0% found this document useful (0 votes)
21 views23 pages

Sensors 24 02297

This article presents a meta ensemble model for industrial fault detection using ultrasonic signals, focusing on pipelines and motors. The study employs various machine learning techniques, including feature extraction and dimensionality reduction methods, to enhance predictive accuracy, achieving approximately 5% improvement in classification performance. The proposed model demonstrates potential for real-time monitoring with efficient execution on microcontroller units.

Uploaded by

salva3.ndongo
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
21 views23 pages

Sensors 24 02297

This article presents a meta ensemble model for industrial fault detection using ultrasonic signals, focusing on pipelines and motors. The study employs various machine learning techniques, including feature extraction and dimensionality reduction methods, to enhance predictive accuracy, achieving approximately 5% improvement in classification performance. The proposed model demonstrates potential for real-time monitoring with efficient execution on microcontroller units.

Uploaded by

salva3.ndongo
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 23

sensors

Article
Industrial Fault Detection Employing Meta Ensemble Model
Based on Contact Sensor Ultrasonic Signal
Amirhossein Moshrefi 1, * , Hani H. Tawfik 2 , Mohannad Y. Elsayed 2 and Frederic Nabki 1

1 Department of Electrical Engineering, Ecole de Technologie Supérieure, ETS, Montreal, QC H3C 1K3, Canada;
[email protected]
2 MEMS-Vision International Inc., Montreal, QC H4P 2R9, Canada; [email protected] (H.H.T.);
[email protected] (M.Y.E.)
* Correspondence: [email protected]

Abstract: Ultrasonic diagnostics is the earliest way to predict industrial faults. Usually, a contact mi-
crophone is employed for detection, but the recording will be contaminated with noise. In this paper,
a dataset that contains 10 main faults of pipelines and motors is analyzed from which 30 different
features in the time and frequency domains are extracted. Afterward, for dimensionality reduction,
principal component analysis (PCA), linear discriminant analysis (LDA), and t-distributed stochastic
neighbor embedding (t-SNE) are performed. In the subsequent phase, recursive feature elimination
(RFE) is employed as a strategic method to analyze and select the most relevant features for the
classifiers. Next, predictive models consisting of k-Nearest Neighbor (KNN), Logistic Regression (LR),
Decision Tree (DT), Gaussian Naive Bayes (GNB), and Support Vector Machine (SVM) are employed.
Then, in order to solve the classification problem, a stacking classifier based on a meta-classifier
which combines multiple classification models is introduced. Furthermore, the k-fold cross-validation
technique is employed to assess the effectiveness of the model in handling new data for the evaluation
of experimental results in ultrasonic fault detection. With the proposed method, the accuracy is
around 5% higher over five cross folds with the least amount of variation. The timing evaluation of
the meta model on the 64 MHz Cortex M4 microcontroller unit (MCU) revealed an execution time of
11 ms, indicating it could be a promising solution for real-time monitoring.

Citation: Moshrefi, A.; Tawfik, H.H.; Keywords: fault detection; ultrasonic signal; feature extraction; meta classification; machine learning;
Elsayed, M.Y.; Nabki, F. Industrial real-time monitoring
Fault Detection Employing Meta
Ensemble Model Based on Contact
Sensor Ultrasonic Signal. Sensors 2024,
24, 2297. https://doi.org/10.3390/
1. Introduction
s24072297
In today’s industry, rapid growth has led to more widespread automated processes,
Academic Editor: Jongmyon Kim as well as an increased demand for advanced equipment and machines. Motor and
Received: 5 February 2024
pipeline spare parts have been increasingly used in machine maintenance due to the
Revised: 21 March 2024 advent of industrial automation. Meanwhile, preventing the breakdown of these types
Accepted: 28 March 2024 of equipment would be beneficial to reduce the cost and time of maintenance. There are
Published: 4 April 2024 some methods to detect industrial faults which have different characteristics. Industrial
equipment fault monitoring with technologies like vibration analysis and temperature
sensing has been available for many years. However, as demonstrated by the installation
potential failure (IPF) curve in Figure 1, ultrasound is alternatively an efficient technology
Copyright: © 2024 by the authors. to sense certain mechanical and electrical faults in order to predict failures much earlier
Licensee MDPI, Basel, Switzerland. than other monitoring technologies [1].
This article is an open access article In this study, common faults of pipes and motors are investigated. In order to transport
distributed under the terms and fluids from one place to another, pipes are widely used. Numerous pipe segments connected
conditions of the Creative Commons
with joints make up the pipeline networks that extend for several kilometers. Various
Attribution (CC BY) license (https://
factors, including traffic and surface loads, may cause the pipes and joints in these pipelines
creativecommons.org/licenses/by/
to be overstressed, resulting in leaks and pipe bursts. On the other hand, due to pipeline
4.0/).

Sensors 2024, 24, 2297. https://doi.org/10.3390/s24072297 https://www.mdpi.com/journal/sensors


Sensors 2024, 24, 2297 2 of 23

Sensors 2024, 24, x FOR PEER REVIEW defects [2] such as cracks, cavitation, corrosion, and other mechanical damage, the problem
2 of 24
of pipeline operation safety can become very difficult to manage [3,4].

Figure
Figure1.1.IPF
IPFcurve
curvefor
forindustrial
industrial fault predictiondelay
fault prediction delay[1].
[1].

InMoreover,
this study, common
bearings need to faults of pipes for
be considered anddiagnosing
motors are investigated.
rotating machineryInfaults orderor to
transport
diagnosing fluids
shaftfrom oneinplace
failures to another,
industrial pipes
processes. are widely
Bearings used. the
are among Numerous pipe seg-
most important
peripheral
ments mechanisms
connected in most
with joints developing
make up the countries and have to
pipeline networks operate
that extend under
for high loads
several kilo-
and speeds.
meters. VariousTheir failure,
factors, mostly due
including to incorrect
traffic lubricant
and surface selection,
loads, may causecontamination,
the pipes and lossjoints
of
inlubricant, or over-greasing,
these pipelines causes malfunctioning
to be overstressed, of theand
resulting in leaks machinery and shutdown,
pipe bursts. On the other which
hand,
in turn affects the quality and cost of the products [5].
due to pipeline defects [2] such as cracks, cavitation, corrosion, and other mechanical dam-
age, the Highlighting
problem ofthe symbiotic
pipeline relationship
operation safetybetween
can become motors
veryand pipelines
difficult in industrial
to manage [3,4].
settings is crucial for understanding their collective impact on system efficiency and safety.
Moreover, bearings need to be considered for diagnosing rotating machinery faults
These components frequently operate as part of a unified system across numerous industrial
or diagnosing shaft failures in industrial processes. Bearings are among the most im-
applications. For instance, in the chemical processing sector, motors power pumps that
portant peripheral mechanisms in most developing countries and have to operate under
facilitate the movement of fluids through pipelines. A malfunction in the motor could
high loads and speeds.
result in reduced pressureTheiror failure, mostly
flow within thedue to incorrect
pipelines, lubricantthe
compromising selection,
efficiencycontami-
and
nation, loss of lubricant,
safety of the process. or over-greasing, causes malfunctioning of the machinery and
shutdown, whichsimultaneous
Moreover, in turn affects thedetection
fault quality and cost motors
in both of the products [5]. fosters an all-
and pipelines
Highlighting
encompassing the symbiotic
approach relationship
to predictive between
maintenance. motors
Through and pipelines
diligent monitoring in of
industrial
these
settings is crucial for understanding their collective impact
components, industries can avert unplanned downtime, refine maintenance schedules, on system efficiency and
safety. Thesesystem
and bolster components frequently
reliability. operate
This strategy is as part of a unified
particularly vital in system across
sectors such asnumerous
oil and
gas, where
industrial the failure ofFor
applications. anyinstance,
component might
in the induce processing
chemical substantial production
sector, motors setbacks
power
and environmental risks. Additionally, the interconnected nature
pumps that facilitate the movement of fluids through pipelines. A malfunction in the mo- of motor and pipeline
failures—where
tor could result ina reduced
fault in one could or
pressure exacerbate issues
flow within theinpipelines,
the other—underscores
compromisingthe theim-
effi-
portance of understanding
ciency and safety of the process. these relationships for effective fault diagnosis and mitigation.
For Moreover,
example, a simultaneous
blockage in a pipeline might overload
fault detection in both amotors
motor, and
causing it to overheat
pipelines fosters and
an all-
fail, whereas motor failure could disrupt fluid flow, leading to increased pressure and
encompassing approach to predictive maintenance. Through diligent monitoring of these
strain on pipeline joints or seals. In scenarios where motors propel compressors or pumps
components, industries can avert unplanned downtime, refine maintenance schedules,
linked to pipelines, like HVAC systems in large buildings or water treatment facilities,
and bolster system reliability. This strategy is particularly vital in sectors such as oil and
maintaining the faultless operation of both elements is imperative for energy conservation
gas,
andwhere the failure of any component might induce substantial production setbacks
cost-efficiency.
and environmental risks. Additionally,
Adherence to regulatory the interconnected
and safety standards frequentlynature of motor
necessitates and pipeline
extensive moni-
failures—where
toring and fault adetection
fault in across
one could exacerbate
all crucial issues
industrial in the
system other—underscores
components. the im-
In the pharma-
portance of understanding these relationships for effective fault diagnosis
ceutical industry, preserving the integrity of motors and pipelines is essential for ensuring and mitigation.
For example, a blockage in a pipeline might overload a motor, causing it to overheat and
fail, whereas motor failure could disrupt fluid flow, leading to increased pressure and
strain on pipeline joints or seals. In scenarios where motors propel compressors or pumps
linked to pipelines, like HVAC systems in large buildings or water treatment facilities,
maintaining the faultless operation of both elements is imperative for energy conservation
Sensors 2024, 24, 2297 3 of 23

product sterility and quality. Any defects in these elements could jeopardize the entire
production cycle, potentially resulting in regulatory breaches and health risks. For example,
in the food and beverage industry, motors drive pumps and agitators that convey ingredi-
ents, mixtures, or final products through pipelines. Failures in these systems may cause
contamination, product wastage, or safety issues, highlighting the importance of unified
fault detection mechanisms [6].
To address the aforementioned faults, many approaches have been investigated in
the literature. In addition to ultrasound and vibration, other signals such as infrared and
audible noise as well as oil leakage can detect faults. As discussed before, among them,
ultrasound signals have the ability to distinguish industrial faults more efficiently [7].
However, in order to analyze ultrasonic signals, conventional methods like FFT [8], filter-
ing, and windowing [9,10], due to the overlapping of noise and signal, are not effective.
Alternatively, intelligent methods like machine learning (ML) can detect subtle changes in
the received signals from the ultrasonic sensor [1,11].
Based on the literature, many authors studied fault detection using artificial intelli-
gence (AI) algorithms by focusing on various derived signals. Quy et al. [3] introduced
a technique using a KNN classifier embedded in a microcontroller unit to classify the
leak faults in real-time. They extracted the hybrid features from acoustic emission signals
through a gas pipeline. Ding et al. [12] introduced an algorithm for detecting faults in
spacecraft, specifically focusing on four types of leaks with varying leakage conditions.
Their approach utilizes acoustic signals and combines SVM with empirical mode decom-
position (EMD) as the underlying technique. According to the assessment of a pipeline
by Coelho et al. [13], a system based on a wireless sensor network designed to monitor
distribution systems using a random forest (RF) algorithm was used to precisely locate
fluid leaks. They concluded that, by employing more features, accuracy could be increased.
Rai et al. [14] highlighted the insufficiency of historical data on pipeline failures for the
existing supervised AI methods. To address this challenge, they introduced an alternative
approach focused on a health index perspective. Their proposed method incorporates
multiscale analysis, the Kolmogorov–Smirnov (KS) test, and a Gaussian mixture model
(GMM) to accurately determine the leakage situation in pipelines.
Heng et al. [15] proposed a method that employs Principal Component Analysis
(PCA) to integrate data from multiple sensors for the accurate prediction of rolling bearing
lifespan. This approach significantly enhances reliability by amalgamating various features
from vibration signals, though it relies heavily on extensive lifecycle data for effective
modeling. A notable limitation of their work is its dependence on complete lifecycle
test data from rolling bearings to achieve precise modeling, necessitating comprehensive
testing to identify optimal sample sizes. This requirement potentially limits the method’s
application in scenarios where extensive historical data is unavailable, likely increasing
the time and cost associated with the predictive maintenance process. Additionally, this
method primarily focuses on time-domain features, which may restrict the depth of data
feature analysis and risk missing valuable insights from other domains, such as frequency.
Jose et al. [16] introduced a condition monitoring methodology for induction mo-
tors, analyzing motor stator currents and vibrations to estimate different features across
multiple domains. Utilizing genetic and PCA algorithms for feature selection, followed
by dimensionality reduction via the Linear Discriminant Analysis (LDA) algorithm, this
approach ultimately evaluates the refined features using a neural network classifier to
achieve global and individual classification ratios. However, the complexity of this method
and its intensive computational resource requirements are significant drawbacks. The
process involves multiple advanced stages, including high-dimensional feature estimation,
optimization with Genetic Algorithms and PCA, and dimensionality reduction through
LDA, before proceeding to fault diagnosis. This complexity not only necessitates consider-
able processing power but also complicates real-time application, potentially limiting its
practicality in scenarios where rapid and efficient diagnostics are essential.
Sensors 2024, 24, 2297 4 of 23

Research on rotating machinery fault diagnosis using ensemble kernel extreme learn-
ing machines based on stacked denoising autoencoders (SDAE) was explored [17]. This
study focused on gear, rotor, and real engine rolling bearing datasets, extracting vibration
characteristics from both time and frequency domains. The PCA algorithm was used to
merge two sets of SDAE features from multiple domains into essential low-dimensional
features, which were then classified for fault patterns using the ensemble kernel extreme
learning machine. Despite its innovative approach, this method encountered challenges
such as lengthy training times, complex implementation, high computational costs, and the
need for large amounts of labeled data for training the models.
Our contribution builds upon these insights, utilizing ultrasonic data for early de-
tection characterized by its minimal computational demands. This enables efficient de-
ployment on MCUs, offering quick responsiveness with reduced memory usage, thereby
addressing some of the limitations highlighted in the aforementioned studies.
Maliuk et al. [18] conducted research on detecting bearing faults through signal pro-
cessing and proposed another approach based on the Gaussian mixture model (GMM).
Their method leverages a fault-frequency-oriented GMM window series for reliable fea-
ture extraction. The classification step is performed using the weighted KNN algorithm.
As discussed before, besides pipeline faults in industrial equipment, rotating machinery
faults commonly occur in many different areas and manufacturing sectors. These include
aerospace, power generation, oil refining, machining, automotive, railway transportation,
pumping systems, etc. [6,19,20]. To prevent operational malfunctions that could potentially
result in catastrophic failures, various condition monitoring techniques have been devel-
oped for the purpose of fault detection and diagnosis in bearings. Abdelrhman et al. [21]
introduced a diagnosis and detection model for bearing faults in rotating machinery. Their
approach involves the utilization of multivariate analysis of variance to extract parameters
from acquired data sets. Time-domain features are employed, and a binary logistic regres-
sion (BLR) modeling technique is utilized for the fault diagnosis and detection. Similarly,
Jiang et al. [22] proposed a method for weak rotating machinery fault diagnosis. They
introduced a multiscale permutation entropy feature extraction approach, which involves
calculating time series with equal overlapping segments. The extracted features are then
used as input for classification using SVMs. To optimize the SVM parameters, the authors
introduced a chaos firefly optimization algorithm as a solution to the parameter optimiza-
tion problem. Another study by Souza et al. [23] proposed an SVM-based classifier by
employing statistical data features for bearing fault recognition. They investigated the
presence of different types of bearing defects and combinations of statistical attributes on
SVM accuracy. In another work by Jung et al. [24], using adaptive wavelet denoising and
statistical–spectral acoustic features, the authors performed a binary classification task to
monitor the health conditions of a ball bearing with a microphone under noisy conditions.
The authors used an adaptive wavelet method based on the kurtosis entropy index and
multiple acoustic features were extracted based on expert knowledge.
However, few works have investigated the use of an ensemble model in fault detection
though ultrasonic monitoring. In [25], we proposed a meta model based on the majority
voting of four classifiers for industrial faults. Accordingly, this study is an extension of
previous work and focuses on fault diagnosis using ultrasound signals by a meta ensemble
model. The research encompasses the classification of ten distinct types of faults in both
pipeline and rotating machinery conditions. Statistical features from both the time and
frequency domains are extracted by the PCA, LDA, and t-SNE approaches. Finally, the
features are refined using RFE. Then, five widely used classification models including
GNB, LR, KNN, SVM, and DT are implemented. Afterward, using an ensemble learning
approach, faults are classified into ten categories. Then, the reliability of the model based on
four evaluation metrics—k-fold validation, confusion matrix, ROC, and learning curve—is
investigated and, ultimately, its potential for real-time application is assessed through
deployment on an MCU.
Sensors 2024, 24, x FOR PEER REVIEW 5 of 24

Sensors 2024, 24, 2297 curve—is investigated and, ultimately, its potential for real-time application is assessed 5 of 23
through deployment on an MCU.
Our study makes significant contributions by meticulously preparing a dataset that
Our study
encapsulates 10makes
distinctsignificant
fault types, contributions
integrating bothby meticulously preparing
motor and pipeline a datasetInthat
anomalies. the
encapsulates 10 distinct
initial preprocessing fault we
phase, types, integrating
employ advanced bothtechniques
motor and to pipeline anomalies.
effectively In
eliminate
the initial preprocessing phase, we employ advanced techniques to effectively
noise, enhancing the data’s clarity. We then leverage feature reduction methods, including eliminate
noise,
RFE, to enhancing the data’s
meticulously selectclarity.
the most We impactful
then leverage feature
features forreduction
subsequent methods, including
analysis. During
RFE, to meticulously select the most impactful features for subsequent
the classification phase, we introduce a novel approach by combining conventional clas- analysis. During the
classification phase, we introduce a novel approach by combining conventional
sifiers into an innovative ensemble model. This model is rigorously evaluated through k- classifiers
into
foldan innovative ensemble
cross-validation, model. This
demonstrating its model
abilityistorigorously
accuratelyevaluated
categorize through k-fold
features withcross-
93%
validation, demonstrating its ability to accurately categorize features
accuracy. The culmination of our efforts is the deployment of this framework on an MCU, with 93% accuracy.
The
where culmination of our efforts
we successfully assess itsis the deployment
capability of this framework
for real-time monitoring. onThis
an MCU, where we
comprehensive
successfully assess its capability for real-time monitoring. This
approach not only showcases the framework’s high accuracy but also its practical comprehensive approach
applica-
not only showcases the framework’s high accuracy
bility in real-world industrial settings for fault detection. but also its practical applicability in
real-world industrial settings for fault detection.
The rest of this paper is organized as follows: preprocessing, feature extraction, and
methods rest
The of this paper
are explained is organized
in Section 2. Then,as Section
follows:3preprocessing, feature extraction,
presents the experimental results and
and
methods are explained in Section 2. Then, Section 3 presents the experimental results and
is followed by a conclusion.
is followed by a conclusion.
2. Proposed
2. Proposed Methodology
Methodology
The proposed
The proposedmethodology’s
methodology’soverviewoverviewis is illustrated
illustrated in Figure
in Figure 2. The
2. The entire
entire proce-
procedure
dure consists of five steps, each of which is explained in a separate
consists of five steps, each of which is explained in a separate subsection. The initialsubsection. The initial
step involves acquiring ultrasonic data from a microphone array
step involves acquiring ultrasonic data from a microphone array module, which is then module, which is then
recordedas
recorded asraw
rawdata
datafor
forfurther
furtherclassification
classification analysis.
analysis. The
The collected
collected datadata
fromfrom
thethe struc-
structure
ture is processed to refine and prepare it for the feature extraction stage.
is processed to refine and prepare it for the feature extraction stage. Feature extraction Feature extraction
is performed
is performed to to transform
transform the the data
data into
into aa set
set of
of numerical
numerical features
features known
known as as aa feature
feature
vector, which provides a concise and informative representation of
vector, which provides a concise and informative representation of the data. This study the data. This study
extracts a comprehensive suite of features from ultrasonic data,
extracts a comprehensive suite of features from ultrasonic data, spanning time, power spanning time, power
spectral density,
spectral density, andand frequency
frequency domains.
domains. These These features
features include
include minimum,
minimum, maximum,
maximum,
mean, zero
mean, zero crossing
crossing times,
times, slope
slope change,
change, impulse
impulse factor, shape factor,
factor, shape margin factor,
factor, margin factor, area
area
under curve, standard deviation, skewness, and kurtosis, offering
under curve, standard deviation, skewness, and kurtosis, offering a detailed analysis ofa detailed analysis of
the data’s
data’s characteristics.
characteristics. The presence of large input vectors often poses a challenge challenge in in
the predictive
predictive modeling
modeling process
process of of machine
machine learning
learning models.
models. This challenge is commonly
known as
known asthe
theproblem
problemofofhigh highdimensionality
dimensionality inin
rawraw data.
data. ToTo address
address this,this, a dimension-
a dimensionality
ality reduction
reduction technique
technique is employed
is employed as a solution.
as a solution. Finally,
Finally, a machine
a machine learning
learning classifica-
classification
tion algorithm
algorithm is applied
is applied to thetoreduced-dimensional
the reduced-dimensional data data forclassification
for the the classification
task. task.

Figure 2.
Figure 2. Overview
Overview of
of the
the workflow
workflow methodology.
methodology.

2.1. Preprocessing
To
To mitigate
mitigate noise
noise in
in ultrasonic
ultrasonic signals,
signals, one technique
technique involves
involves utilizing
utilizing a Butterworth
filter
filter in
inthe
thetime
timedomain,
domain,aimed
aimedatateradicating
eradicatingundesired
undesired spikes,
spikes,trends,
trends, and outliers.
and Butter-
outliers. But-
worth
terworth filters are are
filters favored in control
favored systems
in control for their
systems lack of
for their peaking,
lack ensuring
of peaking, a consistent
ensuring a con-
signal
sistentwithout unduly unduly
signal without amplifying any frequency.
amplifying These filters
any frequency. These arefilters
crafted
aretocrafted
sustaintoasus-
flat
frequency response across the passband, avoiding alterations in signal
tain a flat frequency response across the passband, avoiding alterations in signal ampli- amplitude within
this
tuderange
within[26].
thisThis characteristic
range is crucial for maintaining
[26]. This characteristic is crucial forthe integrity ofthe
maintaining theintegrity
ultrasonicof
signal’s
the ultrasonic signal’s information content. The nth order Butterworthdescribed
information content. The nth order Butterworth filter is mathematically filter is
Sensors 2024, 24, 2297 6 of 23

by its frequency response equation, encapsulating the filter’s design to achieve the desired
signal such that the following holds:

1
H( jω) = q (1)
ω 2n
1 + ε2 ( ω p
)

where H is the passband gain, ω is the angular frequency, ε is the maximum pass band
gain, and n represents the filter order.
The data values are normalized to be comparable for the different signal sets. Among
many different scalers available for feature scaling, the standard scaler in [27] is normalized
as follows:
X−µ
Z= (2)
σ
where the normalized vector, denoted as Z, is derived from the signal vector X. The
normalization process involves calculating the mean (µ) and standard deviation (σ) of the
input acoustic sequence X. These statistical measures are used to standardize the values in
X, resulting in the normalized vector Z.

2.2. Feature Extraction


Firstly, to provide the characteristics of each data, some features should be considered.
In this study, 30 features in time, Fourier, and power spectrum domains are extracted which
are composed of maximum, average, standard deviation, skewness, kurtosis, average of
peaks (peaks_mean), time of peak (peak_index), zero crossing, crest factor, and shape factor.
To distinguish between the classes with the most discernment, the independent and more
important features should be selected. The selected complex features are defined below:
1 N
∑ x ( n ) − x )3
Skewness = qN n=1 (3)
( N1 ∑nN=1 ( x (n) − x )2 )3

1 N
∑ x ( n ) − x )4
Kurtosis = qN n=1 (4)
( N1 ∑nN=1 ( x (n) − x )2 )4
max| x (n)|
Crest_ f actor = q (5)
1 N
N ∑ n =1 x ( n )2
q
1 N
N ∑ n =1 x ( n )
2
Shape_ f actor = 1 N
(6)
N ∑n=1 | x ( n )|
where N is the number of samples in the input vector x(n).

2.3. Dimensionality Reduction


In order to enhance the accuracy and efficiency of the model, it is essential to prune
every dataset. Particularly when dealing with ultrasonic signals, the dimensionality of the
data should be reduced. Dimensionality reduction involves transforming high-dimensional
data into a more concise and meaningful representation with reduced dimensions. One
of the prominent techniques for linear dimensionality reduction is PCA which identifies
orthogonal directions that capture the maximum variance, making it a valuable second-
order statistical method [28]. PCA is a highly effective approach for eliminating redundant
dimensions while retaining the most informative ones. This algorithm operates by comput-
ing the eigenvalues of the covariance matrix, aiming to minimize computation costs while
preserving accuracy. By extracting independent information from the acoustic data, PCA
successfully reduces dimensionality while maintaining the essential aspects of the dataset.
In addition to PCA, another algorithm commonly used for assessing feature distinctiveness
Sensors 2024, 24, 2297 7 of 23

and visualizing data separability is t-SNE, which is capable of preserving the relationships
among data points in a lower-dimensional space, making it particularly useful for visual-
izing complex high-dimensional data. One key distinction between the PCA and t-SNE
techniques is that while PCA focuses on preserving large pairwise distances to maximize
variance, t-SNE primarily preserves local similarities among data points. This difference
in approach makes t-SNE well-suited for visualizing intricate patterns in the data [29,30].
For enhanced comparison and visualization, LDA, a supervised technique, is utilized.
LDA seeks to generate new components that differentiate categories effectively, focusing
on a class feature with the goal of identifying the projection that optimally distinguishes
between faults in the ultrasonic data [31]. By reducing the dimensionality of the data,
LDA strives to improve class separability while retaining as much information as possible.
Although it offers a solution to the challenge of small sample sizes, LDA faces issues such
as singularity, dependence on the assumption of Gaussian distributions, and the difficulty
in determining the best dimensionality reduction.

2.4. Recursive Feature Elimination (RFE)


RFE selects features recursively, considering increasingly smaller sets of features at
each step. By comparing different feature sets, the algorithm aims to identify the set
that achieves the highest accuracy [32,33]. In every iteration of the RFE algorithm, the
importance of each feature is measured, and the least relevant feature is subsequently
removed. In other words, RFE aims to evaluate the significance of features within a
dataset, given an external estimator or classifier that assigns importance weights to these
features. This method systematically selects ultrasonic features from the feature set (FS) by
progressively focusing on increasingly smaller subsets of ultrasonic characteristics through
iterative processes. Features that are deemed less important and thus eliminated in each
iteration are then placed into a ranking set (RS). With each subsequent iteration, the FS
is methodically reduced to further assess the importance of the remaining features. This
procedure is repeated multiple times to exhaustively analyze all possible combinations of
features within the FS. Ultimately, the features are systematically organized and ranked
within the RS based on their importance, as determined through the RFE algorithm. The
RFE procedure is depicted in Figure 3.

2.5. Classification Using k-Fold Cross Validation


In this study, for an industrial fault detection system, we propose an effective moni-
toring method using ultrasonic signals based on well-known classifiers. The classification
algorithm first obtains raw signals, and it then preprocesses and scales the time series ultra-
sonic signals. Consequently, a variety of time- and frequency-domain features are extracted
from the data, which serve as input to the classifier. Then, five classifiers including KNN,
LR, DT, GNB, and SVM are employed to distinguish the fault classes. The selection of these
classifiers is based on their reduced memory requirements with lower output variations
in comparison to deep learning classifiers [34], making them well-suited for ensemble
implementation. Afterwards, a meta classifier based on k-fold cross validation (CV) is
designed to detect and classify the irregularities in the ultrasonic signals. The mentioned
classifiers are described as follows:
SVM is a clustering algorithm specifically designed for small sample sets, demonstrat-
ing exceptional learning capabilities even when the data is limited. It also exhibits good
generalization properties. The SVM algorithm, utilizing a kernel function, aims to identify
the optimal hyperplane or decision boundary (such as a point, line or plane) that effectively
separates different classes of data [35].
GNB algorithm is an extension of the Naive Bayes algorithm. Naive Bayes is a
generative model that assumes the independence of features. In the case of Gaussian
Naive Bayes, the algorithm assumes that the covariance matrices are diagonal, implying
independence between the features. Different functions are employed to estimate the data
distribution and calculate the mean and standard deviation for the training [36].
Sensors2024,
Sensors 2024,24,
24,2297
x FOR PEER REVIEW 88 of 24
of 23

Figure 3. Recursive
Figure3. Recursive feature
feature elimination
eliminationflow
flowdiagram.
diagram.

LR is a machine learning classification algorithm widely used in statistical analysis.


2.5. Classification Using k-Fold Cross Validation
It is employed to describe data and model the relationship between a single dependent
In this
variable andstudy,
two orfor an independent
more industrial fault detection
variables, system,
which can beweofpropose
ordinal,an effective
nominal, moni-
interval,
toring method using ultrasonic signals based on well-known classifiers.
or ratio level. The purpose is to determine the outcome or predict the probability of a The classification
algorithm
specific firstorobtains
event class [37].raw signals, and it then preprocesses and scales the time series
ultrasonic
DT is asignals.
supervisedConsequently, a variety
machine learning of time-used
algorithm and forfrequency-domain
both regression and features are
classifi-
extracted from the data, which serve as input to the classifier. Then,
cation tasks. It is constructed using a directed graph technique. The decision tree follows five classifiers includ-
aing KNN, structure,
tree-like LR, DT, GNB, where and SVMnodes
inner are employed
representtothe distinguish
variables the fault classes.
or features The se-
of a dataset,
lection of these classifiers is based on their reduced memory requirements
branches represent the decision rules, and each leaf node represents the output result [38]. with lower out-
put variations in comparison to deep learning classifiers [34],
KNN is a supervised classification approach that avoids making any assumptions making them well-suited
for ensemble
about the primaryimplementation. Afterwards, a algorithm
data. It is a non-parametric meta classifier based on
that utilizes thek-fold crossbetween
similarity valida-
tion (CV)
new is designed
data and existingtodata detect and classify
to assign the new thedata
irregularities
to the most in the ultrasonic
similar groupsignals.
among The the
mentioned
available classifiers
groups [39]. are described as follows:
SVM isCV
k-fold a clustering
is employed algorithm specifically
to evaluate designed forofsmall
the generalization sample
industrial sets,
fault demon-
detection
strating results
analysis exceptionalon anlearning
independent capabilities
dataset.evenBy when the data
employing k-foldis limited.
CV, we It also
can exhibits
assess the
good generalization
predictive performance properties. The SVM
of our machine algorithm,
learning modelutilizing
on unseena kernel function,
ultrasonic aims
data andto
identify
ensure thethe optimal
model’s hyperplane
robustness or Using
[40]. decisionthisboundary
approach,(such as a point, dataset
our underlying line or plane) that
is divided
effectively
into separates different
k non-overlapping folds. Asclasses
shownof data [35]. 4, in the test phase, the first ultrasonic
in Figure
signalGNBfoldalgorithm
is used toisevaluate
an extension the model
of the performance, while the remaining
Naive Bayes algorithm. Naive Bayes k−is1afolds
gen-
are used for training. Next, the second input signal fold is used
erative model that assumes the independence of features. In the case of Gaussian Naive for the test step and the
rest of our
Bayes, the dataset
algorithm is used
assumes for the
thattraining. The procedure
the covariance matrices is arerepeated
diagonal, k times
implyinguntilinde-
the
last ultra-sonic
pendence fold,the
between and at last, Different
features. the average value of
functions arethese k experiments
employed to estimate is used for dis-
the data the
evaluation
tribution and [39].calculate
In this work,
the meanstratified k-fold CVdeviation
and standard is employed, which
for the makes[36].
training the results less
susceptible
LR is atomachine
data biaslearning
and imbalance. Stepsalgorithm
classification for k-fold widely
CV are usedsummarized as follows:
in statistical analysis.
It is employed to describe data and model the relationship between a single dependent
variable and two or more independent variables, which can be of ordinal, nominal,
signal fold is used to evaluate the model performance, while the remaining k−1 folds are
used for training. Next, the second input signal fold is used for the test step and the rest
of our dataset is used for the training. The procedure is repeated k times until the last
ultra-sonic fold, and at last, the average value of these k experiments is used for the eval-
Sensors 2024, 24, 2297
uation [39]. In this work, stratified k-fold CV is employed, which makes the results less9 of 23
susceptible to data bias and imbalance. Steps for k-fold CV are summarized as follows:

Figure 4. Pseudo-code
Figure for the
4. Pseudo-code fork-fold CV algorithm.
the k-fold CV algorithm.

2.6. Meta
2.6. Meta Classifier
Classifier
In theInrealm
the realm of ensemble
of ensemble learning
learning methodologies,
methodologies, stacking
stacking involves
involves thethe integration of
integration
of multiple learning models through a meta-classifier. As illustrated in Figure 5, the5,initial
multiple learning models through a meta-classifier. As illustrated in Figure the initial
step entails training individual classification models (C
step entails training individual classification models (Ci) using i ) using the entire ultrasonic training
the entire ultrasonic train-
set. Subsequently, the meta-classifier is trained using the labels from these individual
ing set. Subsequently, the meta-classifier is trained using the labels from these individual
machine learning models within the ensemble [41]. The forecasted outputs of the classifiers
machine learning models within the ensemble [41]. The forecasted outputs of the classifi-
(P ) serve as meta-features, which are then combined with weights (wi ). These weighted
ers (Pi)i serve as meta-features, which are then combined with weights (wi). These
meta-features are utilized as inputs for the meta-classifier, ultimately determining the final
weighted meta-features are utilized as inputs for the meta-classifier, ultimately determin-
output (ŷ), which is chosen based on the predicted class output from other classifiers. This
ing the final output ( ŷ ), which is chosen based on the predicted class output from other
process is defined as follows:
classifiers. This process is defined as follows:
Sensors 2024, 24, x FOR PEER REVIEW m 10 of 24
ŷ = arg max ∑ w j Pij (7)
i j =1
𝑦 = 𝑎𝑟𝑔 𝑚𝑎𝑥 𝑤𝑃 (7)

Figure 5. Utilization
Figure 5. Utilization of
of aa fault
fault detection
detection framework employing the
framework employing the k-fold
k-fold CV
CV and
and meta-classifier.
meta-classifier.

Ensemble methods,
methods,whichwhichaggregate
aggregatepredictions from
predictions various
from base
various models,
base often
models, en-
often
hance
en-hance generalization and robustness, outperforming single-model approaches duethe
generalization and robustness, outperforming single-model approaches due to to
diversity in the
the diversity inbase models.
the base In contrast,
models. deep learning
In contrast, modelsmodels
deep learning are susceptible to overfit-
are susceptible to
overfit-ting,
ting, particularly
particularly with limited
with limited datasetsdatasets or excessive
or excessive model model capacity.
capacity. StackingStacking meta-
meta-classi-
classifiers
fiers addresses
addresses this by
this issue issue by integrating
integrating outputs
outputs from several
from several models,models, each trained
each trained on dif-
on different
ferent data subsets
data subsets or utilizing
or utilizing distinctdistinct algorithms,
algorithms, therebythereby
reducingreducing
the risk the risk of
of overfit-
overfit-ting.
ting.
Stacking meta-classifiers is especially advantageous for small datasets, where they
tend to outperform deep learning models that typically require vast amounts of labeled
data for effective training. Moreover, deep learning models demand substantial compu-
tational power and specialized hardware like GPUs or TPUs, making them less accessible.
In contrast, stacking meta-classifiers are less demanding on resources and can be trained
more efficiently on conventional hardware.
Sensors 2024, 24, 2297 10 of 23

Stacking meta-classifiers is especially advantageous for small datasets, where they


tend to outperform deep learning models that typically require vast amounts of labeled data
for effective training. Moreover, deep learning models demand substantial computational
power and specialized hardware like GPUs or TPUs, making them less accessible. In
contrast, stacking meta-classifiers are less demanding on resources and can be trained more
efficiently on conventional hardware.
Therefore, stacking meta-classifiers is often a superior choice in situations demanding
high interpretability, limited data availability, constrained computational resources, sim-
plicity in model complexity, robustness against overfitting, reliability of results, and when
domain expertise is a key factor.

2.7. Implementation
The “micromlgen” library, as depicted in this study, represents a critical tool for the
deployment of machine learning models on microcontrollers, facilitating the conversion
of such models into C code suitable for hardware environments constrained by memory,
processing power, and energy consumption. Despite its utility, the practical implementation
of “micromlgen” for classification tasks reveals several limitations, particularly due to the
restricted computational resources of microcontrollers. The “micromlgen” library exhibits
optimal performance with simpler or specifically tailored lightweight models, given its
support for a finite range of ML algorithms. This includes well-known algorithms such as
decision trees, support vector machines, and certain neural network types, albeit excluding
support for more complex or recent models like extensive deep learning architectures.
In the realm of TinyML, which optimizes AI for low-power, compact microcontrollers,
the nRF52840 microcontroller stands out for its capability to deploy AI models efficiently,
leveraging TinyML libraries. This synergy is particularly conducive for embedding on-
device decision-making capabilities, thus diminishing the reliance on perpetual cloud
connectivity. Utilizing “micromlgen” enables the conversion of ML models into C11code,
Sensors 2024, 24, x FOR PEER REVIEW of 24
rendering it compatible with the development environment of the nRF52840. This process
allows for the direct integration of AI into devices, facilitating real-time data processing and
decision-making.
execution efficiency Theonapplication of TinyML
the nRF52840, libraries
capitalizing on thefurther enhances model
microcontroller’s execution
functionalities
efficiency on the nRF52840,
while managing its limitations. capitalizing on the microcontroller’s functionalities while
managing its limitations.
The integration of “micromlgen” for model conversion and TinyML for deployment
The integration of “micromlgen” for model conversion and TinyML for deployment
on the nRF52840 MCU embodies a strategic methodology for embedding ML into hard-
on the nRF52840 MCU embodies a strategic methodology for embedding ML into hard-
ware systems. This combination exploits the advantages of the aforementioned tools and
ware systems. This combination exploits the advantages of the aforementioned tools and
the MCU’s capabilities, paving the way for the development of intelligent embedded sys-
the MCU’s capabilities, paving the way for the development of intelligent embedded
tems proficient in edge data processing. The implementation flowchart, illustrated in Fig-
systems proficient in edge data processing. The implementation flowchart, illustrated in
ure 6, delineates this comprehensive process, underscoring the potential for developing
Figure 6, delineates this comprehensive process, underscoring the potential for developing
autonomous, cloud-independent devices that prioritize privacy, speed, and reliability.
autonomous, cloud-independent devices that prioritize privacy, speed, and reliability.

Figure 6. Work flow of the model implementation.


Figure 6. Work flow of the model implementation.
2.8. Evaluation Metrics
2.8. Evaluation Metrics
In assessing the efficacy of the fault detection framework put forth, conventional
quality
In metrics including
assessing precision,
the efficacy of therecall,
fault F-measure, and accuracy
detection framework putare employed.
forth, These
conventional
metrics
quality furnish
metrics quantitative evaluations
including precision, of the
recall, framework’s
F-measure, ability to are
and accuracy identify faults, These
employed. draw-
ing from furnish
metrics parameters such as true
quantitative positive (TP),
evaluations falseframework’s
of the negative (FN), true negative
ability (TN),
to identify and
faults,
false positive
drawing from(FP) [42]. Theysuch
parameters are defined as follows:
as true positive (TP), false negative (FN), true negative
(TN), and false positive (FP) [42]. They are defined as follows:
TP
Precision = (8)
TP +TP FP
Pr ecision = (8)
TP + FP

TP
Recall = (9)
TP + FN

2 × Pr ecision × Recall
F1 = (10)
Sensors 2024, 24, 2297 11 of 23

TP
Recall = (9)
TP + FN
2 × Precision × Recall
F1 = (10)
Precision + Recall
TP + TN
Accuracy = (11)
TP + FN + TN + FP
TP signifies the count of accurately classified samples deemed relevant, FN indicates
the count of relevant samples inaccurately classified, TN denotes the count of accurately
classified samples considered irrelevant, and FP represents the count of samples inaccu-
rately classified as relevant. Precision quantifies the ratio of accurately classified relevant
samples to all samples classified as relevant. Recall gauges the ratio of relevant samples
accurately classified to all relevant samples within the dataset. The F1 score calculates
the harmonic mean of precision and recall, furnishing a balanced assessment of classi-
fication performance that considers both precision and recall. Accuracy determines the
proportion of correctly classified samples across all classes, presenting an overall gauge
of the accuracy of classification [43]. In addition, receiver operating characteristic (ROC)
plots offer a comprehensive view of a classifier’s performance across various levels of
specificity. In this research, ROC plots are utilized to further investigate and evaluate the
performance of the classifiers. The ROC plots provide insights into the trade-off between
the true positive rate and the false positive rate, allowing for a more detailed analysis of
the classifiers’ performance.

3. Results and Discussion


Ultrasonic fault detection for pipes and motors is a technique to identify defects or
irregularities. The process involves using a contact ultrasonic microphone sensor, also
known as an ultrasonic transducer. The contact ultrasonic microphone sensor is placed in
direct contact with the surface of the pipe or motor to receive the data. The profile can be
due to friction, impacts, leaks, or cracks that occur naturally during operation or as a result
of damage.
Healthy and uniform material emits the wave in minimal disruption. Some advantages
of contact measuring are that it requires no interruption of normal equipment operation,
allowing for continuous monitoring. It can identify very small or early-stage faults that
might not impact visible performance or be detectable through other non-destructive
testing methods.
Also, it eliminates the cost of an external ultrasonic transmitter and the associated
operational complexities. This contact sensing is particularly suited to environments where
continuous monitoring is desired without the need for frequent manual inspections or where
the operational noise of the equipment itself can be leveraged to identify potential issues.
The dataset extracted from UE Systems Co. by a piezoelectric array called an “Ultra-
probe” consist of 10 classes, including 4 for bearings (i.e., under-lubricated, over-lubricated,
slow speed, and healthy bearing) and 6 for pipelines (i.e., steam, cavitation, motor-boating,
reciprocating, thermostatic, and healthy pipe), with 20,000 samples [44], 70% for training
and 30% for testing, along with a sliding window with length of 1000 to select and augment
the data.
The motor rolling bearing and pipeline faults mainly consist of the main body which
contains the ultrasonic signals of ten patterns: under-lubricated, over-lubricated, slow
speed, steam, cavitation, motor boating, reciprocating, thermostatic, healthy pipe, and
healthy motor signals. The data samples are shown in Figure 7.
Sensors 2024, 24, x FOR PEER REVIEW 13 of 24

Sensors 2024, 24, 2297 12 of 23

Figure 7. Ultrasonic
Figure 7. Ultrasonicsignals
signals of the
of the ten patterns:
ten patterns: under-lubricated,
under-lubricated, over-lubricated,
over-lubricated, slow
slow speed, speed,
steam,
steam, cavitation, motor boating, reciprocating, thermostatic, healthy pipe, and healthy motor sig-
cavitation, motor boating, reciprocating, thermostatic, healthy pipe, and healthy motor signals.
nals.
As mentioned in the introduction, regarding the IPF curve, the rapid identification of
bearing and pipe faults is a primary objective in the operation of industrial mechanisms.
This study focuses on investigating the impact of features on the statistical parameters of
the acquired ultrasonic datasets. The aim is to understand how different features affect
the statistical properties of the collected ultrasonic data, which are crucial for the quick
identification of faults in bearings and pipes. For visualization the features, other methods
like PCA, LDA, and t-SNE can be considered. PCA preserves the variance in the data
and is applied to the data to reduce its dimensionality and readily visualize the faults.
In addition to PCA for dimensionality reduction, t-SNE also is calculated. Figures 8–10
respectively, present the PCA, t-SNE, and LDA visualization of ultrasonic data based on
two components. As mentioned in the previous section, five classifiers for our model
are employed. For GNB (smoothing= 0.008), DT (orientation: ”Gini”, min split sample:
3), LR (sklearn = 0.98), kNN (k = 7), and SVM (cost = 0.520 and Gaussian kernel), m = 5
are considered.

Figure 8. PCA representation of the ultrasonic dataset.


Figure 7. Ultrasonic signals of the ten patterns: under-lubricated, over-lubricated, slow speed,
Sensors 2024, 24, 2297 steam, cavitation, motor boating, reciprocating, thermostatic, healthy pipe, and healthy motor
13 sig-
of 23
nals.

Sensors 2024, 24, x FOR PEER REVIEW 14 of 24

PCArepresentation
Figure8.8.PCA
Figure representationofofthe
theultrasonic
ultrasonicdataset.
dataset.

t-SNErepresentation
Figure9.9.t-SNE
Figure representationofofthe
theultrasonic
ultrasonicdataset.
dataset.

Feature selection is a critical step in the classification process, aimed at identifying


and eliminating irrelevant, less useful, and redundant features, while identifying the most
relevant and informative inputs for a classification model. By reducing the number of
in-put features, feature selection assists in the development of a predictive model. It plays
a crucial role in enhancing the efficiency and accuracy of the classification task by focusing
on the most significant features that contribute to the classification performance while
eliminating noise and irrelevant information. Furthermore, feature selection is desirable to
both decrease the computational cost of classification models and, in some cases, to improve
the performance of the fault detection model. It enables onboard ML for MCU devices by
decreasing the required memory size. However, deploying ML models on MCUs presents
challenges, notably in terms of processing power consumption and limitations in flash
memory, which can restrict the implementation of all available features.
Sensors 2024, 24, 2297 14 of 23

Figure 9. t-SNE representation of the ultrasonic dataset.

Sensors 2024, 24, x FOR PEER REVIEW 15 of 24

MCUs presents challenges, notably in terms of processing power consumption and limi-
Figure 10. LDA representation of
of the ultrasonic
Figurein
tations LDAmemory,
10.flash representation
which the
can restrictdataset.
ultrasonic dataset.
the implementation of all available features.
Figure
Figure1111 shows
showsthetheRFE method performance for
forour ultrasound dataset. Mostly
Mostlyfor
Feature selection is a RFE method
critical step inperformance
the classificationour ultrasound
process, aimeddataset.
at identifying
implementation,
for implementation, the RFE results
the RFE can
results be
can used to
beredundant deploy
used to deploy our desired
our while
desired ML models
ML models onthethe
and eliminating irrelevant, less useful, and features, identifying theon
most
MCU,
MCU,and
relevantand
and
the
thefinal
finalmodel is
is used
modelinputs
informative used for
forfor
real-timeindustrial
real-time industrial
a classification
monitoring.
model.monitoring.
By reducing the number of in-
put features, feature selection assists in the development of a predictive model. It plays a
crucial role in enhancing the efficiency and accuracy of the classification task by focusing
on the most significant features that contribute to the classification performance while
eliminating noise and irrelevant information. Furthermore, feature selection is desirable
to both decrease the computational cost of classification models and, in some cases, to
improve the performance of the fault detection model. It enables onboard ML for MCU
devices by decreasing the required memory size. However, deploying ML models on

Figure 11.11.Feature
Figure Featureselection
selectionusing
using the
the RFE
RFE algorithm.
algorithm.

InInthis
thisgraph,
graph,ititisisclear
clear that
that when
when five
five features,
features, bracketed
bracketedtotothe theright
rightofofFigure
Figure11,11,
are
are eliminated(i.e.,
eliminated (i.e.,kurtosis,
kurtosis,zero
zero crossing,
crossing, skewness,
skewness,shape
shapefactor,
factor,and
andpeak
peak index), thethe
index),
accuracy drops dramatically, indicating that these features play the more important to
accuracy drops dramatically, indicating that these features play the more important role role
todistinguish
distinguishthetheclasses.
classes.
To validate the efficacy of the methods employed, accuracy analyses were conducted
To validate the efficacy of the methods employed, accuracy analyses were conducted
across four techniques for six classifiers. RFE distinguished itself by delivering superior
across four techniques for six classifiers. RFE distinguished itself by delivering superior
accuracy in pattern classification, a claim substantiated by the results depicted in Figure 12.
accuracy in patternanalysis
The comparative classification,
reveals athat
claim
RFEsubstantiated by the results
achieves commendable depicted
diagnostic in Figure
outcomes
12.relative
The comparative
to alternative analysis reveals
data-reduction that RFE
strategies. Theachieves commendable
metric result diagnostic
of the six classifiers out-
based
comes relative to alternative
on RFE is shown in Table 1. data-reduction strategies. The metric result of the six classi-
fiers based on RFE is shown in Table 1.
The results in Table 1 show the superiority of the meta model for all the scores across
the ten classes. The classification results of the selected methods, along with the utilized
meta-classifier technique, are summarized using a confusion matrix, as shown in Figure
13. The confusion matrix provides insights into the ability of the methods to accurately
Sensors 2024, 24, x FOR PEER REVIEW 16 of 24
Sensors 2024, 24, 2297 15 of 23

Figure 12. Applying


Figure 12. Applyingmulti-class
multi-class classification
classification using
using dimensionality
dimensionality reduction
reduction t-SNE, t-SNE, PCA,
PCA, LDA, LDA,
and
and RFE feature selection.
RFE feature selection.

Table 1. Precision, recall, F1 measure, and accuracy of fault detection that are achieved by the ML
Table 1. Precision, recall, F1 measure, and accuracy of fault detection that are achieved by the ML
models.
models.
F1 F1
Classifier Class Precision
Classifier Class
Recall
Precision Recall
F1 Accuracy Classifier Class Precision Recall F1
Accuracy Classifier Class Precision Recall
Accuracy
Accuracy
Score
Score Score
Score
00 0.48
0.48 1.00
1.00 0.65
0.65 0 0 0.560.56 0.80 0.80 0.660.66
11 0.57
0.57 0.45
0.45 0.50
0.50 1 1 0.620.62 0.58 0.58 0.600.60
2 0.89 0.83 0.86 2 0.86 0.84 0.85
23 0.89
0.97 0.83
0.93 0.86
0.95 3 2 0.920.86 0.88 0.84 0.900.85
34 0.97
0.50 0.93
0.46 0.95
0.48 4 3 0.680.92 0.71 0.88 0.690.90
GNB 0.731 KNN 0.798
45 0.82
0.50 0.92
0.46 0.87
0.48 5 4 0.870.68 0.87 0.71 0.870.69
GNB 6 1.00 0.99 1.00 0.731 KNN 6 0.98 1.00 0.99 0.798
57 0.82
0.88 0.92
0.95 0.87
0.91 7 5 0.840.87 0.91 0.87 0.870.87
68 1.00
0.66 0.99
0.58 1.00
0.62 8 6 0.820.98 0.70 1.00 0.750.99
79 0.91
0.88 0.22
0.95 0.35
0.91 9
7 0.80
0.84 0.57 0.91 0.660.87
0 0.58 0.95 0.72 0 0.82 0.78 0.80
81 0.66
0.76 0.58
0.63 0.62
0.69 1 8 0.720.82 0.74 0.70 0.730.75
92 0.91
0.91 0.22
0.82 0.35
0.86 2 9 0.930.80 0.91 0.57 0.920.66
03 0.92
0.58 0.90
0.95 0.91
0.72 3 0 0.960.82 0.99 0.78 0.970.80
4 0.80 0.57 0.66 4 0.79 0.85 0.82
LR 15 0.76 0.63 0.69 0.817 DT 0.886
0.85 0.93 0.89 5 1 0.880.72 0.92 0.74 0.900.73
26 0.91
0.99 0.82
1.00 0.86
0.99 6 2 1.000.93 0.99 0.91 0.990.92
37 0.76
0.92 0.95
0.90 0.84
0.91 7
3 0.99
0.96 0.96 0.99 0.970.97
8 0.66 0.53 0.59 8 0.89 0.85 0.87
49 0.80
0.84 0.57
0.66 0.66
0.74 9 4 0.840.79 0.82 0.85 0.830.82
LR 0.817 DT 0.886
5 0.85 0.93 0.89 5 0.88 0.92 0.90
6 0.99 1.00 0.99 6 1.00 0.99 0.99
7 0.76 0.95 0.84 7 0.99 0.96 0.97
8 0.66 0.53 0.59 8 0.89 0.85 0.87
9 0.84 0.66 0.74 9 0.84 0.82 0.83
Sensors 2024, 24, 2297 16 of 23

Table 1. Cont.

F1 F1
Classifier Class Precision Recall Accuracy Classifier Class Precision Recall Accuracy
Score Score
0 0.71 0.92 0.80 0 0.86 0.93 0.90
1 0.84 0.67 0.75 1 0.95 0.83 0.88
2 0.96 0.84 0.90 2 0.96 0.88 0.92
3 0.98 0.97 0.97 3 0.98 1.00 0.99
4 0.83 0.84 0.83 4 0.89 0.96 0.92
SVM 0.884 Meta 0.932
5 0.88 0.95 0.91 5 0.90 0.96 0.93
6 0.99 1.00 0.99 6 1.00 1.00 1.00
7 0.97 0.97 0.97 7 1.00 0.97 0.98
8 0.84 0.88 0.86 8 0.88 0.91 0.90
9 0.89 0.79 0.84 9 0.92 0.89 0.91

The results in Table 1 show the superiority of the meta model for all the scores across
the ten classes. The classification results of the selected methods, along with the utilized
meta-classifier technique, are summarized using a confusion matrix, as shown in Figure 13.
The confusion matrix provides insights into the ability of the methods to accurately classify
different ultrasonic fault states. Higher values along the diagonal indicate a better model,
as it correctly classifies the majority of the samples. Since the dataset used in this research
includes both pipe and bearing faults, it allows us to evaluate the performance of the
proposed method in handling mixed faults involving both pipes and bearings. Figure 14
illustrates the ROC curves of the selected models, utilizing k-fold CV.
The ROC curve serves as a measure of separability, indicating the capability of the
selected models to distinguish between different fault classes. It provides a visual rep-
resentation of the trade-off between the true positive rate (TPR) and false positive rate
(FPR) for a predictive model across various probability thresholds. Basically, for every
threshold, TPR and FPR are calculated and plotted on one chart. As can be seen, the
lower FPR and the higher TPR for the given RFE thresholds demonstrate the better results.
A steeper and more vertical ROC curve is desirable, as it indicates a model with better
discriminative ability. The ROC curve analysis is instrumental in evaluating the perfor-
mance and effectiveness of the selected models for fault classification tasks. In Figure 14,
it can be observed that the meta model, which refers to a composite model combining
predictions from multiple models, exhibits an area under the ROC curve (AUC) of 0.997
for the ultrasonic test dataset. This AUC is significantly greater than those of the other
classifiers, underscoring the meta model’s superior discriminative capability and predictive
performance for fault classification.
Therefore, it can be concluded that the selected meta model is reliable and demon-
strates strong usability in the context of ultrasonic fault detection. Its high area under the
ROC curve suggests that it can effectively distinguish between different fault classes and
make accurate predictions.
Furthermore, in this research on ultrasonic fault evaluation, stratified k-fold CV is
utilized as a method for rotation estimation to assess the generalizability of the statistical
analysis to an independent ultrasonic dataset. Our simulations provide a quantitative
analysis of the classification performance across five folds and the selected ML models,
as presented in Table 2. The average test accuracy obtained through rotation estimation
for the meta-classifier is 93%. This average accuracy of the meta model surpasses that of
the other models under consideration. To better provide analysis of our simulation results
over the selected models through the k-fold CV method, the accuracy of each model is
visualized in Figure 15.
Sensors 2024, 24, x FOR PEER REVIEW 18 of 24
Sensors 2024, 24, 2297 17 of 23

Figure 13.
Figure Theconfusion
13. The confusionmatrix
matrixbased
basedon
onthe
theachieved
achievedresults
results by
bythe
theSVM,
SVM,GNB,
GNB,LR,
LR,KNN,
KNN,DT,
DT, and
and
meta-classifier models.
meta-classifier models.
Sensors
Sensors 2024,
2024, 24,
24, x2297
FOR PEER REVIEW 19 of 24
18 of 23

Figure
Figure14.
14.ROC
ROCcurves
curvesfor
forsix
sixclassifiers.
classifiers.

Table 2. Performance of SVM, GNB, LR, KNN, and meta-classifier models through 5-fold CV.
Table 2. Performance of SVM, GNB, LR, KNN, and meta-classifier models through 5-fold CV.
k-folds SVM GNB LR KNN DT Meta
k-folds SVM GNB LR KNN DT Meta
Fold_1 0.877174 0.713043 0.811957 0.806522 0.878261 0.925140
Fold_1 0.877174 0.713043 0.811957 0.806522 0.878261 0.925140
Fold_2 0.893478 0.741304 0.856522 0.790217 0.890217 0.933696
Fold_2 0.893478 0.741304 0.856522 0.790217 0.890217 0.933696
Fold_3
Fold_3 0.902174
0.902174 0.736957
0.736957 0.858696
0.858696 0.817391
0.817391 0.871739
0.871739 0.939135
0.939135
Fold_4
Fold_4 0.879348
0.879348 0.731522
0.731522 0.835870
0.835870 0.784783
0.784783 0.876087
0.876087 0.925581
0.925581
Fold_5
Fold_5 0.885870
0.885870 0.747826
0.747826 0.833696
0.833696 0.814130
0.814130 0.916304
0.916304 0.935870
0.935870
Average 0.887609 0.734130 0.839348 0.802609 0.886522 0.931739
Average 0.887609 0.734130 0.839348 0.802609 0.886522 0.931739
Sensors 2024, 24, x FOR PEER REVIEW 20 of 24
Sensors 2024, 24, x FOR PEER REVIEW 20 of 24
Sensors 2024, 24, 2297 19 of 23

Figure 15. Accuracy of the six classifier models using the k-fold CV strategy.
Figure15.
Figure 15.Accuracy
Accuracyofofthe
the
sixsix classifier
classifier models
models using
using the the k-fold
k-fold CV strategy.
CV strategy.
For more comparison, the variations in the results during the CV step for different
folds For
Formore
are morecomparison,
presented throughthe
comparison, variations
athe
boxplot in in
thethe
in Figure
variations results during
16.results the the
during CV CV
stepstep
for different
for different
folds are presented through a boxplot in Figure 16.
folds are presented through a boxplot in Figure 16.

Figure 16. Boxplot for the six classifier models.


Figure 16. Boxplot for the six classifier models.
Figure
The16.distribution
Boxplot for the six data
of the classifier models.
allows for effective comparisons of accuracy across dif-
ferentThe distribution
folds within theof the data dataset.
ultrasonic allows forTheeffective comparisons
boxplot offers of accuracy
a graphical depictionacross
of the dif-
ferent folds
five-number within
The distributionthe
summary of ultrasonic
ofthe
thedataset,dataset.
data allows The boxplot
for effective
encompassing offers
comparisons
the minimum a graphical
value,offirst depiction
accuracy
quartileacross of the
(25th dif-
five-number
percentile), summary
ferent foldsmedian
within the of
(50th the dataset,
ultrasonic
percentile), encompassing
dataset.
third The boxplot
quartile the minimum
offers
(75th percentile), and value,
a graphical
maximum firstvalue.
depiction quartile
of the
(25th
The percentile),
five-number
‘box’ median
summary
visualizes (50th
of the percentile),
dataset,
the interquartile range third
(IQR),quartile
encompassingthe span (75th
the percentile),
minimum
between thevalue,and
first maximum
first
and quartile
third
value. Thewhile
quartiles, ‘box’the
(25th percentile), visualizes
‘whiskers’
median theextend
(50thinterquartile rangequartile
to the furthest
percentile), third (IQR), the span
data points
(75th not betweenand
considered
percentile), the first and
outliers,
maximum
which
third
value. are
Theplotted
quartiles,
‘box’ individually.
while
visualizes theThis
the ‘whiskers’ boxplot
extend to
interquartile succinctly captures
the furthest
range (IQR), the the
data distribution,
points
span central
not considered
between the firstout-
and
tendency,
liers,
thirdwhich variability,
are while
quartiles, plottedand outliers of the
theindividually.
‘whiskers’ ultrasonic
This
extend to the data.
boxplot It isdata
succinctly
furthest evident from
captures
points not the
the analysis out-
distribution,
considered
that
liers,the
central meta-classifier
tendency,
which plotted demonstrates
are variability, theThis
and outliers
individually. highest
of the consistency
ultrasonic
boxplot data.
succinctly in It
accuracy
is evident
captures and the
from
the least
the anal-
distribution,
variability
ysis
central compared
thattendency, to other
the meta-classifier
variability, models.
demonstrates
and outliersthe highest
of the consistency
ultrasonic data. Itinisaccuracy
evident fromand the theleast
anal-
variability compared to other models.
ysis that the meta-classifier demonstrates the highest consistency in accuracy and the least
variability compared to other models.
Sensors 2024, 24, x FOR PEER REVIEW 21 of 24

Sensors 2024, 24, 2297 20 of 23

Accurately labeled data is crucial for the effectiveness of supervised learning algo-
rithms in industrial
Accurately fault
labeled diagnosis.
data is crucialInsufficient labeled data
for the effectiveness complicates
of supervised modelalgo-
learning training
and can lead to decreased accuracy. Additionally, the complexity of ultrasonic
rithms in industrial fault diagnosis. Insufficient labeled data complicates model training data, signal
acquisition
and can leaddelays, and sample
to decreased imbalances
accuracy. Additionally,among faults present
the complexity significant
of ultrasonic challenges
data, signal in
superviseddelays,
acquisition diagnosis. Certain imbalances
and sample classifiers, among
like GNB,
faultspresuppose feature challenges
present significant independence,
which
in complicates
supervised the management
diagnosis. of highly
Certain classifiers, correlated
like GNB, data and
presuppose hinders
feature the attainment
independence,
which complicates the management of highly correlated data and hinders
of optimal performance. Moreover, a limited dataset and uneven distribution of ultrasonicthe attainment
of optimalmay
features performance. Moreover, apredictive
result in substantial limited dataset and
errors. uneven
Figure 17 distribution
highlights the of ultrasonic
issues arising
features may result
from limited samplein substantial predictive
sizes, illustrating howerrors. Figure 17
insufficient highlights
data the issues
can diminish thearising
predictive
from limited sample sizes, illustrating how insufficient data can diminish
accuracy of classification models. This underscores the importance of larger and more bal-the predictive
accuracy of classification
anced datasets models.
for enhancing the This underscores
efficacy the importance
of fault diagnosis models. of larger and more
balanced datasets for enhancing the efficacy of fault diagnosis models.

Figure17.
Figure 17.Accuracy
Accuracycurve
curvevs.vs.the
the number
number of of samples
samples (learning
(learning curve).
curve).

Figure
Figure17 17clearly
clearlydemonstrates
demonstrates that thethe
that metametamodel
modelcancan
predict withwith
predict highhigh
accuracy
accuracy
even
even with a limited number of samples, approximately 300. With an increase number
with a limited number of samples, approximately 300. With an increase in the in the num-
of samples,
ber the meta
of samples, modelmodel
the meta consistently outperforms
consistently the other
outperforms themodels in accuracy.
other models in accuracy.
In the final step, the micromlgen library in Python is used to extract the model and
In the final step, the micromlgen library in Python is used to extract the model and
deploy it onto an nRF52840 64 MHz ARM Cortex-M4F MCU. The five optimal features,
deploy it onto an nRF52840 64 MHz ARM Cortex-M4F MCU. The five optimal features,
specifically kurtosis, zero crossing, skewness, shape factor, and peak index (as identified
specifically kurtosis, zero crossing, skewness, shape factor, and peak index (as identified
in Figure 11), are computed on the MCU to serve as inputs for five classifiers: GNB,
in Figure
LR, KNN, 11),
SVM, areand
computed
DT. The on the MCU
outputs fromtothese
serve as inputsare
classifiers forthen
fivefed
classifiers:
into the GNB,
meta LR,
KNN, SVM,
model, whichand DT. The
executes in aoutputs
total offrom
11 ms these classifiers
on this MCU—thisare then fed into
includes 6.5 the
ms meta model,
for data
which executes in a total of 11 ms on this MCU—this includes 6.5 ms for
sampling, 183 µs for feature calculation, and 4.3 ms for the meta model prediction. These data sampling,
183 μs for
timings werefeature calculation,
verified using a logicand 4.3 ms as
analyzer, fordepicted
the meta model 18,
in Figure prediction.
confirming These timings
that our
were verified
proposed modelusing a logic
is highly analyzer,
suitable as depicted
for real-time in Figure
condition 18, confirming
monitoring that our
with relatively low pro-
posed model requirements.
computational is highly suitable for real-time condition monitoring with relatively low
computational requirements.
Sensors 2024, 24, x FOR PEER REVIEW 22 of 24
Sensors 2024, 24, 2297 21 of 23

Figure18.
Figure 18. Execution
Execution timing analysis of the meta
meta model
model implemented
implementedon
onthe
theMCU.
MCU.

4.4. Conclusions
Conclusions
This
This work
workhas hasintroduced
introducedaaprecise
preciseultrasonic
ultrasonicindustrial
industrialfault
faultdetection
detectionmeta-classifier
meta-classi-
model designed for general-purpose applications, such as
fier model designed for general-purpose applications, such as bearings and bearings and pipelines, utilizing
pipelines, uti-
contact sensors. This model is capable of determining and monitoring
lizing contact sensors. This model is capable of determining and monitoring the health the health condition
of industrial
condition machinery.
of industrial To assessTo
machinery. theassess
model’s performance,
the model’s a k-folda CV
performance, strategy
k-fold CV strat-was
employed to detect and classify ten different conditions within
egy was employed to detect and classify ten different conditions within the ultrasonic the ultrasonic signals.
The reduction
signals. methodsmethods
The reduction of RFE,ofPCA,RFE, t-SNE, and LDA
PCA, t-SNE, were were
and LDA calculated, and and
calculated, RFERFEwas
chosen
was chosenas theasinput for six
the input forclassification models
six classification (namely
models SVM,
(namely SVM, GNB, LR,LR,
GNB, KNN, KNN, DT,DT,and
the meta classifier) to discern between ten distinct classes of irregular
and the meta classifier) to discern between ten distinct classes of irregular conditions. Eachconditions. Each
model’s
model’sperformance
performancewas wasevaluated
evaluatedbased
basedon onvarious
variousmetrics.
metrics.The
Theproposed
proposedmeta metaclassifier
classi-
model demonstrated superiority with impressive results across multiple
fier model demonstrated superiority with impressive results across multiple metrics, metrics, including
in-
ROC
cludingcurves
ROC(i.e., achieving
curves an AUC an
(i.e., achieving of 0.997),
AUC ofconfusion matrices,matrices,
0.997), confusion k-fold validation (with
k-fold valida-
93%
tion accuracy),
(with 93% and learning
accuracy), andcurves.
learningAdditionally, the modelthe
curves. Additionally, boasted
modelaboasted
rapid execution
a rapid
time of justtime
execution 11 ms on a11
of just 64msMHzon aCortex-M4F MCU, confirming
64 MHz Cortex-M4F its suitability
MCU, confirming for real-time
its suitability for
industrial fault monitoring
real-time industrial applications.
fault monitoring In futureInwork,
applications. futurewe aim we
work, to extend
aim to the model’s
extend the
deployment capabilities
model’s deployment to a widertorange
capabilities of MCUs
a wider range ofto MCUs
ensure tobroader
ensureapplicability. We plan
broader applicabil-
to incorporate IoT functionalities, enabling sophisticated remote data
ity. We plan to incorporate IoT functionalities, enabling sophisticated remote data access access and control
mechanisms. Efforts will also be directed towards optimizing both
and control mechanisms. Efforts will also be directed towards optimizing both power con- power consumption
and memory
sumption andutilization, ensuring the
memory utilization, model’sthe
ensuring operation
model’s is both efficient
operation is bothand sustainable.
efficient and
Additionally, we will focus on the development of an optimized
sustainable. Additionally, we will focus on the development of an optimized PCB PCB design, specifically
design,
engineered
specifically to enhance the
engineered reliability
to enhance theand performance
reliability of the monitoring
and performance system within
of the monitoring sys-
diverse
tem withinindustrial
diverseenvironments. These targeted
industrial environments. improvements
These are intendedare
targeted improvements to elevate
intended the
model’s practicality, adaptability, and efficiency in real-world applications.
to elevate the model’s practicality, adaptability, and efficiency in real-world applications.

Author Contributions:Software,
AuthorContributions: A.M.;
software, resources,
A.M.; H.H.T.H.H.T.
resources, and M.Y.E.;
and writing—original draft prepa-
M.Y.E.; writing—original
ration, A.M.; writing—review and editing, F.N.; supervision, F.N.; funding acquisition, H.H.T. and
draft preparation, A.M.; writing—review and editing, F.N.; supervision, F.N.; project ad-
M.Y.E. All authors have read and agreed to the published version of the manuscript.
ministration, X.X.; funding acquisition, H.H.T. and M.Y.E. All authors have read and
agreed toCollaborative
Funding: the published versionand
Research of Development
the manuscript.
Grant number CRDPJ 543712-19 from the Nat-
ural Sciences and Engineering Research Council of Canada (NSERC) with MEMS-Vision as grant
Funding: Collaborative Research and Development Grant number CRDPJ 543712-19 from the Nat-
industrial partner.
ural Sciences and Engineering Research Council of Canada (NSERC) with MEMS-Vision as grant
industrial partner
Institutional Review Board Statement: Not applicable.
Institutional
Informed Review
Consent Board Statement:
Statement: Not applicable.
Not applicable.
Informed
Data ConsentStatement:
Availability Statement:Data
Notare
applicable.
contained within the article.
Data Availability
Conflicts Statement:
of Interest: AuthorsData
HaniareH.contained within
Tawfik and the article.
Mohannad Y. Elsayed were employed by the
MEMS-VISION company. The remaining authors declare that the research was conducted in the
Sensors 2024, 24, 2297 22 of 23

absence of any commercial or financial relationships that could be construed as a potential conflict
of interest.

References
1. Messer, A. Solutions for Improved Equipment Reliability. Available online: https://www.bearing-news.com/mechanical-remote-
monitoring-with-ultrasound/ (accessed on 30 November 2015).
2. Yu, Y.; Safari, A.; Niu, X.; Drinkwater, B.; Horoshenkov, K.V. Acoustic and ultrasonic techniques for defect detection and condition
monitoring in water and sewerage pipes: A review. Appl. Acoust. 2021, 183, 108282. [CrossRef]
3. Quy, T.B.; Kim, J.M. Real-time leak detection for a gas pipeline using a k-nn classifier and hybrid ae features. Sensors 2021, 21, 367.
[CrossRef] [PubMed]
4. Zhang, M.; Guo, Y.; Xie, Q.; Zhang, Y.; Wang, D.; Chen, J. Defect identification for oil and gas pipeline safety based on autonomous
deep learning network. Comput. Commun. 2022, 195, 14–26. [CrossRef]
5. Toma, R.N.; Piltan, F.; Kim, J.M. A deep autoencoder-based convolution neural network framework for bearing fault classification
in induction motors. Sensors 2021, 21, 8453. [CrossRef] [PubMed]
6. Aguayo-Tapia, S.; Avalos-Almazan, G.; Rangel-Magdaleno, J.D.J.; Ramirez-Cortes, J.M. Physical variable measurement techniques
for fault detection in electric motors. Energies 2023, 16, 4780. [CrossRef]
7. Soliman, M.H.A. Ultrasound Analysis for Condition Monitoring: Applications of Ultrasound Detection for Various In-
dustrial Equipment. 2020. Available online: https://books.google.com.my/books?hl=en&lr=&id=oM4uEAAAQBAJ&
oi=fnd&pg=PA3&dq=Ultrasound+Analysis+for+Condition+Monitoring:+Applications+of+Ultrasound+Detection+for+
Various+Industrial+Equipment&ots=LHKcKaTnK-&sig=unKxM3VZp0BRImSrJKK_uv7RJck&redir_esc=y#v=onepage&q=
Ultrasound%20Analysis%20for%20Condition%20Monitoring:%20Applications%20of%20Ultrasound%20Detection%20for%20
Various%20Industrial%20Equipment&f=false (accessed on 29 September 2021).
8. Al-Sagheer, R.H.A.; Mohammed, K.I.; Mezher, A.A.H.; Habeeban, K.A.M. Impact of crack length into pipe conveying fluid
utilizing fast fourier transform computer algorithm. Int. J. Electr. Comput. Eng. 2019, 9, 2541.
9. Jaafar, N.S.M.; Aziz, I.A.; Jaafar, J.; Mahmood, A.K. An approach of filtering to select IMFs of EEMD in signal processing for
acoustic emission [AE] sensors. In Intelligent Systems in Cybernetics and Automation Control Theory 2; Springer: New York, NY, USA,
2019; pp. 100–111.
10. Ting, L.L. Detection of Water Pipeline Leakage Using Time and Frequency Signal; UTAR: Tartu, Estonia, 2017.
11. Pandey, S.; Kumar, P.S.; Amarnath, M.; Kumar, T.T.; Rakesh, P. Incipient Fault Detection in Roller Bearing Using Ultrasonic
Diagnostic Technique. In Advances in Metrology and Measurement of Engineering Surfaces: Select Proceedings of ICFMMP 2019;
Springer: New York, NY, USA, 2021; pp. 243–251.
12. Ding, H.; Liang, Z.; Qi, L.; Sun, H.; Liu, X. Spacecraft Leakage Detection Using Acoustic Emissions Based on Empirical Mode
Decomposition and Support Vector Machine. In Proceedings of the 2021 IEEE International Instrumentation and Measurement
Technology Conference (I2MTC), Glasgow, UK, 17–20 May 2021; IEEE: New York, NY, USA, 2021; Volume 2021, p. 5. [CrossRef]
13. Coelho, J.A.; Glória, A.; Sebastião, P. Precise Water Leak Detection Using Machine Learning and Real-Time Sensor Data. IoT 2020,
1, 474–493. [CrossRef]
14. Rai, A.; Kim, J.M. A novel pipeline leak detection approach independent of prior failure information. Meas. J. Int. Meas. Confed.
2021, 167, 108284. [CrossRef]
15. Wang, H.; Ni, G.; Chen, J.; Qu, J. Research on rolling bearing state health monitoring and life prediction based on PCA and
Internet of things with multi-sensor. Measurement 2020, 157, 107657. [CrossRef]
16. Saucedo-Dorantes, J.J.; Jaen-Cuellar, A.Y.; Delgado-Prieto, M.; de Jesus Romero-Troncoso, R.; Osornio-Rios, R.A. Condition
monitoring strategy based on an optimized selection of high-dimensional set of hybrid features to diagnose and detect multiple
and combined faults in an induction motor. Measurement 2021, 178, 109404. [CrossRef]
17. Pang, S.; Yang, X.; Zhang, X.; Lin, X. Fault diagnosis of rotating machinery with ensemble kernel extreme learning machine based
on fused multi-domain features. ISA Trans. 2020, 98, 320–337. [CrossRef] [PubMed]
18. Maliuk, A.S.; Prosvirin, A.E.; Ahmad, Z.; Kim, C.H. Novel Bearing Fault Diagnosis Using Gaussian Mixture Model-Based Fault
Band Selection. Sensors 2021, 21, 6579. [CrossRef]
19. Piltan, F.; Kim, J.M. Bearing Fault Diagnosis Using a Hybrid Fuzzy V-Structure Fault Estimator Scheme. Sensors 2023, 23, 1021.
[CrossRef] [PubMed]
20. Alidadi, M.; Rahimi, A. Fault Diagnosis of Lubrication Decay in Reaction Wheels Using Temperature Estimation and Forecasting
via Enhanced Adaptive Particle Filter. Sensors 2023, 23, 1474. [CrossRef] [PubMed]
21. Abdelrhman, A.M.; Ying, L.; Ali, Y.H.; Ahmad, I.; Georgantopoulou, C.G.; Nor, F.M.; Kurniawan, D. Diagnosis model for
bearing faults in rotating machinery by using vibration signals and binary logistic regression. In AIP Conference Proceedings; AIP
Publishing: New York, NY, USA, 2020; Volume 2262. [CrossRef]
22. Jiang, Q.; Dai, J.; Shao, F.; Song, S.; Meng, F. Bearing Early Fault Diagnosis Based on an Improved Multiscale Permutation Entropy
and SVM. Shock Vib. 2022, 2022, 2227148. [CrossRef]
23. Da Silva Souza, J.; Santos, M.V.L.D.; Bayma, R.S.; Mesquita, A.L.A. Analysis of Window Size and Statistical Features for
SVM-based Fault Diagnosis in Bearings. IEEE Lat. Am. Trans. 2021, 19, 243–249. [CrossRef]
Sensors 2024, 24, 2297 23 of 23

24. Jung, W.; Bae, J.; Park, Y.H. Acoustic Signal Based Non-Contact Ball Bearing Fault Diagnosis using Adaptive Wavelet Denoising;
Department of Mechanical Engineering, Korea Advanced Institute of Science and Technology: Daejeon, Republic of Korea, 2021;
p. 34141, South. no. 20204030200050.
25. Moshrefi, A.; Gratuze, M.; Tawfik, H.H.; Elsayed, M.Y.; Nabki, F. Ensemble AI Fault Diagnosis Model Using Ultrasonic
Microphone. In 2023 IEEE International Ultrasonics Symposium (IUS), Montreal, QC, Canada, 3–8 September 2023; IEEE: New York,
NY, USA, 2023; pp. 1–3. [CrossRef]
26. Selesnick, I.W.; Burrus, C.S. Generalized digital Butterworth filter design. IEEE Trans. signal Process. 1998, 46, 1688–1694.
[CrossRef]
27. Thara, D.K.; PremaSudha, B.G.; Xiong, F. Auto-detection of epileptic seizure events using deep neural network with different
feature scaling techniques. Pattern Recognit. Lett. 2019, 128, 544–550.
28. Postma, E.; Postma, E. Dimensionality Reduction: A Comparative Review Dimensionality Reduction: A Comparative Review. J.
Mach. Learn. Res. 2009, 10, 13.
29. Agis, D.; Pozo, F. A frequency-based approach for the detection and classification of structural changes using t-SNE. Sensors 2019,
19, 5097. [CrossRef]
30. Devassy, B.M.; George, S.; Nussbaum, P. Unsupervised clustering of hyperspectral paper data using t-SNE. J. Imaging 2020, 6, 29.
[CrossRef] [PubMed]
31. Adebiyi, M.O.; Arowolo, M.O.; Mshelia, M.D.; Olugbara, O.O. A Linear Discriminant Analysis and Classification Model for
Breast Cancer Diagnosis. Appl. Sci. 2022, 12, 11455. [CrossRef]
32. Granitto, P.M.; Furlanello, C.; Biasioli, F.; Gasperi, F. Recursive feature elimination with random forest for PTR-MS analysis of
agroindustrial products. Chemom. Intell. Lab. Syst. 2006, 83, 83–90. [CrossRef]
33. Ratio, E.; Discriminant, L. Fault Diagnosis of a Multistage Centrifugal Pump Using Explanatory Ratio Linear Discriminant
Analysis. Sensors 2024, 24, 1830. [CrossRef] [PubMed]
34. Yadavendra; Chand, S. A comparative study of breast cancer tumor classification by classical machine learning methods and
deep learning method. Mach. Vis. Appl. 2020, 31, 46. [CrossRef]
35. Chandra, M.A.; Bedi, S.S. Survey on SVM and their application in imageclassification. Int. J. Inf. Technol. 2021, 13, 1–11. [CrossRef]
36. Ontivero-Ortega, M.; Lage-Castellanos, A.; Valente, G.; Goebel, R.; Valdes-Sosa, M. Fast Gaussian Naïve Bayes for searchlight
classification analysis. Neuroimage 2017, 163, 471–479. [CrossRef] [PubMed]
37. Ohsaki, M.; Wang, P.; Matsuda, K.; Katagiri, S.; Watanabe, H.; Ralescu, A. Confusion-matrix-based kernel logistic regression for
imbalanced data classification. IEEE Trans. Knowl. Data Eng. 2017, 29, 1806–1819. [CrossRef]
38. Kamiński, B.; Jakubczyk, M.; Szufel, P. A framework for sensitivity analysis of decision trees. Cent. Eur. J. Oper. Res. 2018, 26,
135–159. [CrossRef]
39. Nti, I.K.; Nyarko-Boateng, O.; Aning, J. Performance of Machine Learning Algorithms with Different K Values in K-fold
CrossValidation. Int. J. Inf. Technol. Comput. Sci. 2021, 13, 61–71. [CrossRef]
40. Wong, T.-T.; Yeh, P.-Y. Reliable accuracy estimates from k-fold cross validation. IEEE Trans. Knowl. Data Eng. 2019, 32, 1586–1594.
[CrossRef]
41. Sultana, N.; Islam, M.M. Meta classifier-based ensemble learning for sentiment classification. In Proceedings of the International
Joint Conference on Computational Intelligence: IJCCI 2018, Seville, Spain, 18–20 September 2018; Springer: New York, NY, USA,
2020; pp. 73–84.
42. Sravani, S.; Karthikeyan, P.R. Detection of cardiovascular disease using KNN in comparison with naive bayes to measure
precision, recall and f-score. In AIP Conference Proceedings; AIP Publishing: New York, NY, USA, 2023; Volume 2821, no. 1.
43. Reddy, B.H.; Karthikeyan, P.R. Classification of fire and smoke images using decision tree algorithm in comparison with logistic
regression to measure accuracy, precision, recall, F-score. In Proceedings of the 14th International Conference on Mathematics,
Actuarial Science, Computer Science and Statistics (MACS), Karachi, Pakistan, 12–13 November 2022; IEEE: New York, NY, USA,
2022; pp. 1–5.
44. Available online: https://www.uesystems.com/resources/sound-library/ (accessed on 1 January 2020).

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.

You might also like