. He is currently a Full-Time
Review (Vol. 40). Elsevier Ireland Ltd. Professor and Researcher at the Autonomous University of Queretaro,
https://doi.org/10.1016/j.cosrev.2021.100378 Queretaro, Mexico. His research interests include hardware signal processing
[21] Chayjan, R. A., & Esna-Ashari, M. (2010). Comparison between using field-programmable gate arrays, digital systems, process control,
artificial neural networks and mathematical models for estimating instrumentation, and the application of heuristic techniques to solve
equilibrium moisture content in raisin. In Agric Eng Int: CIGR Journal mechatronics problems.
(Vol. 12, Issue 1). http://www.cigrjournal.org
[22] Turan, N. G., Gümüşel, E. B., & Ozgonenel, O. (2013). Prediction of Israel Zamudio-Ramírez the M.S. degree in mechatronic from the
heavy metal removal by different liner materials from landfill leachate: Autonomous University of Queretaro, Mexico, in 2019. He is currently
Modeling of experimental results using artificial intelligence technique. working toward the Ph.D. degree in mechatronics at the Autonomous
The Scientific World Journal, 2013. University of Queretaro with the Department of Mechatronics, Mexico and in
https://doi.org/10.1155/2013/240158 electrical engineering at the Universitat Politècnica de València with the
Department of Electrical Engineering, Spain. His research interests include
[23] Camarena-Martinez, D., Valtierra-Rodriguez, M., Garcia-Perez, A., monitoring and diagnosis of electrical machines, embedded systems and
Osornio-Rios, R. A., & Romero-Troncoso, R. D. J. (2014). Empirical hardware signal processing for engineering applications on FPGA.
mode decomposition and neural networks on FPGA for fault diagnosis
in induction motors. The Scientific World Journal, 2014.
https://doi.org/10.1155/2014/908140 Jose Antonino-Daviu (S’04, M’08, SM’12) received his M.S. and Ph. D.
[24] Zhu, D., Lu, S., Wang, M., Lin, J. and Wang, Z. Efficient Precision- degrees in Electrical Engineering, both from the Universitat Politècnica de
Adjustable Architecture for Softmax Function in Deep Learning, in València, in 2000 and 2006, respectively. He also received his Bs. in Business
IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 67, Administration from Universitat de Valencia in 2012. He was working for
no. 12, pp. 3382-3386, Dec. 2020, IBM during 2 years, being involved in several international projects.
https://doi.org/10.1109/TCSII.2020.3002564 Currently, he is Full Professor in the Department of Electrical Engineering of
the mentioned University, where he develops his docent and research work.
He has been invited professor in Helsinki University of Technology (Finland)
VII. BIOGRAPHIES in 2005 and 2007, Michigan State University (USA) in 2010, Korea
University (Korea) in 2014 and Université Claude Bernard Lyon 1 (France)
Jonathan Cureño Osornio was born in Mexico, on September 8, 1998. He in 2015.
graduated in Electromechanical Engineering degree from the Engineering
Faculty at the Autonomous University of Queretaro, Mexico in 2021. He is Roque Alfredo Osornio-Rios (M’10, SM’21) received the Ph.D. degree in
currently pursuing an M.S.E. degree in Mechatronics at the Autonomous mechatronics from the Autonomous University of Queretaro, Queretaro,
University of Queretaro (UAQ). His research interests include monitoring Mexico, in 2007. He is a National Researcher level 3 with the Mexican
electromechanical machines with different conditions of damage, the Council of Science and Technology, CONACYT. He is currently a Head
processing of various signals for the monitoring of induction motors, and new Professor with the Department of Mechatronics, University of Queretaro. He
mathematical processing techniques. is an Advisor for more than 80 theses, and has coauthored more than 200
technical papers published in international journals and conferences. His
Juan Jose Saucedo-Dorantes (Member, IEEE), received the M. E. and Ph. research interests include hardware signal processing and mechatronics. Dr.
D degrees in mechatronics from the Autonomous University of Queretaro Osornio-Rios is a Fellow of the Mexican Academy of Engineering. He is part
(UAQ), Queretaro, Mexico, in 2014 and 2018, respectively. Since 2012, he of the editorial board of Journal and Scientific and Industrial Research.
has been doing research work at the HSPdigital research group. He is a