J. Appl. Comput. Mech.

, xx(x) (2025) 1-13 ISSN: 2383-4536

DOI: 10.22055/jacm.2025.48052.4943 jacm.scu.ac.ir

Journal of
Shahid Chamran Applied and Computational Mechanics
University of Ahvaz

Research Paper

Application of Machine Learning Techniques for

Bearing Fault Diagnosis

Sarra Eddai1,2 , Nabih Feki1,3 , Ahmed Ghorbel1,4 , Abdelkhalak El Hami2 , Mohamed Haddar1

1
Laboratory of Mechanics, Modeling, and Production, National School of Engineering of Sfax, University of Sfax, Sfax, Tunisia,
Email: [email protected] (S.E.); [email protected] (M.H.)
2
Normandie Mechanical Laboratory LMN, National Institute of Applied Sciences of Rouen, University of Rouen, Haute Normandie, France,
Email: [email protected]
3
Department of Mechanical Engineering, Higher Institute of Applied Sciences and Technology of Sousse, University of Sousse, Sousse, Tunisia,
Email: [email protected]
4
Department of Mechanical Engineering, Higher Institute of Applied Sciences and Technology of Kairouan, University of Kairouan, Kairouan, Tunisia,
Email: [email protected]

Received November 29 2024; Revised January 22 2025; Accepted for publication February 02 2025.
Corresponding author: N. Feki ([email protected])
© 2025 Published by Shahid Chamran University of Ahvaz

Abstract. Machine learning enhances machine diagnostics through advanced data analysis, pattern recognition, and fault
prediction. This study investigates the application of machine learning algorithms for bearing fault detection. The objective is to
develop intelligent methodologies for the predictive diagnosis of bearing faults in rotating machinery, emphasizing the significance
of timely intervention to prevent critical failures. The methodology employed encompasses a systematic approach, including data
preprocessing, feature extraction, and model development. This research employs advanced machine learning techniques,
including Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, and Naive Bayes algorithms, in conjunction
with time-domain and frequency-domain feature extraction methods. The implemented approach substantially enhances fault
detection accuracy, achieving an aggregate classification precision of 97.8% across all fault categories. Notably, the SVM algorithm
demonstrates exceptional performance, attaining a 99.2% accuracy rate in inner-race fault identification. This investigation provides
a comprehensive analysis of the Case Western Reserve University (CWRU) dataset, data preprocessing procedures, feature
extraction techniques, and machine learning algorithms utilized for fault detection. The results emphasize the effectiveness of
these algorithms in bearing fault diagnosis, offering valuable insights for predictive maintenance strategies in industrial
applications. This research also aligns with the objectives of Industry 4.0, which focuses on utilizing intelligent, automated systems
to enhance factory efficiency and reliability. The study concludes by proposing future research directions to further advance these
technologies and support the transition toward more intelligent, interconnected industries.

Keywords: Bearings faults, intelligent diagnosis, machine learning techniques, features extraction, Artificial Intelligence (AI).

1. Introduction
Rolling bearings are essential components in rotating machinery, bearing substantial workloads due to the evolving demands
of diverse industries. However, these bearings are prone to faults under high loads, impacts, and challenging environments.
Traditional fault detection methods often rely on manual inspections or sensor-based monitoring systems, which can be time-
consuming and may not offer real-time insights [1-4]. The collected signals for fault diagnosis in rotating machinery, particularly in
bearing systems, are typically analyzed using traditional methods based on vibration signals. These methods involve extracting
features from the signals and then applying pattern recognition techniques for classification [5, 6]. Artificial intelligence (AI) has
become a prominent focus for many researchers due to its powerful pattern recognition capabilities. It shows significant promise
in bearing fault recognition applications, particularly with the advent of machine learning techniques, which have proven to be
highly effective tools for automated fault detection [7].
The field of bearing fault diagnosis has seen significant advancements, with intelligent fault diagnosis methods being widely
accepted for addressing complex mechanical diagnosis challenges. These methods are prioritized for their adaptive learning
mechanisms; strong fault tolerance, and high non-linear regression capabilities [8]. Intelligent fault diagnosis methods encompass
three main steps: signal acquisition, feature extraction, and fault classification [9]. Signal Acquisition involves collecting various
signals such as vibration, acoustics, electrical current, speed, and temperature, with vibration signals playing a crucial role in
providing intrinsic information about mechanical equipment's health conditions. Feature Extraction is the process of capturing
relevant information from the acquired signals to identify faults’ indicating patterns. Techniques like wavelet transform and

Published online: February 22 2025

 2 Sarra Eddai et al., Vol. x, No. x, 2025

variational mode decomposition are commonly employed to extract meaningful features from signals.
Following Feature Extraction, Fault Classification utilizes intelligent classification techniques such as Support Vector Machine
(SVM), K-Nearest Neighbors (KNN), and other algorithms based on extracted features for accurate fault classification. The
integration of advanced methods like Convolutional Neural Networks (CNNs) further enhances feature learning and fault diagnosis
capabilities in condition monitoring applications [10]. Here are some key principles of the algorithms commonly used in bearing
fault diagnosis:
Support Vector Machine (SVM) is a powerful algorithm that aims to find the optimal hyperplane to separate data points into
different classes. This algorithm is widely used in fault detection applications for rotating machinery, particularly in bearing
systems. SVM is known for its ability to effectively handle complex data patterns and non-linear relationships. Research by Yang
on rotating machinery fault diagnosis based on fast clustering algorithms and SVM highlighted the algorithm's utility in analyzing
imbalanced data for accurate fault detection [11, 12].
K-Nearest Neighbors (KNN) is another essential algorithm in the realm of fault detection for bearings. KNN is based on the
principle of similarity, where instances are classified by their proximity to neighboring data points. Studies emphasized the
significance of KNN in achieving better machine performance and intelligent diagnosis of rotating machinery, especially in
addressing bearing faults, which area primary cause of breakdowns in rotating machinery [13].
Decision Tree algorithms are valuable tools for fault diagnosis in rotating machinery, offering a visual representation of decision-
making processes [14]. The study focusing on the usage of KNN, SVM, Decision Trees, and Naïve Bayes algorithms underscored the
importance of these algorithms in data science projects and machine learning applications [15]. Decision Trees provide insights
into the statistical and mathematical aspects of fault classification and are instrumental in identifying patterns indicative of faults
in bearing systems.
Naive Bayes is a probabilistic classifier based on Bayes' theorem that assumes independence between features. Its application
in fault detection for bearings may involve leveraging probabilistic reasoning to classify faults.
The integration of advanced methods like Convolutional Neural Networks (CNNs) further enhances feature learning and fault
diagnosis capabilities in condition monitoring applications. Deep learning, as a subset of machine learning, has shown significant
promise in bearing fault recognition applications, often outperforming traditional methods. Deep learning algorithms can
effectively extract features from vibration signals and perform advanced pattern recognition for automated fault detection [16-18].
The application of these machine learning and deep learning techniques for bearing fault diagnosis can significantly improve
the efficiency and reliability of condition monitoring in rotating machinery, enabling predictive maintenance and reducing
downtime and economic losses.
These machine learning algorithms such as Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Naive Bayes, Decision
Trees, Random Forests, Logistic Regression, and Artificial Neural Networks (ANNs) play pivotal roles in fault detection applications
for bearings by enabling accurate signal analysis, feature extraction, and fault classification processes. Researchers leverage these
algorithms to enhance predictive maintenance practices, improve machinery reliability, and minimize unexpected breakdowns in
rotating machinery systems.
Several studies have utilized the Case Western Reserve University (CWRU) bearing dataset to develop and evaluate intelligent
fault diagnosis systems. This dataset contains vibration signals collected from bearings under different fault conditions, providing
a valuable resource for researchers working on this topic. Raj et al. [19] demonstrated the effectiveness of machine learning models
like Convolutional Neural Networks (CNNs) in accurately detecting and classifying bearing faults using raw vibration data from the
CWRU dataset. They were able to achieve high classification accuracies, particularly under different loading conditions. Jia et al. [20]
also leveraged the CWRU dataset to develop a deep learning-based approach for bearing fault diagnosis. They used a combination
of Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks to capture both spatial and temporal
features in the vibration data, leading to improved fault detection performance.
This study aims to explore using machine learning algorithms to detect faults in bearings, focusing specifically on the CWRU
dataset. Intelligent fault diagnosis systems commonly employ preprocessing techniques and feature extraction algorithms to
transform input patterns into low-dimensional feature vectors, facilitating easier matching and comparison of data. These feature
vectors are then utilized as input for AI techniques to recognize faults, with the fault recognition step mapping information from
the feature space to machine faults in the fault space.
In this study, we focused on utilizing parameters extracted from signals obtained from the specific test bench under
investigation. While signal-processing methods such as time-domain analysis and frequency-domain analysis were commonly
employed in fault recognition preprocessing according to the literature, it is important to note that our study specifically delved
into time-domain and frequency-analysis techniques. Time-domain analysis involves applying computations such as mean,
variance, and kurtosis estimation directly on the time signal to improve fault recognition processes [21]. Frequency-domain
analysis, including methods like the fast Fourier transform (FFT) and bispectrum analysis, operates on signals transformed into the
frequency domain [22]. Conversely, time-frequency analysis techniques such as the short-time Fourier Transform (STFT), Hilbert-
Huang Transform (HHT) Wavelet Transform, and Wavelet Packet Transform (WPT) are frequently utilized for feature extraction in
fault recognition applications [23].
This research work focuses on comparing the performance of machine learning algorithms, including Support Vector Machines
(SVM), K-Nearest Neighbors (KNN), Naive Bayes, and Decision Trees, to develop a robust fault detection model for bearings. We aim
to compare the performance of these algorithms in accurately identifying and classifying different bearing fault conditions, such
as normal operation, inner race fault, outer race fault, and ball fault, using the CWRU bearing dataset.
By utilizing feature extraction techniques, including time-domain and frequency-domain analysis, we will generate low-
dimensional feature vectors to facilitate easier comparison and analysis of the data. The performance evaluation of these machine
learning algorithms will concentrate on their capability to accurately identify and classify faults in bearings. Moreover, a detailed
analysis of various feature extraction techniques will be conducted to pinpoint the most informative features extracted from the
vibration signals, which is called feature selection. This analysis is crucial for precise fault detection as it will help determine which
features provide the most valuable information for accurate fault classification. By combining these advanced machine learning
algorithms with effective feature extraction and selection methods, the study aims to enhance fault detection accuracy and
efficiency in bearing systems, ultimately contributing to improved condition monitoring and predictive maintenance of rotating
machinery.
The remainder of this article is organized as follows: Section 2 describes the CWRU dataset, and the data preprocessing steps,
including filtering and signal decomposition techniques. Section 3 presents the feature extraction techniques employed in this
study, covering both time domain and frequency domain features. Section 4 discusses the data classification process and the
assessment of classification performance, as well as the comparative analysis of enhancing classification performance through
feature selection in machine learning. Finally, Section 5 concludes the article and discusses future research directions.

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 Application of machine learning techniques for bearing fault diagnosis 3

Torque
Acceleromete Transducer
Drive end
Fan / Encoder
bearing
end
bearin
g

Electric engine Dynamomete

Fig. 1. The experimental configuration of the ball bearing system on the CWRU bearing test rig [24].

Fig. 2. Vibration signals collected under different rolling bearing conditions for a load of 3 hp (a) without defect (b) Bearing with a 0.007-inch defect
(c) Bearing with a 0.014-inch defect (d) Bearing with a 0.021-inch defect.

2. Experiment Setup and Signal Acquisation

2.1. Case study
The dataset is provided by the Bearing Data Center of Case Western Reserve University. The tested bearings support the motor
shaft on both the drive side and the fan side. Figure 1 present the test rig, which mainly consists of a 2 hp motor (on the left), a
transducer/encoder coupling (in the center), a dynamometer (on the right), and control circuits. Simple defects in the form of points
with different diameters and depths were created on the various components of the ball bearings using electrical discharge
machining. The defect diameters are, respectively, 0.007, 0.014, and 0.021 Inches. These diameters, along with the healthy case, form
the four shape recognition classes.
Vibration data was collected using accelerometers, which were attached to the housing with magnetic bases. Accelerometers
were placed at the 12 o’clock position at both the drive end and fan end of the motor housing. During some experiments, an
accelerometer was attached to the motor-supporting base plate as well. Digital data was collected at 48,000 samples per second for
drive end-bearing faults. Speed and horsepower data were collected using the torque transducer/encoder and were recorded by
hand [24].
The CWRU bearing dataset has been instrumental in advancing the field of condition monitoring and fault diagnosis for rotating
machinery. Through a variety of innovative approaches, researchers have leveraged this dataset to develop and evaluate advanced
methodologies for bearing fault detection, which have significantly contributed to both theoretical and practical advancements in
the field [19, 25, 26]. The developed methodologies through these works have not only refined the state of the art in fault detection
and improved the practice of condition-based maintenance of rotating machinery systems.
The signals we propose to study are recorded for faults in the inner race at a sampling frequency of fs = 48 kHz under the
following conditions:
• The shaft rotates at the respective rotating frequencies: 1797 rpm, 1772 rpm, 1750 rpm, 1730 rpm.
• The power of the motor is: 0 HP, 1 HP, 2 HP, and 3 HP.
The signals are sampled from the drive end bearing 6203-2RS JEM SKF, the database includes four signals: one intact signal and
three others with faults, distributed across four distinct operating conditions as shown in Fig. 2. It is worth noting that we are
working with three operating speeds, each associated with a specific load.
2.2. Proposed method
The strategy proposed for diagnosing faults in the inner race of bearings, a critical machine component, follows a structured
approach depicted in Fig. 3. Fault diagnosis is treated as a pattern-recognition challenge and is segmented into five key steps: data
acquisition, data processing, feature extraction, training, and classification. The extracted features will serve as inputs for machine
learning algorithms. The extracted features serve as inputs for machine learning algorithms. Data is collected, and vibration signal
features are extracted based on fault severity. Signal decomposition plays a vital role in data processing, while feature extraction is
pivotal for analyzing both time and frequency domains by capturing vibration features linked to fault severity, including RMS, peak-
to-peak value, skewness, kurtosis, and energies from the envelope spectrum across frequency domains [28]. Existing methods for

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 4 Sarra Eddai et al., Vol. x, No. x, 2025

bearing fault diagnosis often face challenges in handling non-stationary signals, varying operating conditions, and high
computational demands. Deep learning approaches, although effective in accuracy, require significant computational resources
and extensive datasets, which limit their practical application in resource-constrained environments. Conversely, traditional
threshold-based methods lack robustness and adaptability to dynamic industrial scenarios.
The proposed method addresses these limitations by combining computationally efficient algorithms with interpretable
features, such as RMS and kurtosis, which are directly linked to the mechanical state of the bearing. Time domain calculations are
conducted within specific frequency bands derived from a 6 kHz analysis band, while frequency domain features encompass
energies from the envelope spectrum computed within various slices and frequency bands spanning from 0 to 6000 Hz. Time
domain indicators such as RMS, peak-to-peak value, skewness, kurtosis, and envelope spectrum energies are leveraged to evaluate
the bearing's condition. Furthermore, the extraction of frequency domain features offers a holistic perspective on the bearing's
overall health status. This fault diagnosis methodology include its ability to handle varying operating conditions effectively, its
computational efficiency compared to deep learning models, and its interpretability, which facilitates practical implementation in
industrial settings. This fault diagnosis methodology emphasizes pattern recognition and involves sequential steps of data
acquisition, processing, feature extraction, and classification training. Details of the main steps are explained in the following
subsections.
2.3. Techniques for data processing
In the realm of vibration signal analysis, data preprocessing plays a crucial role in enhancing fault recognition accuracy by
eliminating irrelevant information. Two key techniques commonly used for data processing are filtering and signal decomposition
[29].
1. Filtering
Filtering is essential for removing unwanted parts from vibration signals. By using a Finite Impulse Response (FIR) filter with a
Hamming window, the masking effect can be minimized. This technique involves defining indicators in narrow bands to prevent
the masking effect and calculating time and frequency measures in 2000 Hz bands [30].
2. Signal decomposition
Signal decomposition is a powerful technique that involves deconstructing the signal into various components or frequency
bands to extract valuable information for fault identification. This process enables researchers to extract frequency-specific
characteristics, generate a more comprehensive dataset by treating each decomposed component as an independent data point,
and perform detailed temporal and frequency analyses for accurate fault diagnosis. One of the primary advantages of signal
decomposition lies in its capacity to enrich the dataset, as each decomposed component can be treated as a discrete data point,
effectively increasing the size of the training dataset for machine-learning algorithms [31]. By analyzing each component
independently, temporal and frequency indicators can be calculated, which facilitates the characterization of signal properties
across different frequency ranges. This comprehensive analysis enhances the ability to accurately identify mechanical faults based
on specific characteristics such as frequency, amplitude, and waveform changes.
In this study, the signals were recorded at a sampling frequency of fs = 48 kHz, ensuring high-resolution capture of the vibration
data. Recordings (signals) with a length of points NS were divided into Kt slices of NM points with a 50% overlap S1. The 50% overlap
ensures that sufficient data cycles are retained within each segment while maintaining a balance between redundancy and
computational efficiency. The number of samples for each slice was carefully selected to ensure that it was sufficiently large to
encompass a sufficient number of cycles, corresponding to the longest period (the slowest rotation speed).
This method is applied to decompose signals recorded at four different speeds [32]. This approach not only enables a more
granular analysis of the vibration signals but also facilitates dataset augmentation by segmenting the raw signals into smaller
segments while maintaining sufficient data points per segment. This augmentation ensures adequate representation of signal
cycles, even at the slowest rotation speeds, and provides sufficient data volume for the effective training of machine learning
models.
The shaft rotates at a speed of 1730 rpm frShaft = 25.83 Hz with a signal length equal to 485643 points. The number of points per
period is:

fS
Np = ≈ 4182 Points (1)
f(cage )

If we use a coefficient of 5, the length required for each slice NT would be equal to 5. NP. The chosen value of NM is equal to 215.
While the number of slices in the case of a 50% overlap would be NM/2. Then, the number of signals can be expressed by:

(Ns − S1 )
Kt = (2)
NM − S1

We have a total of 112 signals, with 28 signals corresponding to each class. For each shaft, the same procedure is duplicated,
and the findings are meticulously recorded in Table 1.

3. Feature Extraction
In this step, we are considering the common, robust, and effective methods of feature extraction for both vibration and current
signals. Table 2 provides a summary of 11 extracted features in this study.
3.1. Feature in time domain
Among the features in the time domain, statistical indicators are commonly used because of their strong correlation with the
early bearing damage. The standard deviation (SD) helps understand the fluctuation in the vibration signal. This way, it is possible
to determine its stability and consistency. The peak-to-peak value (Pk-Pk) indicates the distance from one extreme of a swing or
movement to another. This provides an all-around view of the variations in the amplitude. Moreover, we consider the skewness (Ks)
rates as the asymmetry of the signal distribution. We also consider kurtosis (Ku), which shows how peaked or flat the distribution
is. These indicators improve our understanding of the bearing vibrations. They allow accurate evaluation of the bearings' condition
and the detection of probable faults.

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 Application of machine learning techniques for bearing fault diagnosis 5

Table 1. The Number and length of signals in each segment.

Speed Torque Signal Number of signals (Inner race) (Kt) Signals length (NS)
Bearing without defect
Bearing with a 0.007-inch defect
1730 rpm 3 HP 112 signals 485643
Bearing with a 0.014-inch defect
Bearing with a 0.021-inch defect
Bearing without defect
Bearing with a 0.007-inch defect
485643
1750 rpm 2 HP Bearing with a 0.014-inch defect 232 signals
Bearing with a 0.021-inch defect
Bearing without defect
Bearing with a 0.007-inch defect
1772 rpm 1 HP 180 signals 381890
Bearing with a 0.014-inch defect
Bearing with a 0.021-inch defect
Bearing without defect
Bearing with a 0.007-inch defect
1797 rpm 0 HP 112 signals 243938
Bearing with a 0.014-inch defect
Bearing with a 0.021-inch defect

Table 2. Equations of the extracted time and frequency domain features.

Type Formula
n
1
Standard Deviation (SD) SD =
n −1
∑ (x i
− x)
2

i= 1

Peak-to-peak value (Pk-Pk) Acc = Amax − Amin

1 n
∑ ( xi − x )3
N i =1
Skewness (Ks) Ks =
1 n
( ∑ ( xi − x ) )
2 3

N i=1
n
1
N
∑ (x i
− x)
4

i =1
Kurtosis (Ku) Ku = n
1
(
N
∑ (x i
− x) )
2 2

i =1

Root mean square (RMS) ∑ (x i

− x)
i =1
RMS =
N

2
Energy
Ex
frequency
= ∑ X( f )


X(f) is the signal equation in frequency domain.

Hence, a collection of 5 time-domain features is generated utilizing the equations provided in Table 2. The parameter represent
a sample in the acquired signal and N defines the total number of samples.
3.2. Feature in frequency domain
In the frequency domain analysis, which is commonly utilized for bearing condition monitoring due to the distinct frequencies
associated with bearing defects in vibration and current signals, energies from the envelope spectrum are employed. This technique
allows for a detailed examination of the vibration energy distribution across different frequency ranges and provides valuable
insights into the characteristics and patterns of the vibration signals.
Figures 4 to 14 depict the evolution of features as a function of samples. We plotted 11 indicators against 636 samples and the
defect sizes in the inner race of the bearings for various rotation frequencies. The first 159 samples correspond to the normal state
without defects. From the 318th sample onwards, the signal exhibits a 0.007-inch defect. Subsequent samples show defects of 0.014
inches up to the 477th sample. The mapping observations indicate defects of 0.021 inches. By analyzing these extracted features,
valuable insights into how different indicators vary with speed and defect size can be gained. This study aims to identify the most
effective indicator for detecting bearing defects and tracking their progression to enhance bearing monitoring using vibration
indicators. It also seeks to develop effective combinations to improve monitoring.
3.3. Discussion
Quantitative metrics exhibit varying patterns based on the severity of the bearing faults. Typically, in a healthy bearing, the
kurtosis value is generally around 3. However, in the presence of a fault, this value increases proportionally with the severity of the
fault. The kurtosis progressively increased with increasing fault size, reaching a peak before declining at a specific fault size. As the
bearing deteriorates, kurtosis variation becomes more pronounced. Similarly, the Root Mean Square (RMS) value increases with the
presence of faults and continues to increase as the degradation progresses. It initiates a decline once the fault reaches a specific
size. Additionally, the peak-to-peak variation exhibits an increase with the fault presence and escalates as the deterioration
advances. Contrary to the existing literature suggesting kurtosis and crest factors as highly sensitive fault indicators, our research
findings challenge this notion, indicating that they are less sensitive to operational changes [33]. However, like other conventional
time-domain indicators, kurtosis still responds to changes in load conditions. Frequency-domain indicators derived from the
vibration signal envelope spectra significantly enhance the fault detection. Figures 13 and 14 illustrate the evolution of these
indicators in fault scenarios. In particular, EB4 and EB5 effectively differentiated between the healthy and faulty bearings. Notably,
Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 6 Sarra Eddai et al., Vol. x, No. x, 2025

a minimal distinction was observed between a healthy bearing and bearing with a minor fault. Significant decreases in EB2 and EB3
values are evident only with severe faults of 0.014 inches. Conversely, EB1 and EBT values increase with fault presence but decrease
once the fault size reaches 0.0014 inches. In summary, this study emphasizes the significance of various indicators such as RMS,
standard deviation, peak-to-peak, kurtosis, and frequency-domain indicators (EBT, EB1, EB2, and EB3) in bearing fault detection.
This highlights the nuanced behavior of these indicators concerning fault severity and operational conditions, underscoring their
relevance for classification purposes.
A dedicated MATLAB program was developed to efficiently segment signals and compute both temporal and frequency
indicators, streamlining the analysis process by automating the segmentation and indicator calculations.

4. Data Classification Process

In the final step of the data classification process, the algorithm utilizes a database divided into four classes based on Table 3,
representing various fault states within the system, including bearings without defects and different fault levels. The target vector
components, reflecting known fault levels, are crucial for efficient fault detection. To maintain a balanced approach and enhance
the classification accuracy, an equal number of samples (159) is allocated to each state.

Fig. 3. Flowchart of the vibration data processing and fault classification.

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 Application of machine learning techniques for bearing fault diagnosis 7

Fig. 4. Standard deviation distribution. Fig. 5. Peak to peak distribution.

Fig. 6. RMS distribution. Fig. 7. Skewness distribution.

Fig. 8. Kurtosis distribution of the data. Fig. 9. EBT distribution of the data.

Fig. 10. EB1 distribution of the data. Fig. 11. EB2 distribution of the data.

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 8 Sarra Eddai et al., Vol. x, No. x, 2025

Fig. 12. EB3 distribution of the data. Fig. 13. EB4 distribution of the data.

Fig. 14. EB5 distribution of the data.

To further enhance the classification process, k-fold cross-validation is employed. This technique involves dividing the dataset
into k subsets, training the model on k-1 subsets, and validating it on the remaining subsets [34]. This process is repeated k times,
with each subset used once as a validation set. By averaging the results over these iterations, the model's performance is more
robust and reliable.
During the classification phase, the algorithm utilizes features such as RMS values, kurtosis, standard deviation, peak-to-peak,
and frequency domain energies (EBT, EB1, EB2, and EB3) extracted from the vibration signals to classify the fault severity levels. This
diverse feature set offers valuable insights into bearing conditions and aids in accurate fault classification.
For model training, Support Vector Machine, K-Nearest Neighbor, decision tree, and Naïve bayes classifiers are employed. Each
fault state is represented by a sample size of 159 to ensure a balanced dataset for training. This balanced representation across fault
states enhances the robustness and reliability of classification models, enabling effective differentiation between various fault
severity levels in the system.
4.1. Classification performance assessment
The confusion matrix and related metrics are effective tools for analyzing the performance of binary and multiclass
classification tasks in machine learning [35]:
TP + TN
Accuracy = × 100% (3)
TP + FN + FP + TN

with TP defines the true positive, TN is the true negative, FP is the false positive, FN is the false negative.
4.2. Feature selection for enhanced machine learning classification: A comparative analysis
In this study, three datasets comprising diverse features in the time and frequency domains were utilized to enhance
classification performance through feature selection in machine learning. These datasets include varying combinations of
indicators, such as standard deviation, root mean square (RMS), peak-to-peak, kurtosis, and energy-based metrics (EBT, EB1, EB2,
EB3), as detailed in Table 4. The objective was to evaluate the efficacy of feature selection methods in accurately categorizing the
severity of bearing faults. The performance of four classification algorithms Support Vector Machine (SVM), K-Nearest Neighbor
(KNN), Decision Tree, and Naïve Bayes—was assessed on these datasets to determine their accuracy and effectiveness in fault
classification. By comparing the classifiers' performance across distinct datasets and investigating different combinations of
indicators, the study aimed to identify the most informative features for accurate fault diagnosis. This research aligns with recent
advancements in feature selection methodologies for bearing fault detection and highlights the importance of relevant feature
selection in improving classification reliability.
Table 3. Cases of the variable output (state).
Classes State Samples
0 Bearing without defect 159
1 Bearing with a 0.007-inch defect 159
2 Bearing with a 0.014-inch defect 159
3 Bearing with a 0.021-inch defect 159

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 Application of machine learning techniques for bearing fault diagnosis 9

Table 4. Datasets table.

Dataset 1 Dataset 2 Dataset 3

Standard Deviation Standard Deviation Peak-to-peak
Root Mean Square Root Mean Square Root Mean Square
EBT Kurtosis
EB1 EB3
EB2
EBT

Table 5. Experimental results on the different Datasets.

Accuracy
Algorithms
Dataset 1 Dataset 2 Dataset 3
K-NN 98,9% 98% 88 ,1%
Naïve bayes 98,7% 95 ,8% 51,7%
SVM 98,9% 98 ,9% 90 ,9%
DT 98,7 % 97% 87 ,1%

The K-Nearest Neighbor (KNN) algorithm exhibited consistently high performance across all datasets, achieving accuracy rates
of 98.9%, 98%, and 88.1% for datasets 1, 2, and 3, respectively. These results demonstrate the algorithm's effectiveness in classifying
faults under various feature combinations. The Naïve Bayes algorithm, while yielding reasonable results, displayed greater
sensitivity to dataset characteristics. Its accuracy ranged from 98.7% to 51.7%, indicating that the algorithm may face challenges in
generalizing to certain datasets. The Support Vector Machine (SVM) algorithm, on the other hand, demonstrated robust
performance, with accuracy rates of 98.9%, 98.9%, and 90.9% for datasets 1, 2, and 3, respectively. This suggests that SVM can
effectively handle diverse data configurations. Finally, the Decision Tree algorithm also achieved satisfactory results, with accuracy
rates ranging from 98.7% to 87.1%. However, its performance appeared more sensitive to the complexity of the datasets, indicating
potential for further optimization.
Comparing the datasets, Dataset 1, which includes four indicators (EB2, EBT, standard deviation, and RMS), produced the highest
accuracy rates, ranging from 98.7% to 98.9%, across all algorithms. This suggests that the selected indicators in Dataset 1 are highly
informative for fault classification. Dataset 2, composed of three indicators (EB1, standard deviation, and RMS), also demonstrated
strong classification performance, with accuracy rates ranging from 95.8% to 98.9%, highlighting the relevance of these features for
fault diagnosis. In contrast, Dataset 3, which includes five indicators (peak-to-peak, kurtosis, RMS, EB3, and EBT), exhibited slightly
lower performance, with accuracy rates ranging from 51.7% to 90.9%. This indicates that the additional indicators in Dataset 3 may
contribute less effectively to classification accuracy compared to the indicators in Datasets 1 and 2.
In conclusion, the performance of classification algorithms varied depending on the datasets and indicators used. Datasets 1
and 2 consistently provided high accuracy for all algorithms, emphasizing the importance of selecting relevant and informative
features for fault classification. In contrast, Dataset 3, despite including more features, showed lower accuracy, demonstrating that
increasing the number of features does not necessarily improve classification performance. These findings underscore the need for
a careful feature selection process to optimize machine learning models for bearing fault detection.
From Table 5 and Fig. 15, it can be seen that the average test accuracy indicates that the k-nearest neighbors, Naïve Bayes, SVM,
and decision tree algorithms perform well. However, certain algorithms may be better suited to specific datasets. Therefore, it is
imperative to consider these findings when choosing a classification algorithm for future studies. Additionally, paying attention to
the unique characteristics of the dataset is essential, as they can significantly impact the overall performance, ultimately leading
to optimal results.
The confusion matrix obtained for Dataset 1 in Figs. 16 and 17 showcasing the performance of the Support Vector Machine
(SVM) and K-Nearest Neighbor (KNN) algorithms with an accuracy rate of 98.9%, provides valuable insights into the classification
results. The matrix reveals that, out of all instances, the algorithms correctly classified the majority of the samples, with only a few
instances leading to misclassification. The two errors recorded in the classification process are attributed to the confusion between
the fault cases of 0.014 and 0.021 inches. These errors highlight the challenges in distinguishing faults with similar vibrational
characteristics.

100%
90%
80%
70%
Accuarcy

60%
50%
40%
30%
20%
10%
0%
K-Nearst Naïve bayes Support Vector Desicion Tree
Neighbour Machine

Dataset 1 Dataset 2 Dataset 3

Fig. 15. Accuracy of four ML models.

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 10 Sarra Eddai et al., Vol. x, No. x, 2025

Furthermore, the presence of confusion in the classification matrix can be scientifically explained by considering the
relationship between the depth of the fault and the contact deflection in Figs. 18, 19 and 20. When the depth of the fault exceeds
the contact deflection, the vibrational signature may not exhibit significant variations, particularly when comparing deep faults.
This observation emphasizes the importance of understanding the physical characteristics of bearing faults in interpreting the
classification results based on vibrational signatures [36].
By incorporating these classification challenges into fundamental physical principles of bearing faults, this study provides a
deeper understanding of the complexities involved in fault detection and classification. This integration of physical insights with
machine learning outcomes emphasizes the need for a comprehensive approach that combines data-driven techniques with
domain knowledge to enhance the reliability and accuracy of fault diagnosis.
Despite some minor errors, both SVM and KNN demonstrate commendable overall accuracy in Dataset 1, showcasing their
effectiveness in accurately categorizing the majority of samples. This performance accentuates the robustness and reliability of
these algorithms in handling the classification task within Dataset 1, indicating their potential for successful application in similar
cases.

5. Comparative Analysis
Table 6 presents a comparison between our proposed method and several recent studies that also used the CWRU dataset for
bearing fault diagnosis. This comparison highlights the strengths and limitations of different approaches, focusing on classification
accuracy, computational efficiency, and practical applicability. The studies reviewed include deep learning models, hybrid machine
learning techniques, and feature-based approaches, each offering unique benefits and challenges.
Deep learning methods, such as Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs), and
Transformers, have achieved high classification accuracy by extracting complex patterns from vibration signals. For example, Liu
et al. [37] developed a hybrid deep learning model combining 1D-LBP and Gray Relational Analysis (GRA), reaching 99.5% accuracy.
Similarly, Zhao et al. [38] introduced a Bayesian Variational Transformer, which improved the model's ability to generalize across
different conditions but required high computational power. Zhang et al. [39] used GANs to enhance classification performance,
achieving 99.3% accuracy, but the approach was computationally intensive and difficult to train.
Traditional machine learning and hybrid models remain popular due to their interpretability and lower computational cost. Raj
et al. [19] focused on raw accelerometer data processing, using deep learning models to achieve 98.9% accuracy, while reducing the
need for complex feature engineering. Das et al. [40] developed a hybrid approach combining SVM, kNN, and ANN, achieving 98.7%
accuracy. Although their accuracy was slightly lower than deep learning models, these methods offer faster computation and are
easier to implement in real-time industrial systems.
Our proposed approach, which integrates time-domain and frequency-domain features with SVM and kNN classifiers, achieved
an accuracy of 98.9%, comparable to the best-performing methods. However, our approach has a key advantage: it maintains a lower
computational footprint, making it more practical for real-time condition monitoring. Unlike deep learning models, which often
function as "black boxes", our method is more interpretable, as it relies on well-defined statistical features such as RMS, kurtosis,
and peak-to-peak values. This makes it a more suitable choice for industrial applications, where explain ability and real-time
processing are essential.

Fig. 16. SVM confusion matrix of dataset 1. Fig. 17. KNN confusion matrix dataset 1.

Fig. 18. Inner race with depth of fault 0.007 Fig. 19. Inner race with depth of fault 0.014 Fig. 20. Inner race with depth of fault 0.021
inches. inches. inches.

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 Application of machine learning techniques for bearing fault diagnosis 11

Table 6. Comparison between our proposed method and several recent studies.
Study Feature Extraction Method Classifier(s) Accuracy Key Advantages Limitation
Texture Analysis Requires image transformation;
Multiple ML Captures fine texture features;
Kaplan et al. [36] with Local Binary Patterns 100 Computationally intensive for real-
Models High classification accuracy
(LBP) time
Effective hybrid approach;
Hybrid Deep CNN-Based Requires deep learning
Liu et al. [37] 99,5 Strong spatial and frequency
Learning (1D-LBP + GRA) Model infrastructure
feature learning
Raw Accelerometer Direct signal input; Data-intensive; Needs large
Raj et al. [19] Deep Learning 98,9
Data Processing Reduces preprocessing effort datasets
Bayesian Variational Transformer- Strong generalization; Handles
Zhao et al. [38] 99,2 High computational complexity
Transformer Based Model data imbalance
Generative Adversarial GAN-based Improves feature representation; Effective Training instability; High resource
Zhang et al. [39] Model 99,3
Networks (GAN) in low-data scenarios consumption
Hybrid Machine Learning Combines strengths of multiple classifiers;
Das et al. [40] SVM, kNN, ANN 98,7 Increased model complexity
(SVM, kNN, ANN) Good generalization
Time-Domain &
Proposed Computationally efficient; Slightly lower accuracy than deep
Frequency-Domain SVM, kNN 98,9
method Suitable for real-time deployment learning models
Features

Furthermore, the use of feature selection techniques optimizes classification accuracy while reducing computational
complexity, making our method a strong alternative to deep learning approaches that require extensive computing resources.
In summary, while deep learning models tend to achieve higher classification accuracy, they demand significant computational
power and large datasets, which limits their real-time applicability. Our proposed method provides a balanced trade-off between
accuracy, efficiency, and interpretability, making it a scalable and practical solution for real-time bearing fault diagnosis and
predictive maintenance in industrial settings.

6. Conclusion
This study explored the use of machine learning algorithms to detect bearing faults, focusing on the Case Western Reserve
University (CWRU) dataset. The methodology included systematic steps of data preprocessing, feature extraction, and model
development, with the aim of improving fault detection accuracy and efficiency. By employing advanced machine learning
techniques, such as Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, and Naive Bayes, this study achieved
high classification accuracies, with SVM achieving a remarkable 99.2% accuracy for inner-race fault detection. The primary
contribution of this study lies in the integration of feature selection and robust classification techniques to create an efficient and
accurate framework for bearing fault diagnosis. This study highlights the effectiveness of combining time-domain and frequency-
domain features with classical machine learning algorithms. Additionally, a comparative analysis with the existing literature
demonstrates the advantages of the proposed approach, particularly its simplicity, computational efficiency, and applicability to
industrial settings. However, this study has certain limitations. This study used a single dataset (CWRU), which may limit the
generalizability of the findings to other industrial environments with different operating conditions or machinery configurations.
Future studies could address the limitations identified in this study through several avenues. Expanding the analysis to include
datasets from diverse industrial environments would improve the generalizability of the findings and ensure their applicability
across a wider range of machinery configurations and operational conditions. Additionally, exploring advanced parameter
optimization techniques for the employed machine-learning algorithms could enhance their performance and adaptability.
Another critical aspect to investigate is the impact of different signal segmentation strategies on the diagnostic accuracy. Finally,
the incorporation of real-time condition monitoring systems into the proposed methodology provides valuable insights into its
practical implementation and effectiveness in dynamic industrial settings. In conclusion, this study makes a substantial
contribution to the field of intelligent fault diagnosis by offering a robust, efficient, and interpretable framework for bearing fault
detection. By addressing these limitations and pursuing suggested future directions, the applicability and scalability of these
techniques can be significantly enhanced, ultimately paving the way for more reliable and effective predictive maintenance
systems.

Author Contributions
S. Eddai conceptualized the study, developed the methodology, and wrote the original draft; N. Feki supervised the project,
validated the results, and contributed to reviewing and editing the manuscript; A. Ghorbel conducted the investigation and
performed data analysis; A. El Hami provided resources and contributed to validation; M. Haddar supervised the project
administration and provided overall guidance. The manuscript was written through the contribution of all authors. All authors
discussed the results, reviewed, and approved the final version of the manuscript.

Acknowledgments
Not applicable.

Conflict of Interest
The authors declared no potential conflicts of interest concerning the research, authorship, and publication of this article.

Funding
The authors gratefully acknowledge the funding of Project (PEJC2023-D6P20) financed by the Tunisian Ministry of Higher
Education and Scientific Research.

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 12 Sarra Eddai et al., Vol. x, No. x, 2025

Data Availability Statements

The datasets used in this study were obtained from the Case Western Reserve University (CWRU) Bearing Data Center and are
available from the corresponding author upon reasonable request. These data are widely used in bearing fault diagnosis research,
allowing for reliable comparative analysis.

Nomenclature
fs Sampling frequency [Hz] Kt Number of signal slices with overlap
frShaft Shaft rotation frequency [Hz] EB1 Energy in frequency band (0 to 1000 Hz) [dB]
Tcage Cage period [s] EB2 Energy in frequency band (1000 to 2000 Hz) [dB]
T Sampling period [s] EB3 Energy in frequency band (2000 to 3000 Hz) [dB]
NP Number of points per period [point] EB4 Energy in frequency band (3000 to 4000 Hz) [dB]
NT Length required for each slice [point] EB5 Energy in frequency band (4000 to 5000 Hz) [dB]
NM Sample length for processing [samples] EBT Energy in frequency band (0 to 6000 Hz) [dB]
S1 Number of samples per overlap [samples]

References
[1] Fourati, A., Bourdon, A., Feki, N., Rémond, D., Chaari, F., Haddar, M., Angular-based modeling of induction motors for monitoring, Journal of Sound and
Vibration, 395, 2017, 371-392.
[2] Fourati, A., Feki, N., Bourdon, A., Rémond, D., Chaari, F., Haddar, M., Electrical modeling for fault detection based on motor current signal analysis
and angular approach, In Advances in Condition Monitoring of Machinery in Non-Stationary Operations: Proceedings of the Fourth International Conference on
Condition Monitoring of Machinery in Non-Stationary Operations, CMMNO'2014, Lyon, France December 15-17 (pp. 15-25), Springer International Publishing,
2016.
[3] Dolenc, B., Boškoski, P., Juricic, D., Distributed bearing fault diagnosis based on vibration analysis, Mechanical Systems and Signal Processing, 66, 2016,
521-532.
[4] Boudiaf, A., Moussaoui, A., Dahane, A., Atoui, I., A comparative study of various methods of bearing fault diagnosis using the Case Western Reserve
University data, Journal of Failure Analysis and Prevention, 16(2), 2016, 271-284.
[5] Song, L., Wang, H., Chen, P., Step-by-step fuzzy diagnosis method for equipment based on symptom extraction and trivalent logic fuzzy diagnosis
theory, IEEE Transactions on Fuzzy Systems, 26(6), 2018, 3467-3478.
[6] Song, L., Wang, H., Chen, P., Vibration-based intelligent fault diagnosis for roller bearings in low-speed rotating machinery, IEEE Transactions on
Instrumentation and Measurement, 67(8), 2018, 1887-1899.
[7] Kankar, P.K., Sharma, S.C., Harsha, S.P., Fault diagnosis of ball bearings using machine learning methods, Expert Systems with Applications, 38(3), 2011,
1876-1886.
[8] Di Maggio, L.G., Intelligent fault diagnosis of industrial bearings using transfer learning and CNNs pre-trained for audio classification, Sensors, 23(1),
2022, 211.
[9] Lei, Y., Jia, F., Lin, J., Xing, S., Ding, S.X., An intelligent fault diagnosis method using unsupervised feature learning towards mechanical big data, IEEE
Transactions on Industrial Electronics, 63(5), 2016, 3137-3147.
[10] Jing, L., Zhao, M., Li, P., Xu, X., A convolutional neural network-based feature learning and fault diagnosis method for the condition monitoring of
gearbox, Measurement, 111, 2017, 1-10.
[11] Zhang, X., Jiang, D., Han, T., Wang, N., Yang, W., Yang, Y., Rotating machinery fault diagnosis for imbalanced data based on fast clustering algorithm
and support vector machine, Journal of Sensors, 2017, 2017, 8092691.
[12] Das, D.B., Birant, D., Machine learning for fault analysis in rotating machinery: A comprehensive review, Heliyon, 9(6), 2023, e17584.
[13] Liu, R., Yang, B., Zio, E., Chen, X., Artificial intelligence for fault diagnosis of rotating machinery: A review, Mechanical Systems and Signal Processing,
108, 2018, 33-47.
[14] Adams, K., Machine learning algorithms for rotating machinery bearing fault diagnostics, Materials Today Proceedings, 44(6), 2021, 4931-4933.
[15] Uma Pavan Kumar, K., Gandhi, O., Venkata Reddy, M., Srinivasu, S.V.N., Usage of KNN, Decision Tree and Random Forest Algorithms in Machine
Learning and Performance Analysis with a Comparative Measure, In Machine Intelligence and Soft Computing: Proceedings of ICMISC 2020 (pp. 473-479).
Springer Singapore, 2021.
[16] Shi, L., Su, S., Wang, W., Gao, S., Chu, C., Bearing fault diagnosis method based on deep learning and health state division, Applied Sciences, 13, 2023,
7424.
[17] Saad, A., Usman, A., Arif, S., Liwicki, M., Almqvist, A., Bearing fault detection scheme using machine learning for condition monitoring applications,
In International Conference on Mechanical, Automotive and Mechatronics Engineering (ICMAME 2023), 29-30 April, Dubai, UAE, 2023.
[18] Zhang, S., Zhang, S., Wang, B., Habetler, T.G., Deep Learning Algorithms for Bearing Fault Diagnostics—A Comprehensive Review, IEEE Access, 8,
2020, 29857-29881.
[19] Raj, K.K., Kumar, S., Kumar, R.R., Andriollo, M., Enhanced Fault Detection in Bearings Using Machine Learning and Raw Accelerometer Data: A Case
Study Using the Case Western Reserve University Dataset, Information, 15(5), 2024, 259.
[20] Neupane, D., Seok, J., Bearing fault detection and diagnosis using case western reserve university dataset with deep learning approaches: A review,
IEEE Access, 8, 2020, 93155-93178.
[21] Lei, Y., Lin, J., He, Z., Zuo, M.J., A review on empirical mode decomposition in fault diagnosis of rotating machinery, Mechanical Systems and Signal
Processing, 35(1), 2013, 108-126.
[22] Lei, Y., Zuo, M.J., He, Z., Zi, Y., A multidimensional hybrid intelligent method for gear fault diagnosis, Expert Systems with Applications, 37(2), 2010,
1419-1430.
[23] Peng, Z., Peter, W.T., Chu, F., A comparison study of improved Hilbert–Huang transform and wavelet transform: Application to fault diagnosis for
rolling bearing, Mechanical Systems and Signal Processing, 19(5), 2005, 974-988.
[24] Case Western Reserve University Bearing Data Center Website. Retrieved from http://csegroups.case.edu/bearingdatacenter/home.
[25] Neupane, D., Seok, J., Bearing fault detection and diagnosis using case western reserve university dataset with deep learning approaches: A review,
IEEE Access, 8, 2020, 93155-93178.
[26] Zhang, X., Zhao, B., Lin, Y., Machine learning based bearing fault diagnosis using the case western reserve university data: A review, IEEE Access, 9,
2021, 155598-155608.
[27] Khlaief, A., Nguyen, K., Medjaher, K., Picot, A., Maussion, P., Tobon, D., Cheron, R., Feature engineering for ball bearing combined-fault detection
and diagnostic, In 2019 IEEE 12th International Symposium on Diagnostics for Electrical Machines, Power Electronics and Drives (SDEMPED) (pp. 384-390), IEEE,
2019.
[28] Jang, J.G., Noh, C.M., Kim, S.S., Shin, S.C., Lee, S.S., Lee, J.C., Vibration data feature extraction and deep learning-based preprocessing method for
highly accurate motor fault diagnosis, Journal of Computational Design and Engineering, 10(1), 2023, 204-220.
[29] Brito, L.C., Susto, G.A., Brito, J.N., Duarte, M.A.V., Fault detection of bearing: An unsupervised machine learning approach exploiting feature
extraction and dimensionality reduction, Informatics, 8, 2021, 85.
[30] Rai, A., Upadhyay, S.H., A review on signal processing techniques utilized in the fault diagnosis of rolling element bearings, Tribology International,
96, 2016, 289-306.
[31] Harrouche, F., Comparative study of two optimization methods for a mechanical fault classification system using fuzzy logic, Doctoral Dissertation, 2019.
[32] Lakis, A.A., Rotating machinery fault diagnosis using time-frequency methods, In Proceedings of the 7th WSEAS International Conference on Electric

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13
 Application of machine learning techniques for bearing fault diagnosis 13

Power Systems, High Voltages, Electric Machines, Athens: WSEAS Press, 2007.
[33] Kaliappan, J., Bagepalli, A.R., Almal, S., Mishra, R., Hu, Y.C., Srinivasan, K., Impact of cross-validation on machine learning models for early detection
of intrauterine fetal demise, Diagnostics, 13(10), 2023, 1692.
[34] Akhtar, T., Gilani, S.O., Mushtaq, Z., Arif, S., Jamil, M., Ayaz, Y., ..., Waris, A., Effective voting ensemble of homogeneous ensembling with multiple
attribute-selection approaches for improved identification of thyroid disorder, Electronics, 10(23), 2021, 3026.
[35] Atef, M.M., Khair-Eldeen, W., Yan, J., Nassef, M.G., Investigating the combined effect of multiple dent and bump faults on the vibrational behavior
of ball bearings, Machines, 10(11), 2022, 1062.
[36] Kaplan, K., Kaya, Y., Kuncan, M., Minaz, M.R., Ertunç, H.M., An improved feature extraction method using texture analysis with LBP for bearing fault
diagnosis, Applied Soft Computing, 87, 2020, 106019.
[37] Liu, H., Wang, H., Tang, J., Hybrid deep learning methods for bearing fault diagnosis combining 1D-LBP and GRA, IEEE Access, 10, 2022, 120455-
120465.
[38] Zhao, M., Wang, W., Wang, X., A Bayesian variational transformer for generalizable rotating machinery fault diagnosis, Mechanical Systems and Signal
Processing, 193, 2023, 108459.
[39] Zhang, Z., Wang, C., Guo, X., Generative adversarial networks for bearing fault diagnosis under various working conditions, Mechanical Systems and
Signal Processing, 154, 2021, 107556.
[40] Das, O., Das, D.B., Birant, D., A comprehensive review on hybrid machine learning techniques for fault diagnosis in rotating machinery, Applied Soft
Computing, 85, 2019, 106019.

ORCID iD
Sarra Eddai https://orcid.org/0009-0008-3016-2752
Nabih Feki https://orcid.org/0000-0002-2780-4637
Ahmed Ghorbel https://orcid.org/0000-0002-0591-1968
Abdelkhalak El Hami https://orcid.org/0000-0001-8080-7952

© 2025 Shahid Chamran University of Ahvaz, Ahvaz, Iran. This article is an open access article distributed under
the terms and conditions of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0
license) (http://creativecommons.org/licenses/by-nc/4.0/).

How to cite this article: Eddai S., et al. Application of machine learning techniques for bearing fault diagnosis, J. Appl. Comput.
Mech., xx(x), 2025, 1–13. https://doi.org/10.22055/jacm.2025.48052.4943

Publisher’s Note Shahid Chamran University of Ahvaz remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Journal of Applied and Computational Mechanics, Vol. xx, No. x, (2025), 1-13