sensors

Article
The Edge Application of Machine Learning Techniques for
Fault Diagnosis in Electrical Machines
Javier de las Morenas 1, * , Francisco Moya-Fernández 2 and Julio Alberto López-Gómez 1

1 Mining and Industrial Engineering School of Almadén, University of Castilla-La Mancha,

13400 Almadén, Spain
2 Mantis Research Group, EIIA Toledo, University of Castilla-La Mancha, 45005 Toledo, Spain
* Correspondence: [email protected]

Abstract: The advent of digitization has brought about new technologies that enable advanced
condition monitoring and fault diagnosis under the Industry 4.0 paradigm. While vibration signal
analysis is a commonly used method for fault detection in literature, it often involves the use
of expensive equipment in difficult-to-reach locations. This paper presents a solution for fault
diagnosis of electrical machines by utilizing machine learning techniques on the edge, classifying
information coming from motor current signature analysis (MCSA) for broken rotor bar detection.
The paper covers the process of feature extraction, classification, and model training and testing
for three different machine learning methods using a public dataset to then export the results to
diagnose a different machine. An edge computing approach is adopted for the data acquisition,
signal processing and model implementation on an affordable platform, the Arduino. This makes it
accessible for small and medium-sized companies, albeit with the limitations of a resource-constrained
platform. The proposed solution has been tested on electrical machines in the Mining and Industrial
Engineering School of Almadén (UCLM) with positive results.

Keywords: fault diagnosis; edge computing; machine learning; motor current signature analysis

Citation: de las Morenas, J.;

Moya-Fernández, F.; López-Gómez,
1. Introduction
J.A. The Edge Application of Machine Society is currently deeply immersed in the fourth industrial revolution, where digi-
Learning Techniques for Fault talization has led to the emergence of new technologies and lines of research, such as the
Diagnosis in Electrical Machines. Internet of Things, Big Data, Smart technology, and Cloud Computing, all with the goal of
Sensors 2023, 23, 2649. https:// digitalizing information.
doi.org/10.3390/s23052649 The digitalization process is having a significant impact not only in industry but
Academic Editors: Salvatore also in other areas, such as smart cities and buildings [1], energy systems [2], pollution
Cavalieri and Nunzio Marco Torrisi monitoring [3], health tracking [4], and intelligent transportation systems [5]. However, its
full realization within industry has yet to be fully achieved, and its impact on small and
Received: 2 February 2023 medium-sized enterprises (SMEs) is considerably lower.
Revised: 22 February 2023
In Spain, Industry 4.0 is known as Connected Industry 4.0 [6]. In order to assist SMEs
Accepted: 26 February 2023
in the development of digitalization, the government has created the Digital Transformation
Published: 28 February 2023
Offices within red.es (a public entity attached to the Ministry of Economic Affairs and Digi-
tal Transformation) [7]. This is not insignificant, as SMEs make up 99.6% of total companies
in Spain according to data from the Social Welfare, as of December 2022, including 59.7% of
Copyright: © 2023 by the authors.
the total number of workers [8].
Licensee MDPI, Basel, Switzerland. The present paper is focused on one of the fundamental concepts of Industry 4.0,
This article is an open access article namely predictive maintenance, with a specific emphasis on fault diagnosis in electrical
distributed under the terms and machines. As per the International Energy Agency report [9], approximately 70% of all
conditions of the Creative Commons energy expended in industries in developed countries is consumed by electric motors.
Attribution (CC BY) license (https:// Therefore, there has been a significant increase in interest in predicting and detecting faults
creativecommons.org/licenses/by/ in these machines [10–13], and the advent of Industry 4.0 provides an additional boost for
4.0/). this trend [14].

Sensors 2023, 23, 2649. https://doi.org/10.3390/s23052649 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 2649 2 of 20

Induction motors are protected by devices such as relays, circuit breakers, etc., but the
performance of these protections is limited either to protecting the motor against an external
problem or to a significant level of motor deterioration when the fault is often irreversible.
On the other hand, the demand from users to increase the reliability of electric motors is
constantly growing due to the importance of many applications, advances in technology
and the need to obtain better operating results. Therefore, the ability to predict and detect
the onset of failures becomes crucial in applications where the complexity, criticality, cost of
shutdown or repair, and safety considerations do not permit unexpected motor failure [15].
This paper presents a solution for the fault diagnosis of electrical machines in an edge
computing approach. The novelty of this work is the constant condition monitoring of the
machines and the edge application of machine learning classifying methods based on a low-
cost and resource-constrained platform, with the objective of reducing maintenance costs
and equipment downtime through the detection of equipment faults, from the perspective
of making it accessible for SMEs. Another innovative aspect is the extrapolation of the
classification model. This model has been developed using a dataset from a specific
machine, and it is now being applied to other machines. This process requires the selection
of features that are independent of the absolute magnitudes of each motor. Novel features
have been proposed.
The subsequent sections of this paper are structured as follows: In Section 2, the
prevalent methods for fault diagnosis in electric induction motors are discussed, along with
a brief overview of Machine Learning techniques and Industrial Cyber-Physical Systems
(ICPS). Section 3 presents the components of the ICPS system responsible for fault diagnosis
and emphasizes their implementation on platforms with limited resources using an edge
computing approach, which enables local data processing. Section 4 presents the public
dataset used, the development of the classification models, the practical implementation of
the proposed solution, and the detailed analysis of the results. Lastly, Section 5 concludes
the paper by summarizing our findings and outlining potential future avenues for research.

2. State of the Art

Several diagnostic techniques have been established for three-phase induction motors.
These techniques are primarily based on the examination of the magnitudes that are
accessible at the motor terminals or those faults that can be detected through appropriate
equipment by performing measurements under normal operating conditions of the motor
or when the machine is out of service.

2.1. Out-of-Service Diagnostics

The techniques for diagnosing three-phase induction motors typically involve conduct-
ing evaluations when the machine is not in operation. This often requires the motor to be
shut off and, in some cases, partially or fully disassembled. These evaluations are typically
performed during scheduled maintenance intervals. The majority of these methods focus
on assessing the condition of the insulation system and the state of the magnetic core. The
methods utilized for the diagnosis of three-phase induction motors include, but are not
limited to [16]:
• Measurement of insulation resistance to ground, including the determination of polar-
ization index, absorption current, tangent delta capacitance, loss factor, and overvolt-
age tests.
• Analysis of insulation between turns, utilizing techniques such as the shock wave test,
measurement of insulation resistance, reverse sequence impedance, electrical circuit
analysis, and parameter estimation.
• Examination of the magnetic core through the application of nominal or reduced flux.
• Analysis of bearing lubricants, including the examination of physical characteristics
and particle content.
Sensors 2023, 23, 2649 3 of 20

2.2. In-Service Diagnostics

In-service diagnostic methods offer the advantage of providing information on the
state of the machine without interrupting its operation. This is a significant advantage,
as these methods typically include a series of analysis utilities, commonly in the form
of software, which can be utilized in predictive maintenance programs, making their
implementation highly desirable.
The most prominent methods for detecting faults through Condition Monitoring (CM)
are [10]:
• Stator current analysis or motor current signature analysis (MCSA) [17,18].
• Vibration analysis [19–21].
• Temperature measurement (thermography) [22–24].
• Partial discharges.
• Gas analysis.
• Axial dispersion flux analysis.
• Torque and speed fluctuations.
• Input power analysis.
Motor Current Signature Analysis (MCSA) is a non-invasive diagnostic technique
used to analyze the health of induction motors. It involves analyzing the current waveform
generated by the motor during normal operation and comparing it to a known good signa-
ture to detect anomalies and faults. It represents the first step in fault diagnosis, typically
accompanied by signal processing techniques such as Fast Fourier Transform (FFT) [19,25],
wavelet [17,26] or Hilbert–Huang Transform [27,28]. MCSA can provide inaccurate results
with the occurrence of saturation, interbar currents or magnetic asymmetry [29].
Reference [17] studied two techniques of signal processing to diagnose inter-turn
short circuit (ITSC) and the unbalanced voltage supply (UVS) using MCSA and processing
the information first with Fast Fourier Transform (FFT) and then with Discrete Wavelet
Energy Ratio (DWER). Reference [18] used MCSA for detecting the combination of bearing
faults by applying Discrete wavelet transform (DWT) to de-noise the signal, and then a
pre-fault component cancellation using an adaptive filter (Wiener filter) to finally estimate
the fault using Matrix Pencil method (MPM). A similar approach with MCSA and Wiener
filter is taken in [30] for several types of fault detection. In [31] is presented a guideline
for avoiding false fault detection based on MCSA. The guideline includes studying the
relationship between other faults and the potential for misidentification, as well as strategies
for overcoming such issues. MCSA is mainly applied in stationary conditions, but it is also
interesting to analyze nonstationary conditions as in [32], which is focused on detecting
eccentricities in the startup time modeling the induction motor, or in [27], which uses new
transient-based diagnosis approaches involving time-frequency transformations.
There are alternatives to the application of MCSA, mainly the analysis of vibration
signals. Accelerometers are used for the acquisition of vibration data, and their adequate
location conveys a great part of the result. Reference [19] proposed to utilize vibration
signals for detecting faults in rotors, with the support of classification machine learning
techniques. It is typical to utilize FFT in vibration data signal processing [33]. The fracture
of races in bearings can be detected by utilizing radial and axial accelerometers [20]. In
the event of bearing faults or imbalance, it is possible to measure the vibration of the
machine through an antenna and subsequently classify the issue using deep learning
techniques [34]. The combined usage of vibration and stator current signals shows great
promise in accurately classifying broken rotor bars [21].
Thermography is a method commonly used for diagnosing faults in machinery. To
effectively apply this technique, it is necessary firstly to pre-process the images, for example
using the region of interest (ROI). Once this has been done, the images can then be classified
using machine learning algorithms [23]. It is a common method in short-circuit faults or
inter-turn faults [24]. Model-based diagnosis is another alternative for fault diagnosis in
electrical machines [11,35].
 Thermography
Thermographyisisa amethod methodcommonly
commonlyused usedfor fordiagnosing
diagnosingfaults
faultsininmachinery.
machinery.ToTo
effectively
effectivelyapply
applythis
thistechnique,
technique,it itisisnecessary
necessaryfirstly
firstlytotopre-process
pre-processthetheimages,
images,for forexam-
exam-
ple
pleusing
usingthe theregion
regionofofinterest
interest(ROI).
(ROI).Once
Oncethis
thishashasbeen beendone,
done,thetheimages
imagescan canthen
thenbebe
Sensors 2023, 23, 2649 classified
classifiedusing
usingmachine
machinelearning
learningalgorithms
algorithms[23].[23].It Itisisa acommon
commonmethod
methodininshort-circuit
short-circuit
4 of 20
faults
faultsororinter-turn
inter-turnfaults
faults[24].
[24].Model-based
Model-baseddiagnosis
diagnosisisisanother anotheralternative
alternativefor forfault
faultdiag-
diag-
nosis in electrical machines
nosis in electrical machines [11,35].[11,35].
InInthis
Inthispaper,
this paper,
paper, MCSA
MCSA
MCSA isisis
selected asas
selected
selected asa anon-invasive,
a non-invasive,
non-invasive, cost-effective,
cost-effective,
cost-effective, and
and
and reliable
reliable
reliable technol-
technol-
technol-
ogy
ogy[12,36].
[12,36].The
The next
next subsections
subsections briefly
briefly present
presentsome
some faults
faultsdetected
detected
ogy [12,36]. The next subsections briefly present some faults detected with Motor Current with
withMotor
Motor Current
Current
Signature
Signature Analysis.
Analysis.
Signature Analysis.

2.2.1.
2.2.1.Eccentricity
2.2.1. EccentricityFaults
Eccentricity Faults
Faults
TwoTwodistinct
Two distincttypes
distinct ofof
types eccentricity
of faults
eccentricity
eccentricity can
faults
faultscanoccur:
occur:
can static
occur: and
static anddynamic.
static dynamic.
and Figure 1 1illus-
Figure
dynamic. Figureillus-
1
trates both.
illustrates
trates both.both.

STATOR
STATOR STATOR
STATOR STATOR
STATOR

AIR GAP
AIR GAP
AIR GAP
AIR GAP AIR GAP
AIR GAP
ROTOR
ROTOR
ROTOR
ROTOR ROTOR
ROTOR

Figure
Figure 1.1.1.
Figure Healthy
Healthy
Healthymotor
motor
motor (left),
(left),
(left),static
static eccentricity
eccentricity
static fault
fault
eccentricity (center),
(center),
fault and
and
(center), dynamic
dynamic
and eccentricity
eccentricity
dynamic (right).
(right).
eccentricity (right).

Eccentricity is related to air gap distortion. When a mixed eccentricity fault occurs,
Eccentricity
Eccentricityisisrelated
relatedtotoairairgapgapdistortion.
distortion.WhenWhena amixed
mixedeccentricity
eccentricityfault
faultoccurs,
occurs,
sidebands of the fundamental frequency of the power supply appear. These sideband
sidebands
sidebandsofofthe
thefundamental
fundamentalfrequency
frequencyofofthe thepower
powersupply
supplyappear.
appear.These
Thesesideband
sidebandfre-fre-
frequencies are given by [10,11,32]:
quencies
quenciesare
aregiven
givenbyby[10,11,32]:
[10,11,32]:
11−−s−𝑠 𝑠
  
f𝑓ecc𝑓 ===
f𝑓s 𝑓11± 1
1±m±𝑚𝑚
p𝑝 𝑝
(1)
(1)(1)

where𝑓fs𝑓is
where
where isis
the
the supply
the supply
supply frequency,
frequency,
frequency, s 𝑠is𝑠isthe
the
is theslip,
slip,
slip, 𝑝 𝑝the
p is isisthe
thenumber
numbernumber ofofpole
of pole polepairs,
pairs,
pairs, and m =𝑚1,
and
and 𝑚=2,=1,3,
1,
2,.2,
.3,. 3,
.…. ….In
In Incontrast
contrast
contrast totoother
to other otherapproaches,
approaches,
approaches, thisthis
thismethod
methodmethod doesdoesnotnot
does notnecessitate
necessitate
necessitate knowledge
knowledge
knowledge ofof
of the
thethemechanical
mechanical
mechanical properties
properties
properties ofofthe
of the themotor.
motor.motor.

2.2.2.
2.2.2. Broken
2.2.2.Broken Rotor
BrokenRotor Bars
RotorBarsBars
The
TheThebreaking
breaking
breaking ofof
ofrotor
rotor
rotorbars
bars
barsisisis one
one
oneofof the
the
of the primary
primary
primary causes
causes
causesofof failure
failure
of inin
failure induction
induction
in induction motors,
motors,
motors,
particularly
particularly in high-power motors that are frequently started under load. Figure 2shows
particularly inin high-power
high-power motors
motors that
that are
are frequently
frequently started
started under
under load.
load. Figure
Figure 22shows
shows
an
an example
exampleofof
anexample ofaaabroken
broken rotor
brokenrotor bar.
rotorbar.
bar.

Figure
Figure 2.2.2.
Figure Broken
Broken rotor
rotor
Broken bar.
bar.
rotor bar.

Thepresence
The presence of broken rotor
rotor bars in a motor results in the destruction of rotor symme-
The presenceofofbroken
broken rotorbars barsinina amotor
motorresults
resultsininthe
thedestruction
destructionofofrotor
rotorsym-
sym-
try, which,
metry, in in
which, turn, generates a rotating field that produces harmonics in thethe
stator current
metry, which, inturn, generates
turn, generates a arotating
rotatingfield
fieldthat
thatproduces
producesharmonics
harmonicsinin stator
the statorcur-
cur-
as described
rent by [36,37]:
rentasasdescribed
describedbyby[36,37]:
[36,37]:
f brb2 = f s [1 ± k2s] (2)
𝑓 𝑓 ==𝑓 𝑓 [1[1
±± 𝑘2𝑠]
𝑘2𝑠] (2)(2)
where k = 1, 2, 3, 4, . . . .
where𝑘 𝑘==1,1,
where 2,2,
3,3,
4,4,
….….
2.2.3. Bearing Damage
The prevalent cause of installation issues is the improper application of force when
installing the bearing onto the shaft or within the housing. This can result in physical
damage, such as brinelling or false brinelling of the raceways, leading to an early failure of
the bearing. Figure 3 shows two examples of bearing damage.
 2.2.3. Bearing Damage
The prevalent cause of installation issues is the improper application of force when
installing the bearing onto the shaft or within the housing. This can result in physical
Sensors 2023, 23, 2649 5 of 20
damage, such as brinelling or false brinelling of the raceways, leading to an early failure
of the bearing. Figure 3 shows two examples of bearing damage.

Figure3.
Figure Bearingdamage,
3.Bearing damage,inner
innerring
ring(left)
(left)and
and outer
outer ring
ring (right)
(right) [38].
[38].

The mechanical
The mechanical displacement
displacement caused
caused byby damaged
damaged bearings
bearings leads
leads to
tovariations
variations in
inthe
the
machine air gap, which can be characterized by a combination of rotating eccentricities
machine air gap, which can be characterized by a combination of rotating eccentricities that
move
that in both
move directions.
in both Similar
directions. to the to
Similar deviation in the air
the deviation in gap, these
the air fluctuations
gap, generate
these fluctuations
stator currents at specific frequencies [10,35,37]:
generate stator currents at specific frequencies [10,35,37]:
𝑓 ==| f|𝑓
f bng s ±±m
𝑚𝑓f i,o, | | (3)(3)

where 𝑚
where m == 1,2,3,
1,2,3,… . and𝑓 , fi,oisisone
. . and oneofofthethe
distinctive vibration
distinctive frequencies
vibration which
frequencies are de-
which are
termined by the dimensions of the
determined by the dimensions of the bearing. bearing.
𝑛  𝑏𝑑
𝑓 , =n 𝑓 1 ±bd 𝑐𝑜𝑠𝛽

f i,o = 2f r 1 ± 𝑝𝑑cosβ (4)(4)
2 pd
where 𝑛 is the number of bearing balls, 𝑓 is the mechanical rotor speed in Hz, 𝑏𝑑 is the
where
ball n is the𝑝𝑑
diameter, number
is the of bearing
bearing pitchballs, fr is theand
diameter, mechanical rotor speed
𝛽 is the contact the bd
angleinofHz, is the
balls on
ball diameter,
the races. pd is the bearing pitch diameter, and β is the contact angle of the balls on
the races.
In this paper, the focus is on the analysis of broken rotor bars.
In this paper, the focus is on the analysis of broken rotor bars.
2.3. Machine Learning Algorithms
2.3. Machine Learning Algorithms
Machine learning (ML) techniques have been proven effective and dependable for a
Machine learning (ML) techniques have been proven effective and dependable for a
wide
wide range
range ofof applications, including identifying
applications, including identifyingpatterns,
patterns,forecasting,
forecasting,modeling,
modeling,optimiza-
optimi-
zation and data analysis. ML methods can be categorized into
tion and data analysis. ML methods can be categorized into supervised, unsupervised, supervised, unsupervised,
semi-supervised
semi-supervisedand andreinforcement
reinforcement learning learning [39],[39], depending
depending on on the
the characteristics
characteristics ofof the
the
data used and the type of system to
data used and the type of system to be developed. be developed.
Supervised
Supervisedlearning
learninguses useslabeled
labeledinputinputdatadata toto
predict anan
predict output
output variable andand
variable cancan
be
further divided into classification and regression techniques.
be further divided into classification and regression techniques. The standard procedure The standard procedure in
supervised learning is to first train a classification model using a
in supervised learning is to first train a classification model using a labeled dataset that labeled dataset that en-
compasses
encompasses thetherelevant
relevant classes.
classes. Then,
Then, thethe
model
model is istested
testedon
onanother
anotherdata
dataset,
set,which
which is is
unlabeled, in order to evaluate its
unlabeled, in order to evaluate its performance. performance.
The
The use
use ofof machine
machine learning
learning methodsmethods is is widely
widely used used for
for fault
fault diagnosis
diagnosis of
of induction
induction
machines
machinesin in literature.
literature. ItItisis common
commonto to find
find fault
fault diagnosis
diagnosis solutions
solutions based
based on
on Artificial
Artificial
Neural
Neural Networks (ANN) and Decision Tree (DT) methods, such as Random Forest(RF)
Networks (ANN) and Decision Tree (DT) methods, such as Random Forest (RF)
and
andSupport
SupportVector
VectorMachine
Machine (SVM)
(SVM) [40], among
[40], among others.
others.Renowned
Renowned methods, such as
methods, Ran-
such as
dom Forest,
Random are continuously
Forest, are continuously beingbeingadvocated
advocatedfor detecting broken
for detecting rotor bars
broken rotor[41]
bars or[41]
otheror
Decision Trees models
other Decision for short
Trees models for circuit faultsfaults
short circuit [42]. Several models,
[42]. Several including
models, Random
including For-
Random
est andand
Forest Support
Support Vector
VectorMachine,
Machine, were
were used
used totomonitor
monitorswitch
switchopen-circuit
open-circuitfaults
faultsinin an
an
inverter-fed
inverter-fed induction
induction motor
motor in in [28].
[28]. An
Anexample
exampleof ofanother
anothertype
typeof ofsolution
solutionthat
thatemploys
employs
AI
AI and
and machine learning is is [43],
[43], where
whereaagenetic
geneticalgorithm
algorithm(GA) (GA)was waspresented
presented forfor bear-
bearing
ing fault
fault diagnosis
diagnosis basedbased on information
on information obtained
obtained from from
MCSA.MCSA.
The research community is devoting significant efforts to real-time and online so-
lutions [10]. In [44], the application of 1-D Convolutional Neural Networks (CNN) is
presented, which offers real-time fault detection in a powerful Field-Programmable Gate
Array (FPGA) deployment without requiring a feature extraction algorithm using current
measurements. In [19], Nearest Neighbour (NN), Linear Discriminant Analysis (LDA), and
Support Vector Machine were used to classify rotor faults from vibration data.
 The research community is devoting significant efforts to real-time and online solu-
tions [10]. In [44], the application of 1-D Convolutional Neural Networks (CNN) is pre-
sented, which offers real-time fault detection in a powerful Field-Programmable Gate Ar-
Sensors 2023, 23, 2649 ray (FPGA) deployment without requiring a feature extraction algorithm using current 6 of 20
measurements. In [19], Nearest Neighbour (NN), Linear Discriminant Analysis (LDA),
and Support Vector Machine were used to classify rotor faults from vibration data.
Deep Learning techniques can also be applied to classifying faults, such as the use of
Deep Learning techniques can also be applied to classifying faults, such as the use of
Deep Convolutional Neural Networks (DCNN) in the case of data obtained from vibra-
Deep Convolutional Neural Networks (DCNN) in the case of data obtained from vibration
tion signals for broken rotor bar analysis [21] or for the diagnosis of bearing faults [45].
signals for broken rotor bar analysis [21] or for the diagnosis of bearing faults [45].
In
In[23],
[23], aa Support
Support Vector
Vector Machine
Machine was was applied
applied toto classify
classify thermography
thermography imagesimagesin in
order to detect short-circuit faults. Meanwhile, in [33], regression models
order to detect short-circuit faults. Meanwhile, in [33], regression models were presented for were presented
for classifying
classifying the the severity
severity of eccentricity
of eccentricity faults,
faults, usingusing
data data
from from
currentcurrent harmonics,
harmonics, vi-
vibration,
bration,
rotationrotation
speed, and speed, andwith
torque, torque, withnot
the aim theonly
aimofnot only of detecting
detecting but also classifying
but also classifying the gravity
the
of the fault. More sophisticated approaches have been taken for classifyingfor
gravity of the fault. More sophisticated approaches have been taken classifying
nonstationary
nonstationary
operations without operations without
a dataset a dataset such
for training, for training,
as using such as using anframework,
an Ensemble Ensemble frame-Fuzzy
work, Fuzzy Rough Active Learning, and drift detection
Rough Active Learning, and drift detection for detecting broken bars [46]. for detecting broken bars [46].
The
Theavailability
availabilityofoftraining
trainingsets setscan
canbebea challenge
a challenge in both
in bothindustrial settings
industrial andand
settings la-
boratories,
laboratories, as as
there may
there may bebe a alimited
limitednumber
numberofoffaulty
faultymachines
machinesfor fortraining
trainingpurposes,
purposes,
and
andcollecting
collectingdata datawith
withmultiple
multiple faults
faultsin in
a single machine
a single machine cancanbe difficult [39].[39].
be difficult It is also
It is
observed
also observed that many of the research studies presented in the literature are basedda-
that many of the research studies presented in the literature are based on on
tasets
datasetsfrom laboratory
from tests.tests.
laboratory It would be interesting
It would to starttorecording
be interesting real data
start recording from
real dataactual
from
environments
actual environments[47]. [47].
The
Thefollowing
following subsections present present aaconcise
conciseoverview
overviewofofthe themachine
machine learning
learning meth-
methods
ods employed
employed in this
in this paper.These
paper. Thesemethods
methodswere werechosen
chosen due due to their high
high performance,
performance,
explainability
explainabilityand andsuitability
suitabilityforforimplementation
implementationon onresource-constrained
resource-constrainedplatforms,
platforms,as asisis
the
thecase
caseat athand.
hand.

2.3.1.Decision
2.3.1. DecisionTrees
Trees
AAdecision
decisiontreetree(DT)
(DT) is aiscommonly
a commonly usedused machine
machine learning
learning method method for decision
for decision sup-
support in data analysis and statistics, with a particular focus on artificial
port in data analysis and statistics, with a particular focus on artificial data mining. The data mining. The
goal of DTs is to create a model that predicts the target value based
goal of DTs is to create a model that predicts the target value based on multiple inputs. on multiple inputs.
DecisionTree
Decision Tree methods,
methods, therefore,
therefore, are are aa widely
widely used
used model
model forfor solving
solving classification
classificationand and
regression problems in the context of supervised learning. The structure
regression problems in the context of supervised learning. The structure of DTs is depicted of DTs is depicted
bybranches
by branchesand andleaves,
leaves,where
wherebranches
branchescontain
containthe
theattributes
attributesthat
thatthethefunction
functionrelies
relieson,
on,
and leaves contain the function’s value. Other nodes contain attributes
and leaves contain the function’s value. Other nodes contain attributes that distinguish that distinguish the
decision
the decisioncases. AnAn
cases. illustration
illustrationof the DTDT
of the algorithm
algorithmis presented
is presentedin Figure
in Figure4. 4.

FIRST FEATURE
attr. unfulfilled attr. fulfilled

SECOND FEATURE Decicion 1

attr. unfulfilled attr. fulfilled

THIRD FEATURE Decicion 2

attr. unfulfilled attr. fulfilled
Decision 3 Decision 4

Decisiontree
Figure4.4.Decision
Figure treediagram.
diagram.

Comparedto
Compared toother
otherdecision
decisionmodels,
models,DTsDTsare
aresimple
simpleand
andrequire
requireonly
onlyaasmall
smallamount
amount
of data to produce results. They can also be combined with other decision
of data to produce results. They can also be combined with other decision models to in- models to
increase accuracy. However, undoubtedly, the main feature that makes
crease accuracy. However, undoubtedly, the main feature that makes these methods onethese methods
one
of theofmost
the most
widelywidely used
used is is theofease
the ease of explanation
explanation of the decision
of the decision processprocess thanks
thanks to to
its tree
its tree structure,
structure, which inherently
which inherently allows allows for reproducing
for reproducing the decision
the decision processprocess obtained
obtained after
after training the model. For this reason, decision trees are known as white-box
training the model. For this reason, decision trees are known as white-box models, as models, as
opposed to other models, such as Artificial Neural Networks or Support Vector Machines,
which are known as black-box models, since they do not offer such an understandable
explanation of the decision process. However, DTs are inherently unstable, and even a
small change in input data can result in a significant change in the decision tree structure
and potentially lead to inaccurate results.
 opposed to other models, such as Artificial Neural Networks or Support Vector Machines,
which are known as black-box models, since they do not offer such an understandable
explanation of the decision process. However, DTs are inherently unstable, and even a
Sensors 2023, 23, 2649 small change in input data can result in a significant change in the decision tree structure
7 of 20
and potentially lead to inaccurate results.

2.3.2. Random
2.3.2. Random Forest
Forest
Decision trees
Decision trees present
present aa dilemma.
dilemma. A A deep
deep tree
tree with
with numerous
numerous leaves
leaves can
can result
result in
in
overfitting as the prediction is based solely on the few features present in
overfitting as the prediction is based solely on the few features present in its leaves. On theits leaves. On
the other
other hand,hand, a shallow
a shallow tree tree
withwith
few few leaves
leaves lackslacks the ability
the ability to capture
to capture distinctions
distinctions in
in the
the raw data and thus performs
raw data and thus performs poorly. poorly.
In contrast,
In contrast, the
the random
random forest
forest approach
approach utilizes
utilizes multiple
multiple trees,
trees, which
which are
are built
built using
using
random subset
random subset attributes.
attributes. This
This way,
way, anan uncorrelated
uncorrelated forest
forest is
is generated.
generated. Then,
Then, the
the predic-
predic-
tion is
tion is built
built by
by taking
taking the
the average
average ofof the
the predictions
predictions made
made byby each
each individual
individual tree
tree or
or using
using
other criteria. As a result, it generally exhibits improved predictive accuracy compared to to
other criteria. As a result, it generally exhibits improved predictive accuracy compared a
a single
single decision
decision tree.
tree.

2.3.3.
2.3.3. Support
Support Vector
VectorMachines
Machines
Intuitively,
Intuitively,Support
SupportVector
VectorMachines
Machines(SVMs)
(SVMs)arearemodels
modelsthat
thatmap
mapsample
samplepoints
points into
into
aa multi-dimensional
multi-dimensional space, separating them into two distinct classes. This division is
space, separating them into two distinct classes. This division is
achieved
achieved byby aa hyperplane,
hyperplane, as as demonstrated
demonstrated in in Figure
Figure 5.
5. The
The hyperplane
hyperplane is is determined
determined by by
the
thetwo
twoclosest
closestsample
samplepoints
pointsfrom each
from class,
each referred
class, to as
referred to support vectors.
as support These
vectors. support
These sup-
vectors determine
port vectors the boundaries
determine of theofclasses,
the boundaries withwith
the classes, the distance between
the distance betweenthe support
the sup-
vectors of each
port vectors of class, referred
each class, to as the
referred margin.
to as Although
the margin. there are
Although multiple
there hyperplanes
are multiple hyper-
that can effectively classify the samples, it is common practice to select the hyperplane
planes that can effectively classify the samples, it is common practice to select the hyper- with
the greatest margin.
plane with the greatest margin.

Cl a s s 1
Cl a s s 2
H Hy p e rp l a n e

Figure 5.
Figure 5. Support
Support Vector
VectorMachine.
Machine.

In the
In thefigure,
figure,twotwo
datadata classes
classes are represented:
are represented: Class 1Class 1 (triangles)
(triangles) and Classand Class 2
2 (squares).
In additionIntoaddition
(squares). linear classification, SVMs can
to linear classification, dealcan
SVMs withdealnon-linear classification
with non-linear using
classification
the kernel trick, using polynomial, radial or sigmoid kernels to map
using the kernel trick, using polynomial, radial or sigmoid kernels to map the data to a the data to a higher
dimension
higher dimensionwhere data whereis separable.
data is separable.
When
Whenaamulticlassification
multiclassificationproblem
problem is addressed, let n
is addressed, letbenthe
be number
the number of classes, a totala
of classes,
n ( n −1) ( )
of
total 2of binary binary classifiers must be trained to separate all the pairs of n classes. way,
classifiers must be trained to separate all the pairs of n classes. This This
the
way, the predicted for
predicted class a new
class for ainstance is theismost
new instance votedvoted
the most among all the
among allclassifiers built.built.
the classifiers For
that reason,
For that when
reason, the number
when the numberof classes is high,
of classes andand
is high, the the
dataset is large,
dataset SVMs
is large, SVMsinvolve an
involve
expensive computational cost in terms of time
an expensive computational cost in terms of time and memory. and memory.

2.3.4. Evaluation of Methods

2.3.4. Evaluation of Methods
The Confusion Matrix is utilized to assess the effectiveness of various classification
The Confusion Matrix is utilized to assess the effectiveness of various classification
techniques by determining the accuracy of a classifier. It is represented as a square ma-
techniques by determining the accuracy of a classifier. It is represented as a square matrix,
trix, with the dimensions being determined by the number of classes present. In the
with the dimensions being determined by the number of classes present. In the current
current scenario, there are two classes: healthy and faulty. The confusion matrix shows the
next values:
• True Positives (TP): This occurs when the prediction that an observation belongs to a
specific class is accurate, as the observation indeed belongs to that class.
• True Negatives (TN): This occurs when the prediction that an observation does not
belong to a specific class is accurate, as the observation indeed does not belong to
that class.
Sensors 2023, 23, 2649 8 of 20

• False Positives (FP): This occurs when the prediction that an observation belongs to a
specific class is incorrect, as the observation actually does not belong to that class.
• False Negatives (FN): This occurs when the prediction that an observation does not
belong to a specific class is incorrect, as the observation actually belongs to that class.
It is necessary to indicate the analysis parameters to measure the performance of the
different classification methods. The accuracy is the ratio of true cases to all cases, as shown
in (5). The recall is focused on the positive class. It is the ratio of the correct positive
predictions to all observations in the positive class, given by (6). The precision is the ratio
of the correct positive predictions to all positive predictions, given by (7).

TP + TN
Accuracy = (5)
TP + TN + FP + FN
TP
Recall = (6)
TP + FN
TP
Precision = (7)
TP + FP
F1 score is a measure given by (8), being the harmonic mean of precision and recall. F1
varies from 0 to 1, with zero as the worst value.

Precision·Recall
F1 = 2 (8)
Precision + Recall

2.4. Industrial Cyber-Physical Systems

A Cyber-Physical System (CPS) combines computational applications with physical
devices and is structured as a network of interacting cyber and physical elements [48].
Unlike embedded systems, which primarily focus on computational elements housed
in standalone devices, CPSs are primarily designed as a larger network of interacting
computational and physical elements [49]. When CPSs are used in industrial environments
emergence the Industrial Cyber-Physical Systems (ICPSs) [50] or Cyber-Physical Production
Systems (CPPSs) [51], in combination with the widespread use of the Internet of Things
(IoT) [52].
Under the subject matter discussed in this paper, the ICPS will consist of various
components, but those related to fault diagnosis can be summarized as follows:
• Physical component: it comprises the hardware and all the sensors involved in data
acquisition, which is necessary to conduct fault diagnosis.
• Cyber component: it is the intelligent component where data collected from the
physical world is analyzed and interpreted. This is where emerging technologies
are utilized.
Two different approaches can be taken at this stage: Edge or Fog/Cloud Computing.

2.4.1. Fog/Cloud Computing

In these scenarios, data is collected at the local level and then transmitted to the fog
or cloud for further processing. This approach offers high computing power. Researchers
often focus on testing or developing new machine learning models using simulation
environments such as MATLAB/Simulink [53] or by modifying datasets to include potential
failures in order to evaluate the performance of various classification techniques [54]. It is
typical for machines to be positioned in a fixed location, although there are exceptions. An
example of this is the predictive maintenance of mobile machinery in a mining complex
utilizing Bluetooth Low Energy vibration sensors, as presented in [55].

2.4.2. Edge Computing

When the cyber and physical components are situated in close proximity to the
electrical machine under examination, this approach can be challenging to implement
Sensors 2023, 23, 2649 9 of 20

in industrial settings due to the computational resources required for data processing,
necessitating the presence of a nearby personal computer.
In controlled environments, such as laboratories, this setup is prevalent. It typically
involves the use of data acquisition cards and specialized software running on a personal
computer to provoke faults in an electrical machine and study their effects. In [18], a
National Instruments Data Acquisition (NI DAQ) device is utilized to acquire information
from a current sensor in order to test various pre-filter techniques for data preparation.
In [21], a NI DAQ, a current probe, and a tachometer are utilized to evaluate an induction
motor and investigate various machine learning techniques, including neural networks
and deep multimodal learning. Reference [17] suggests the utilization of a deep network-
based approach to examine the characteristics of thermograms with the aid of a thermal
camera. NI DAQ and current sensors are used for short-circuit detection [42]. These
methods are suitable for academic settings to conduct experiments and evaluate outcomes
to advance new techniques. However, their implementation in industrial environments
may prove challenging.
In a limited number of instances, edge computing has been utilized to its fullest
potential. As demonstrated in reference [56], a method for detecting bearing faults through
vibration analysis utilizing ultra-low power wireless sensors is proposed. These sensors
collect and analyze the data and only transmit information pertaining to the operational
status of the machine to the rest of the ICPS. This application’s feasibility is made possible
through the employment of single-axis accelerometers and the creation of a lightweight
neural network classifier.
An edge computing solution for fault diagnosis based on an ICPS architecture imple-
mented over the Arduino platform is presented in the following sections.

3. Proposal
This paper presents a rotor bar-breaking fault diagnosis method implemented in an
ICPS approach. The subsequent subsections introduce the components of the ICPS. It
functions using MCSA data and employs a machine learning-based classification method.
Its purpose is to operate at the edge of a platform with limited resources.

3.1. ICPS Physical Part

The physical part of the ICPS oversees the data acquisition system. It must collect
enough current measurements to extract the components of the power spectrum.
The Arduino platform has been selected because it is looking for an affordable data
acquisition system. It includes an analog-to-digital converter (ADC) with a default sampling
rate of 9600 Hz. To ensure accurate detection of the faults is required a sampling frequency
adheres to the Shannon-Nyquist frequency, thereby avoiding potential issues of aliasing or
inaccurate signal reconstruction. Therefore, with the default configuration, it is possible to
correctly sense signals up to 4800 Hz. It is also possible to increase the sampling frequency
of Arduino, setting the prescaler of the clock. For example, the Arduino Uno is equipped
with a microcontroller ATMega328P that includes a clock speed of 16 MHz and a prescale
factor of 128 by default. In case of necessity, it is possible to achieve a higher rate of 615 kHz
by adjusting the prescale factor and using interrupts. The sampling frequency will be set
depending on the physical characteristics of the monitored machine.
In order to measure the current consumed by induction motors, a non-invasive clamp
sensor, model SCT-013-030 by YHDC® , was selected. As shown in Figure 6, this sensor is
a current transformer with a range of 0–30 A and provides an output voltage range of 0
to 1 V.
 machine.
sampling frequency will be set depending on the physical characteristics of the monitored
machine. In order to measure the current consumed by induction motors, a non-invasive clamp
In order tosensor, model SCT-013-030 by YHDC ®, was selected. As shown in Figure 6, this sensor is
measure the current consumed by induction motors, a non-invasive clamp
sensor, model aSCT-013-030
current transformer
by YHDCwith a range
®, was of 0–30
selected. A and provides
As shown in Figurean
6, output voltage
this sensor is range of 0 to
1 V.
Sensors 2023, 23, 2649a current transformer with a range of 0–30 A and provides an output voltage range of 0 to 10 of 20

1 V.

Figure 6. Signal conditioning circuit.

Figure 6. SignalFigure
conditioning
6. Signalcircuit.
conditioning circuit.
Additional hardware has been added to adapt the signal from the sensor to the input
of Arduino’s
Additional
Additional hardware ADC,
has beenincluding
hardware
added an offset
hastobeen added
adapt and amplification
to adapt
the signal fromthethe
signalstage
from
sensor to by
themeans
sensor of
input AD623
to the by
input
Analog
of Arduino’s of
ADC, Devices
Arduino’s
including
® and an OP113 operational amplifier was configured in voltage follower
ADC,an including
offset andan offset and amplification
amplification stage by means stage
of by means
AD623 by of AD623 by
to further
Analog improve
Devices ® andthean characteristic
OP113 curve
operational of the sensor.
amplifier was The final
configured
Analog Devices and an OP113 operational amplifier was configured in voltage follower
® characteristic
in voltage curve of
follower
the
to further improve sensor
to further is as follows:
improve the characteristic
the characteristic curve of
curve of the sensor. thefinal
The sensor. The final characteristic
characteristic curve of curve of
thefollows:
the sensor is as sensor is as follows: 30
𝑖(𝑡) 30
= (𝑣(𝑡) − 0.5) (9)
30 i ( t ) = (4v(t) − 0.5) (9)
𝑖(𝑡) = (𝑣(𝑡) − 0.5) 4 (9)
4
3.2.
3.2.ICPS
ICPSCyber
CyberPart
Part
The cyber component
3.2. ICPS Cyber PartThe cyber component is responsible
is responsibleforfor
processing the the
processing datadata
obtained fromfrom
obtained the sensors,
the sen-
extracting
sors, relevant
extracting features,
relevant classifying
features, them,
classifyingand communicating
them, and the status
communicating of
the the machine
status of the
The cyber component is responsible for processing the data obtained from the sen-
tomachine
the rest to
of the
the rest
ICPS,ofas depicted
the ICPS, asindepicted
Figure 7.in Figure 7.
sors, extracting relevant features, classifying them, and communicating the status of the
machine to the rest of the ICPS, as depicted in Figure 7.
Raw data
Raw data
Signal processing
Signal processing
Feature extraction
Feature extraction
Classification
Classification
Dissemination
Dissemination
Figure7.7.Flowchart
Figure Flowchartofofthe
thecyber
cyberpart.
part.
Figure 7. Flowchart
3.2.1.ofSignal
the cyber part.
Processing
3.2.1. Signal Processing
Raw
Rawdatadatacoming
coming from
from thethe sensors
sensorsis aistime-based
a time-based signal. A frequency-based
signal. signal
A frequency-based is
signal
3.2.1. Signal Processing
needed to extract the convenient information.
Raw data iscoming
neededfrom
FFT
to extract
isisthe chosen
the convenient
the sensors is a in
method this
information.
time-based
paper. signal. A frequency-based signal
FFT the chosen method
is needed to extract the convenient information. in this paper.So, So,the
thesignal
signalhas hastotobebewindowed:
windowed:
∗ ∗
(𝑡)has to
(𝑡) =R (𝑖t) f or𝑓𝑜𝑟 [𝑡0,𝜖windowed:
[0,
T ] 𝑇]

FFT is the chosen method in this paper. So,∗ the isignal t ebe
i∗R (t)𝑖 = (10)
0 𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
∗ (𝑡) 𝑖 ∗ (𝑡) 𝑓𝑜𝑟 𝑡 𝜖0[0, 𝑇] Otherwise
𝑖 = (10)
And then compute the 0Fourier 𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
Transform:
And then compute the Fourier Transform:
And then compute the Fourier Transform:
Z∞
IR∗ (ω ) = i∗R (t)e− jωt dt (11)
−∞

The output signal obtained after the FFT calculation of the acquired current signal is
presented in Figure 8.
 𝐼 ∗ (𝜔) = 𝑖 ∗ (𝑡) 𝑒 𝑑𝑡 (11)

Sensors 2023, 23, 2649 The output signal obtained after the FFT calculation of the acquired current signal
11 of 20is
presented in Figure 8.

30
30
20

20 10

10 0

-10
0
-20
-10
-30

-20 -40
0 50 100 150 200 0 100 200 300 400 500
time (in ms) Frequency (Hz)

Figure8.8.Input
Figure Inputsignal
signal(left)
(left)and
andoutput
outputsignal
signal(right)
(right)ofofthe
thesignal
signalprocessing.
processing.

3.2.2.
3.2.2.Feature
FeatureSelection
Selection
As
Asstated
statedininprevious
previousSection
Section2.2.2,
2.2.2,ititisisnecessary
necessarytotomeasure
measurethe thesidebands
sidebandsofofthe the
main frequency to detect broken rotor bars. Taking that information as
main frequency to detect broken rotor bars. Taking that information as starting point and starting point and
given
giventhat
thatwewewillwillbebetraining
trainingthe themodel
modelbasedbasedon ondata
datafrom
fromaaspecific
specificmotor,
motor,but butour
our
objective
objectiveisistotodesign
designaamodel
modelthatthatcan
canbebeused
usedon ondifferent
differentmotors,
motors,ititisisessential
essentialthat
thatthe
the
definition
definitionofofthe features
the featuresis extrapolatable
is extrapolatable to any motor.
to any Hence,
motor. the following
Hence, features
the following have
features
been
haveproposed:
been proposed:
•• Difference
Differencebetween
betweenthe theamplitude
amplitudeofofthe themain
mainfrequency
frequencyand
andthe
thesidebands
sidebandsup uptoto5th
5th
order, in Db. The selection of sidebands is according to Equation (2)
order, in Db. The selection of sidebands is according to Equation (2) in Section 2.2.2. in Section 2.2.2.
The
Thefrequency
frequencyofofthesethesesidebands
sidebandsdepends
dependson onthethecharacteristics
characteristicsofofthe
themotor.
motor.
•• TheThe relative root mean square (RMS) error of the current signal in comparisonwith
relative root mean square (RMS) error of the current signal in comparison with
the
theexpected
expectedRMS RMSvalue,
value,its itsnominal
nominalvalue valueaccording
accordingtotothe
themanufacturer.
manufacturer.ItItisisthe the
relative
relativerootrootmean
meansquare
squareerrorerror(RRMSE).
(RRMSE).The TheRMSRMSisiscalculated
calculatedininthe thefrequency
frequency
range
rangecovering
coveringthe theabove
abovesidebands.
sidebands.
q
n
Irms𝐼 = = ∑∑j=1 i2j,rms
𝑖, (12)
(12)
Irms,nominal − Irms
RRMSE = 𝐼 −𝐼 (13)
Irms,nominal
,
RRMSE = (13)
𝐼 ,
where ij,rms is the rms current value of the n frequency from the FFT, and Irms,nominal is the
where 𝑖current
nominal , is of
the the
rmscharacteristics
current value of
ofthe
the motor.
n frequency from the FFT, and 𝐼 , is
•the nominal current of the characteristics of the motor.
The total harmonic distortion (THD) calculated among the range covering the above
• sidebands:
The total harmonic distortion (THD) calculated among the range covering the above
sidebands: q
∑nj=1 i2j, rms
THD( I ) f = (14)
I f∑ 𝑖
tal,rms , (14)
THD(𝐼) =
where iftal,rms is the rms current value of the main𝐼frequency. ,
These
where 𝐼 , two parameters are chosen to provide the model
is the rms current value of the main frequency. with an idea of the number of
harmonics in the signal that could lead to a proper classification.
These two parameters are chosen to provide the model with an idea of the number
of harmonics
3.2.3. in the signal that could lead to a proper classification.
Classification
It is proposed to apply supervised learning techniques for classifying new observations
between healthy or faulty states. The classification is based on Decision Trees (DTs),
Random Forest (RF), and Support Vector Machine (SVM), paying special attention to their
possible implementation on Arduino.
 tions between healthy or faulty states. The classification is based on Decision Trees (DTs),
Random Forest (RF), and Support Vector Machine (SVM), paying special attention to their
possible implementation on Arduino.

Sensors 2023, 23, 2649 3.2.4. Dissemination 12 of 20

Upon determining the state of the machine, the system will disseminate the infor-
mation to the rest of the ICPS that control the system. The mode of communication that is
3.2.4. Dissemination
used, either LAN or wireless, will depend on the scenario. For systems that utilize an Ar-
duino Uno,determining
Upon the state of of
the implementation theamachine, the system
communication will disseminate
shield is a feasiblethe informa-
option. However,
tion to the rest of the ICPS that control the system. The mode of communication that is used,
it may be more efficient to use a UART-LAN/WIFI interface module to handle the com-
either LAN or wireless, will depend on the scenario. For systems that utilize an Arduino
munication protocol and allow the Arduino to only transmit the information via the
Uno, the implementation of a communication shield is a feasible option. However, it may
USART port in to
be more efficient order
use atoUART-LAN/WIFI
avoid interfering with the
interface normal
module operation
to handle of the microcontrol-
the communication
ler.
protocol and allow the Arduino to only transmit the information via the USART port in
order to avoid interfering with the normal operation of the microcontroller.
4. Case Study
4. Case Study
The proposed fault detection method based on Arduino boards has been tested on
The proposed fault detection method based on Arduino boards has been tested on
equipment of the Mechanical Engineering Laboratory at the Mining and Industrial Engi-
equipment of the Mechanical Engineering Laboratory at the Mining and Industrial En-
neering School of Almadén (UCLM-Spain), specifically, a lathe and a milling machine
gineering School of Almadén (UCLM-Spain), specifically, a lathe and a milling machine
(Figure 9) equippedwith
(Figure 9) equipped with a three
a three phases
phases electrical
electrical motor
motor AEG AEG AM
AM 100 LZ100 LZa4Siemens
4 and and a Siemens
1LA3 106-4AA20respectively,
1LA3 106-4AA20 respectively, with
with the the characteristics
characteristics shownshown in1.Table 1.
in Table

Figure 9. Lathe
Figure 9. Lathe(left)
(left)and
and milling
milling machine
machine (right)
(right) tested.
tested.

Table 1.Characteristics
Table 1. Characteristicsof of
thethe Electrical
Electrical Motor
Motor tested.
tested.

Description
Description Lathe
Lathe MillingMilling
MachineMachine
Manufacture
Manufacture AEGAEG SiemensSiemens
Power
Power 4 kW4 kW 2 kW 2 kW
Frequency
Frequency 50 Hz
50 Hz 50 Hz 50 Hz
Voltage 230/400 V 230/400 V
Voltage 230/400 V 230/400 V
Current 13.94/8.04 A 6.97/4.03 A
NumberCurrent
of revolutions 13.94/8.04
1420 rpm A 14106.97/4.03 A
Number
Numberofof
revolutions
poles 1420
4 rpm 4 1410
Number of poles 4 4
4.1. Datasets
4.1. Datasets
In this research, there was no starting dataset available, so the one from the Laboratory
of Intelligent Automation
In this research, of Processes
there and Systems
was no starting andavailable,
dataset Laboratorysoofthe
Intelligent
one fromControl
the Labora-
of Electrical Machines, School of Engineering of São Carlos of the University
tory of Intelligent Automation of Processes and Systems and Laboratory of of São Paulo
Intelligent
(USP), Brazil [57], is used to train and test the machine learning models.
Control of Electrical Machines, School of Engineering of São Carlos of the University of
São
4.1.1.Paulo (USP), Brazil [57], is used to train and test the machine learning models.
Raw Data
Their setup consists of a three-phase induction motor coupled with a direct-current
4.1.1. Rawwhich
machine, Data works as a generator connected by a shaft containing a rotary torque
wrench, simulating the load torque. In their laboratory they artificially produced rotor
broken bars faults and tested the motor from 12.5% to 100% load under a stationary state.
The USP dataset includes 10 tests for each of the 8 operational setups (loads every
5 Nm from 5 Nm to 40 Nm) for 5 states: a healthy motor or a motor with 1, 2, 3 or 4 broken
rotor bars. Eleven variables are measured for 19 s:
 Nm from 5 Nm to 40 Nm) for 5 states: a healthy motor or a motor with 1, 2, 3 or 4 broken
rotor bars. Eleven variables are measured for 19 s:
• 3-phase motor current at 50 kHz sampling frequency.
Sensors 2023, 23, 2649
• 3-phase motor voltage at 50 kHz sampling frequency. 13 of 20
• 5 vibration sensors at 7.6 kHz sampling frequency.

4.1.2.
•
Feature Extraction
3-phase motor current at 50 kHz sampling frequency.
• The raw
3-phase datavoltage
motor from the motor
at 50 current was
kHz sampling loaded in MATLAB, split into a four sec-
frequency.
•
onds 5 test,
vibration sensors at 7.6
and processed tokHz sampling
calculate the frequency.
power spectrum by the FFT. By considering the
characteristics of the motor tested in the laboratory of USP (with a nominal speed of 1715
4.1.2. Feature Extraction
rpm and four poles), it is possible to calculate the synchronous speed and the slip. As a
The
result, the raw data from
relevant the motor according
sidebands, current wastoloaded in MATLAB,
literature alerts ofsplit into arotor
broken four seconds
bar, are shown
test, and processed to calculate the power spectrum by the FFT. By considering the charac-
in Table 2.
teristics of the motor tested in the laboratory of USP (with a nominal speed of 1715 rpm and
four poles), it is possible to calculate the synchronous speed and the slip. As a result, the
Table 2. Sidebands of interest for developing the classification model.
relevant sidebands, according to literature alerts of broken rotor bar, are shown in Table 2.
K=1 K=2 K=3 K=4 K=5
2. Sidebands of interest
TableUpper 65.67 for developing the
71.33 classification model.
77.00 82.67 88.33
Lower 54.33
K=1 K=2 48.67 K=3 43.00 K=4 37.33 K=5 31.67
Upper 65.67 71.33 77.00 82.67 88.33
Then,
Lower features54.33
were calculated48.67 following43.00
the indications in Section 3.2.2
37.33 31.67 in the range
from 4.5 Hz to 94.5 Hz. This range encompasses all sidebands and mitigates the inaccura-
cies that
Then,arise in the
features extreme
were endsfollowing
calculated of the FFTthecalculation, particularly
indications in Section 3.2.2when
in thewindowing
range is
utilized.
from 4.5 HzThe final
to 94.5 Hz.dataset consists
This range of 12 features
encompasses andand
all sidebands anmitigates
additional column indicating
the inaccuracies
whether
that arise the machine
in the extremeis ends
in a healthy
of the FFT(1) calculation,
or faulty (−1) state. It comprises
particularly 474 cases,
when windowing is equally
divided
utilized. between
The final healthy and faulty
dataset consists of 12cases (237 and
features cases
aneach). The data
additional is organized
column indicating as a 475
×whether the machine
13 matrix with 475isrows
in a healthy
and 13(1) or faulty (−1) state. It comprises 474 cases, equally
columns.
divided between healthy and faulty cases (237 cases each). The data is organized as a
475 ×
4.2. 13 matrix
Model with 475 rows and 13 columns.
Development
Python
4.2. Model programming language and the sci-kit-learn library in the Kaggle environ-
Development
mentPython
were programming
used to develop the models,
language which were
and the sci-kit-learn laterindeployed
library the Kaggleon the Arduino plat-
environment
form. Decision Trees (DT), Random Forest (RF), and Support Vector
were used to develop the models, which were later deployed on the Arduino platform. Machine (SVM) mod-
els wereTrees
Decision trained using
(DT), Randomthe Forest
created dataset.
(RF), The dataset
and Support was divided
Vector Machine (SVM)into twowere
models parts, with
trained
355 using
cases the created
utilized dataset. (which
for training The dataset was divided
corresponds into two
to 75% parts,
of the with 355data)
available casesand 119
utilizedreserved
cases for training
for(which corresponds
testing to 75% 25%).
(the remaining of the available data) and
The selection of 119 cases reserved
training and test set in-
for testing (the remaining 25%). The selection of training and test set instances has been
stances has been uniformly performed. The aim of the models is to classify new observa-
uniformly performed. The aim of the models is to classify new observations as healthy or
tions as healthy or faulty depending on the 12 presented features, as Figure 10 shows.
faulty depending on the 12 presented features, as Figure 10 shows.

Features Healthy/Faulty

Figure 10.Input
Figure 10. Inputand
and output
output of the
of the models.
models.

After all the classification algorithms were applied to the test dataset, the classification
metrics listed in Section 2.3.4 were derived from the confusion matrix data for each model,
shown in Table 3. The results of these metrics are presented in Table 4. It is worth
mentioning that the best results obtained with SVM are based on the radial kernel with a
regularization parameter, C, equals 1000 and with no maximum leaves defined for DT.
Sensors 2023, 23, 2649 14 of 20

Table 3. Comparison of classifier performance using a confusion matrix.

DT RF SVM
TP 60 53 44
TN 55 51 44
FP 2 9 18
FN 2 6 13

Table 4. Comparison of classifier performance using evaluation metrics.

DT RF SVM
Accuracy 0.958 0.866 0.714
Recall 0.952 0.895 0.755
Precision 0.967 0.836 0.656
F1 0.959 0.864 0.702

In order to validate these metrics, each model was subjected to a K-fold cross-validation
test, yielding consistent results.
It is observed that the DT model provides the highest accuracy, correctly classifying
96.6% of the observation, followed by RF at 87.4% and SVM at 73.9%. With regard to
recall, the DT model demonstrates the best performance. Recall measures the ability to
correctly predict positive class samples. In certain scenarios, it is imperative to avoid
misclassifying any positive class, such as in the case of misclassifying cancer samples. DT
achieves a recall score of 96.8%, followed by the RF model at 89.8% and the SVM model at
77.2%. Precision reflects the model’s ability to correctly identify positive predictions, such
as avoiding labeling a regular email as spam. In terms of precision, SVM and RF models
perform with a score of 71.0% and 85.5%, respectively. The DT model is ahead with a score
of 96.8%. Ultimately, the F1 score measures the overall performance of the models, and the
DT model attains the highest score of 96.8%, followed by RF at 87.6% and SVM at 73.9%.
Throughout the process of training and testing the classification model, we have
observed that the RMS value of the signal provides precise results. Notably, an accuracy of
nearly 100% was achieved when using the RMS value. However, it should be noted that
this may not be applicable to motors with differing characteristics. Therefore, since the goal
is to find a versatile model that can be used with different types of motors, an equivalent
value, RRMSE, has been used, as was presented in Section 3.2.2.

4.3. Implementation of Resource-Constrained Platform

Given that we are working with the resource-constrained platform, Arduino, it is
important to optimize the current measurement for better performance and to ensure proper
processing of the FFT signal. In order to achieve this, three variables must be optimized:
sampling frequency, number of samples, and frequency resolution. The sampling frequency
is restricted by the Shannon–Nyquist frequency and by the range of frequencies associated
with the fault to be detected. According to the equations described in Section 2.2 and the
motor characteristics presented in Table 1, this range can go up to 618.00 Hz (bearing fault),
168.33 Hz (eccentricity) or 80.00 Hz (broken rotor bar).
The number of samples is limited by the memory resources of the Arduino board.
The Arduino Uno has 2 KB of RAM and can store a maximum of 128 samples, while the
Arduino Mega, with 8 KB of RAM, can store up to 512 samples.
The frequency resolution refers to the spacing between two consecutive frequencies in
the power spectrum and, therefore, the frequencies that would be detected. It is calculated
as follows:
f sampling
Res f = (15)
samples
There are two options to improve the resolution: use a smaller fsampling or increase
the number of samples. Both are restricted by the characteristics already commented on.
Sensors 2023, 23, 2649 15 of 20

Table 5 shows the achievable resolutions for measuring up to the fifth sideband frequency
of the lathe and milling machine for the three types of faults discussed, using either the
Arduino Uno (with 128 samples) or the Arduino Mega (with 512 samples).

Table 5. Resolution frequency for each type of fault using the Arduino Uno or the Arduino Mega for
the lathe (L) and milling machine (M).

Max Frequency Sampling Freq. Num. of Samples Resolution

618.00 Hz (L) 128 9.66
Bearing fault 1236.00
614.00 Hz (M) 512 2.41
168.33 Hz (L) 128 2.63
Eccentricity 336.67
167.5 Hz (M) 512 0.66
73.61 Hz (L) 128 1.15
Broken rotor bar 147.22
80.00 Hz (M) 512 0.29

Based on the results shown in Table 5, it can be observed that it is not possible to carry
out the bearing and eccentricity fault diagnosis using the Arduino Uno, as the resolution
frequency is quite poor. Using the Arduino Mega, it is possible to carry out the three types
of fault diagnosis.
The micromlgen library was utilized to export the classification models from Python to
C code. Table 6 summarizes the impact of each model on Arduino. It can be observed that
RF has many more code lines than the rest. This is because it calculates 100 DT models and
provides the mean value of all of them. SVM has a bigger impact on program space, using
13% more space than the other due to the kernel calculation conducted. Dynamic memory
is the same in all the models, as it is related to the FFT calculations. As the results in
Tables 4–6 show, we can conclude that DT is the best model to be implemented on Arduino.

Table 6. Impact of the models on Arduino sketch.

Code Lines Program Space Dynamic Memory

DT 133 6924 bytes (2%) 4341 bytes (52%)
RF 26,571 6924 bytes (2%) 4341 bytes (52%)
SVM 224 40,294 bytes (15%) 4341 bytes (52%)

A graphical representation of the DT model with the decisions made to classify a new
observation is shown in Figure 11. It is worth noting the importance that the RRMSE value
takes, appearing in 9 out of the 22 decisions made by the decision tree. The theory would
suggest that the presence of sidebands is the determining factor, but in this case, it is not
as significant.

4.4. Real Case

After developing and programming the classification model in the Arduino Mega, a
series of tests were conducted on the lathe and milling machine under varying loads to
verify their optimal operating conditions. The lathe was tested under no-load conditions,
turning 1 mm and 2 mm. The milling machine was tested under no-load conditions, milling
0.1 mm and 1 mm.
Arduino measured the current feeding the motor, calculated the FFT of the signal,
extracted and computed the features, and executed the classification model. After that,
Arduino concluded that the lathe was faulty, while the milling machine was healthy.
In order to check these results, the process followed by Arduino was replicated in
the laboratory. Debug codes were loaded into the Arduino to download the values of the
current measurements and the locally calculated FFTs. From there, features were extracted,
and the decision tree model was applied.
Sensors 2023, 23,
Sensors 2023, 23, 2649
x FOR PEER REVIEW 16 of
16 of 20
20

Figure 11. Decision tree model developed.

4.4. Real Case

After developing and programming the classification model in the Arduino Mega, a
series of tests were conducted on the lathe and milling machine under varying loads to
verify their optimal operating conditions. The lathe was tested under no-load conditions,
turning 1 mm and 2 mm. The milling machine was tested under no-load conditions, mill-
ing 0.1 mm and 1 mm.
Arduino measured the current feeding the motor, calculated the FFT of the signal,
extracted and computed the features, and executed the classification model. After that,
Arduino concluded that the lathe was faulty, while the milling machine was healthy.
In order to check these results, the process followed by Arduino was replicated in the
laboratory. Debug codes were loaded into the Arduino to download the values of the cur-
rent measurements
Figure 11. Decision treeand thedeveloped.
model locally calculated FFTs. From there, features were extracted,
Figure
and the 11. Decisiontree
decision tree model
model developed.
was applied.
4.4.Figure
Real 12
Case
Figure 12 shows the power
shows the powerspectrum
spectrum of of
thethe
lathelathe
andand milling
milling machine
machine duringduring
tests tests
with a higher
with After load.
developing
a higher There
load. There are
andare no noticeable
programming
no noticeablethe peaks in the
classification
peaks sidebands
model in
in the sidebands ofthe of either
Arduino
either machine.
Thea The
Mega,
machine.
windowing
windowing effect
series of tests were
effect can
can beobserved
be observed
conducted atlathe
onatthe
thethe ends
endsand offigures,
the figures,
milling
of the machine leading
leadingunder to inaccurate
varying
to inaccurate tovalues
loads in
values
verify
inthese their
theseregions,optimal
regions, as as operating
mentioned
mentioned conditions.
before.
before. Those The
Those lathe
values arewas
values tested
are under no-load conditions,
discarded.
discarded.
turning 1 mm and 2 mm. The milling machine was tested under no-load conditions, mill-
ing 0.1 mm and 1 mm.
Arduino measured the current feeding the motor, calculated the FFT of the signal,
extracted and computed the features, and executed the classification model. After that,
Arduino concluded that the lathe was faulty, while the milling machine was healthy.
In order to check these results, the process followed by Arduino was replicated in the
laboratory. Debug codes were loaded into the Arduino to download the values of the cur-
rent measurements and the locally calculated FFTs. From there, features were extracted,
and the decision tree model was applied.
Figure 12 shows the power spectrum of the lathe and milling machine during tests
with a higher load. There are no noticeable peaks in the sidebands of either machine. The
windowing effect can be observed at the ends of the figures, leading to inaccurate values
in these regions, as mentioned before. Those values are discarded.
Figure
Figure12.
12.Power
Powerspectrum ofthe
spectrum of thelathe
lathe(left)
(left) and
and thethe milling
milling machine
machine (right)
(right) tested.tested.

Thefirst
The firststep
step to
to extracting
extractingthe
thefeatures is to
features is determine the relevant
to determine sidebands
the relevant of the of the
sidebands
lathe and the milling machine. The sidebands are shown in Table
lathe and the milling machine. The sidebands are shown in Table 7. 7.

Table 7. Sidebands of interest for lathe and milling machine testing.

K=1 K=2 K=3 K=4 K=5

Upper 55.33 60.67 66.00 71.33 76.67
Lathe
Lower 44.67 39.33 34.00 28.67 23.33
Milling Upper 56.00 62.00 68.00 74.00 80.00
machine Lower 44.00 38.00 32.00 26.00 20.00
Figure 12. Power spectrum of the lathe (left) and the milling machine (right) tested.
For both machines, the features defined in Section 3.2.2 and shown in Tables 8 and 9
were The first step
extracted. to extracting
Based the features
on these values, is to determine
it is possible the relevant
to reproduce the stepssidebands of the
followed by the
lathe and the milling machine. The sidebands are shown in Table 7.
decision tree model. In the case of the lathe, the first leaf compares whether f5d is less than
or equal to 49.6737. This first comparison is true, so the process continues to the second
leaf. In the second leaf, f5d is compared to 43.5598, and the result is false. The next leaf
compares whether f3d is less than or equal to 51.549. In this case, the result is true, so the
Sensors 2023, 23, 2649 17 of 20

model classifies the machine as faulty. This result was completely unexpected since it was
thought that the machine was in proper condition. The same result is achieved using data
from other tests in which the load is varied.

Table 8. Extracted features for lathe classification.

Feature Value Feature Value Feature Value

f1d 40.2260 f3d 45.8927 f5d 45.9219
f1u 38.1941 f3u 49.5721 f5u 46.4160
f2d 35.5643 f4d 51.9447 THD 1.0342
f2u 38.8301 f4u 47.1394 RRMSE 0.6803

Table 9. Extracted features for milling machine classification.

Feature Value Feature Value Feature Value

f1d 45.8208 f3d 40.9054 f5d 52.3244
f1u 42.2207 f3u 43.6715 f5u 49.1608
f2d 47.9009 f4d 46.8002 THD 1.0255
f2u 49.6617 f4u 48.6237 RRMSE 0.4917

In the milling machine classification, the process begins similarly to the previous case
by comparing if f5d is less than or equal to 49.6737. The comparison result is negative, so
we proceed to the leaf where it is compared if RRMSE is less than or equal to 0.5573, which
is true. We then continue to compare if f5d is less than or equal to 65.5342, which is correct,
and therefore it is determined that the machine is healthy.
Following the successful implementation of the models on the Arduino 2560, it has
been demonstrated that platforms with limited resources, such as Arduino, can execute
machine learning models to undertake fault diagnosis tasks.

5. Conclusions
In this paper, an edge computing approach for fault diagnosis is proposed. The
system has been designed as an Integrated Cyber-Physical System (ICPS) embracing the
physical and the cyber world. Both parts have been developed using the Arduino platform,
specifically the Arduino Mega board.
A classification method was developed by utilizing a publicly available dataset con-
taining information on a three-phase inductive motor under both healthy and broken bar
conditions. The purpose was to create a model that could be applied to other three-phase
induction motors with varying features. For this, novel features were defined based on
well-known MCSA sidebands. However, during the development of the model, it was
discovered that the sidebands are not as crucial to the classifiers as previously believed.
The research has shown that machine learning models, including Decision Trees,
Random Forests, and Support Vector Machines, can be executed satisfactorily on resource-
constrained platforms such as the Arduino Mega. However, due to memory limitations,
the amount of data that can be stored is limited, which in turn limits the features that can
be extracted using the Fast Fourier Transform.
The process of feature extraction, classification, and model training and testing has
been presented and implemented in Arduino Mega, testing two electric motors in the
laboratories of the Industrial and Mining School of Almadén.
The proposed solution is cost-effective for small to medium-sized enterprises, enabling
them to embrace the digitalization era and reap its benefits.
Further research is aimed at incorporating more complex classifiers that draw infor-
mation from multiple sensors, not only to detect faults but also to make prognoses.
Sensors 2023, 23, 2649 18 of 20

Author Contributions: Conceptualization, J.d.l.M., F.M.-F. and J.A.L.-G.; methodology, J.d.l.M. and
J.A.L.-G.; software, J.d.l.M. and J.A.L.-G.; validation, J.d.l.M., F.M.-F. and J.A.L.-G.; formal analysis,
J.d.l.M. and J.A.L.-G.; investigation, J.d.l.M.; resources, J.d.l.M., F.M.-F. and J.A.L.-G.; data curation,
J.d.l.M.; writing—original draft preparation, J.d.l.M. and F.M.-F.; writing—review and editing, J.d.l.M.
and J.A.L.-G.; visualization, J.d.l.M.; supervision, J.d.l.M. and F.M.-F.; project administration, J.d.l.M.
and F.M.-F.; funding acquisition, J.d.l.M. and F.M.-F. All authors have read and agreed to the published
version of the manuscript.
Funding: This research was funded by the 2022 Departmental Research Funds of the Department of
Electrical, Electronic, Automatic and Communications Engineering at UCLM.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Marcu, I.; Suciu, G.; Bălăceanu, C.; Vulpe, A.; Drăgulinescu, A.-M. Arrowhead technology for digitalization and automation
solution: Smart cities and smart agriculture. Sensors 2020, 20, 1464. [CrossRef] [PubMed]
2. Singh, R.; Akram, S.V.; Gehlot, A.; Buddhi, D.; Priyadarshi, N.; Twala, B. Energy System 4.0: Digitalization of the energy sector
with inclination towards sustainability. Sensors 2022, 22, 6619. [CrossRef]
3. Wang, J.; Xu, Y. How Does Digitalization Affect Haze Pollution? The Mediating Role of Energy Consumption. Int. J. Environ. Res.
Public Health 2022, 19, 11204. [CrossRef] [PubMed]
4. Nguyen, T.; Gosine, R.; Warrian, P. A review of the role of digitalization in health risk management in extractive industries—A
study motivated by COVID-19. J. Eng. Des. Technol. 2022, 20, 475–496. [CrossRef]
5. Cañas, V.; García, A.; Blanco, J.; de las Morenas, J. The Internet of Things Applied to the Automotive Sector: A Unified Intelligent
Transport System Approach. In Service Orientation in Holonic and Multi-Agent Manufacturing. Studies in Computational Intelligence;
Springer: Cham, Switzerland, 2016; Volume 640, pp. 53–60. [CrossRef]
6. Ministerio de Industria. Comercio y Turismo Portal Industria Conectada 4.0. Available online: https://www.industriaconectada4
0.gob.es/ (accessed on 10 January 2023).
7. Red.es Transformación Digital en Pymes. Available online: https://www.red.es/redes/es/que-hacemos/transformaci%C3%B3
n-digital-en-pymes/oficinas-de-transformaci%C3%B3n-digital (accessed on 10 January 2023).
8. Ministerio de Trabajo y Economía Social Estadística de empresas inscritas en la Seguridad Social. Available online: https:
//www.mites.gob.es/estadisticas/Emp/welcome.htm (accessed on 10 January 2023).
9. International Energy Agency. Energy Eficiency 2020; IEA Publications: Paris, France, 2020.
10. Frosini, L. Novel Diagnostic Techniques for Rotating Electrical Machines—A Review. Energies 2020, 13, 5066. [CrossRef]
11. Liu, Z.; Zhang, P.; He, S.; Huang, J. A Review of Modeling and Diagnostic Techniques for Eccentricity Fault in Electric Machines.
Energies 2021, 14, 4296. [CrossRef]
12. Choudhary, A.; Goyal, D.; Shimi, S.L.; Akula, A. Condition Monitoring and Fault Diagnosis of Induction Motors: A Review. Arch.
Comput. Methods Eng. 2019, 26, 1221–1238. [CrossRef]
13. Yakhni, M.F.; Cauet, S.; Sakout, A.; Assoum, H.; Etien, E.; Rambault, L.; El-Gohary, M. Variable speed induction motors’ fault
detection based on transient motor current signatures analysis: A review. Mech. Syst. Signal Process. 2023, 184, 109737. [CrossRef]
14. Asad, B.; Vaimann, T.; Rassõlkin, A.; Kallaste, A.; Belahcen, A. Review of Electrical Machine Diagnostic Methods Applicability in
the Perspective of Industry 4.0. Electr. Control Commun. Eng. 2018, 14, 108–116. [CrossRef]
15. Hadi Ashraf, R.; Karolina, K.; Bilal, A.; Toomas, V. Fault Detection and Predictive Maintenance of Electrical Machines. In New
Trends in Electric Machines—Technology and Applications; Miguel Delgado, P., José Alfonso, A.-D., Roque, A.O.-R., Eds.; IntechOpen:
Rijeka, Croatia, 2022; Chapter 4.
16. Duque, Ó.; Pérez, M. Motores de Inducción. Técnicas de Mantenimiento Predictivo; @becedario: Badajoz, Spain, 2005.
17. Khechekhouche, A.; Cherif, H.; Benakcha, A.; Menacer, A.; Chehaidia, S.E.; Panchal, H. Experimental diagnosis of inter-turns
stator fault and unbalanced voltage supply in induction motor using MCSA and DWER. Period. Eng. Nat. Sci. 2020, 8, 1202–1216.
18. Kompella, K.C.D.; Rongala, N.S.; Rayapudi, S.R.; Mannam, V.G.R. Robustification of fault detection algorithm in a three-phase
induction motor using MCSA for various single and multiple faults. IET Electr. Power Appl. 2021, 15, 593–615. [CrossRef]
19. Glowacz, A.; Glowacz, W.; Kozik, J.; Piech, K.; Gutten, M.; Caesarendra, W.; Liu, H.; Brumercik, F.; Irfan, M.; Khan, Z.F. Detection
of deterioration of three-phase induction motor using vibration signals. Meas. Sci. Rev. 2019, 19, 241–249. [CrossRef]
20. Sudhakar, I.; AdiNarayana, S.; AnilPrakash, M. Condition Monitoring of a 3-Ø Induction Motor by Vibration Spectrum anaylsis
using Fft Analyser—A Case Study. Mater. Today Proc. 2017, 4, 1099–1105. [CrossRef]
21. Fu, P.; Wang, J.; Zhang, X.; Zhang, L.; Gao, R.X. Dynamic Routing-based Multimodal Neural Network for Multi-sensory Fault
Diagnosis of Induction Motor. J. Manuf. Syst. 2020, 55, 264–272. [CrossRef]
Sensors 2023, 23, 2649 19 of 20

22. Choudhary, A.; Goyal, D.; Letha, S.S. Infrared Thermography-Based Fault Diagnosis of Induction Motor Bearings Using Machine
Learning. IEEE Sens. J. 2021, 21, 1727–1734. [CrossRef]
23. Khanjani, M.; Ezoji, M. Electrical fault detection in three-phase induction motor using deep network-based features of thermo-
grams. Measurement 2021, 173, 108622. [CrossRef]
24. Singh, G.; Anil Kumar, T.C.; Naikan, V.N.A. Induction motor inter turn fault detection using infrared thermographic analysis.
Infrared Phys. Technol. 2016, 77, 277–282. [CrossRef]
25. Mustafa, M.O.; Varagnolo, D.; Nikolakopoulos, G.; Gustafsson, T. Detecting broken rotor bars in induction motors with model-
based support vector classifiers. Control Eng. Pract. 2016, 52, 15–23. [CrossRef]
26. Gao, H.; Wang, Z.; Han, C.; Tang, A.; Guo, F.; Li, B. Feature Extraction Method of Series Arc Fault Occurred in Three-Phase Motor
With Inverter Circuit. IEEE Trans. Power Electron. 2022, 37, 11164–11173. [CrossRef]
27. Antonino-Daviu, J. Electrical Monitoring under Transient Conditions: A New Paradigm in Electric Motors Predictive Maintenance.
Appl. Sci. 2020, 10, 6137. [CrossRef]
28. Cherif, B.D.E.; Chouai, M.; Seninete, S.; Bendiabdellah, A. Machine-Learning-Based Diagnosis of an Inverter-Fed Induction Motor.
IEEE Lat. Am. Trans. 2022, 20, 901–911. [CrossRef]
29. Bellini, A.; Filippetti, F.; Tassoni, C.; Capolino, G.A. Advances in Diagnostic Techniques for Induction Machines. IEEE Trans. Ind.
Electron. 2008, 55, 4109–4126. [CrossRef]
30. Deekshit, K.K.C.; Rao, M.V.G.; Rao, R.S. Fault Indexing Parameter Based Fault Detection in Induction Motor via MCSA with
Wiener Filtering. Electr. Power Compon. Syst. 2021, 48, 2048–2062. [CrossRef]
31. Lee, S.B.; Hyun, D.; Kang, T.J.; Yang, C.; Shin, S.; Kim, H.; Park, S.; Kong, T.S.; Kim, H.D. Identification of False Rotor Fault
Indications Produced by Online MCSA for Medium-Voltage Induction Machines. IEEE Trans. Ind. Appl. 2016, 52, 729–739.
[CrossRef]
32. Yassa, N.; Rachek, M.; Houassine, H. Motor current signature analysis for the air gap eccentricity detection in the squirrel cage
induction machines. Energy Procedia 2019, 162, 251–262. [CrossRef]
33. Zheng, X.; Inoue, H.; Kanemaru, M.; Liu, D. Eccentricity Severity Estimation of Induction Machines Using a Sparsity-Driven
Regression Model. In Proceedings of the 2022 IEEE Energy Conversion Congress and Exposition (ECCE), Detroit, MI, USA, 9–13
October 2022; pp. 1–6.
34. Dutta, S.; Basu, B.; Talukdar, F.A. Classification of Induction Motor Fault and Imbalance Based on Vibration Signal Using Single
Antenna’s Reactive Near Field. IEEE Trans. Instrum. Meas. 2021, 70, 1–9. [CrossRef]
35. Zhang, S.; Wang, B.; Kanemaru, M.; Lin, C.; Liu, D.; Miyoshi, M.; Teo, K.H.; Habetler, T.G. Model-based analysis and quantification
of bearing faults in induction machines. IEEE Trans. Ind. Appl. 2020, 56, 2158–2170. [CrossRef]
36. Hassan, O.E.; Amer, M.; Abdelsalam, A.K.; Williams, B.W. Induction motor broken rotor bar fault detection techniques based on
fault signature analysis–a review. IET Electr. Power Appl. 2018, 12, 895–907. [CrossRef]
37. Benbouzid, M.E.H. A review of induction motors signature analysis as a medium for faults detection. IEEE Trans. Ind. Electron.
2000, 47, 984–993. [CrossRef]
38. SKF Group. Bearing Damage and Failure Analysis. 2017. Available online: https://www.skf.com/binaries/pub12/Images/0901
d1968064c148-Bearing-failures---14219_2-EN_tcm_12-297619.pdf (accessed on 10 January 2023).
39. Kudelina, K.; Vaimann, T.; Asad, B.; Rassõlkin, A.; Kallaste, A.; Demidova, G. Trends and Challenges in Intelligent Condition
Monitoring of Electrical Machines Using Machine Learning. Appl. Sci. 2021, 11, 2761. [CrossRef]
40. Zhang, W.; Jia, M.-P.; Zhu, L.; Yan, X.-A. Comprehensive Overview on Computational Intelligence Techniques for Machinery
Condition Monitoring and Fault Diagnosis. Chin. J. Mech. Eng. 2017, 30, 782–795. [CrossRef]
41. Quiroz, J.C.; Mariun, N.; Mehrjou, M.R.; Izadi, M.; Misron, N.; Mohd Radzi, M.A. Fault detection of broken rotor bar in LS-PMSM
using random forests. Measurement 2018, 116, 273–280. [CrossRef]
42. Bazan, G.H.; Scalassara, P.R.; Endo, W.; Goedtel, A.; Palácios, R.H.C.; Godoy, W.F. Stator Short-Circuit Diagnosis in Induction
Motors Using Mutual Information and Intelligent Systems. IEEE Trans. Ind. Electron. 2019, 66, 3237–3246. [CrossRef]
43. Toma, R.N.; Prosvirin, A.E.; Kim, J.-M. Bearing Fault Diagnosis of Induction Motors Using a Genetic Algorithm and Machine
Learning Classifiers. Sensors 2020, 20, 1884. [CrossRef]
44. Ince, T.; Kiranyaz, S.; Eren, L.; Askar, M.; Gabbouj, M. Real-Time Motor Fault Detection by 1-D Convolutional Neural Networks.
IEEE Trans. Ind. Electron. 2016, 63, 7067–7075. [CrossRef]
45. Zhang, S.; Zhang, S.; Wang, B.; Habetler, T.G. Deep Learning Algorithms for Bearing Fault Diagnostics—A Comprehensive
Review. IEEE Access 2020, 8, 29857–29881. [CrossRef]
46. Hosseinpour, Z.; Arefi, M.M.; Mozafari, N.; Luo, H.; Yin, S. An Ensemble-Based Fuzzy Rough Active Learning Approach for
Broken Rotor Bar Detection in Nonstationary Environment. IEEE Trans. Instrum. Meas. 2022, 71, 1–8. [CrossRef]
47. Khan, M.A.; Asad, B.; Kudelina, K.; Vaimann, T.; Kallaste, A. The Bearing Faults Detection Methods for Electrical Machines—The
State of the Art. Energies 2023, 16, 296.
48. Kagermann, H.; Wahlster, W.; Helbig, J. Securing the Future of German Manufacturing Industry: Recommendations for Implementing the
Strategic Initiative Industrie 4.0; Technical Report; ACATECH German National Academy of Science and Engineering: Munich,
Germany, 2013; Volume 4.
Sensors 2023, 23, 2649 20 of 20

49. Barbosa, J.; Leitão, P.; Trentesaux, D.; Colombo, A.W.; Karnouskos, S. Cross benefits from cyber-physical systems and intelligent
products for future smart industries. In Proceedings of the 2016 IEEE 14th International Conference on Industrial Informatics
(INDIN), Poitiers, France, 19–21 July 2016; pp. 504–509.
50. Karnouskos, S.; Leitao, P.; Ribeiro, L.; Colombo, A.W. Industrial Agents as a Key Enabler for Realizing Industrial Cyber-Physical
Systems: Multiagent Systems Entering Industry 4.0. IEEE Ind. Electron. Mag. 2020, 14, 18–32. [CrossRef]
51. Monostori, L. Cyber-physical production systems: Roots, expectations and R&D challenges. Procedia CIRP 2014, 17, 9–13.
52. Wollschlaeger, M.; Sauter, T.; Jasperneite, J. The Future of Industrial Communication: Automation Networks in the Era of the
Internet of Things and Industry 4.0. IEEE Ind. Electron. Mag. 2017, 11, 17–27. [CrossRef]
53. Dekhandji, F.Z.; Halledj, S.E.; Zaboub, O. Predictive Maintenance Applied to Three phase Induction Motors. Alger. J. Signals Syst.
2019, 4, 71–88. [CrossRef]
54. Jove, E.; González-Cava, J.M.; Casteleiro-Roca, J.-L.; Alaiz-Moretón, H.; Baruque, B.; Leitão, P.; Méndez Pérez, J.A.; Calvo-Rolle,
J.L. An intelligent system for harmonic distortions detection in wind generator power electronic devices. Neurocomputing 2021,
456, 609–621. [CrossRef]
55. Aqueveque, P.; Radrigan, L.; Pastene, F.; Morales, A.S.; Guerra, E. Data-Driven Condition Monitoring of Mining Mobile Machinery
in Non-Stationary Operations Using Wireless Accelerometer Sensor Modules. IEEE Access 2021, 9, 17365–17381. [CrossRef]
56. Markiewicz, M.; Wielgosz, M.; Bocheński, M.; Tabaczyński, W.; Konieczny, T.; Kowalczyk, L. Predictive Maintenance of
Induction Motors Using Ultra-Low Power Wireless Sensors and Compressed Recurrent Neural Networks. IEEE Access 2019,
7, 178891–178902. [CrossRef]
57. Treml, A.E.; Flauzino, R.A.; Suetake, M.; Maciejewski, N.A.R. Experimental Database for Detecting and Diagnosing Rotor Broken
Bar in a Three-Phase Induction Motor. IEEE Dataport 2022. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.