Saddam Bensaoucha
| Gharib Mousa Gharib | Maha Al Soudi | Ali Teta | Samia Benzita | Ahmed Saadeddine Bertal | Ahlam Guiatni | Abdelaziz Rabehi | Mohamed Benghanem*
© 2025 The authors. This article is published by IIETA and is licensed under the CC BY 4.0 license (http://creativecommons.org/licenses/by/4.0/).
OPEN ACCESS
This study investigates bearing fault diagnosis using the Case Western Reserve University (CWRU) dataset through a systematic approach that includes vibration signal acquisition, feature extraction, and dataset preparation. To evaluate model performance, several predefined Artificial Neural Network (ANN) architectures were trained and compared alongside an automatically optimized neural network (ONN). The predefined models provide a base- line for performance assessment, while the optimized model leverages Bayesian optimization and cross-validation to enhance generalization and reduce overfitting. Experimental results demonstrate that the ONN achieves the highest testing accuracy of 98.6% and an F1-score of 0.99, with the lowest misclassification cost, outperforming all predefined architectures, which achieved accuracies ranging from 92% to 95%. The finding of this study highlights the significance of network architecture selection and hyperparameter optimization in improving ANN-based fault detection. The proposed methodology provides a robust and efficient framework for automated bearing fault diagnosis, offering promising prospects for predictive maintenance and reliability enhancement in rotating machinery. Overall, the study confirms that combining systematic feature extraction with ONN design significantly improves diagnostic accuracy and operational reliability in industrial applications.
rotating machines, bearing faults, fault detection, artificial intelligence, neural networks
Rotating machines, such as motors, turbines, pumps, and compressors, play a vital role in modern industry by enabling energy conversion and fluid transfer processes. Their reliable operation is critical to the efficiency and safety of industrial systems. However, these machines are subject to various electrical and mechanical faults, including misalignment, unbalance, and bearing defects, which can lead to reduced performance, unexpected shutdowns, and costly maintenance interventions. Among these, bearing faults are the most common, accounting for approximately 40-45% of rotating machine failures, followed by stator (30-37%) and rotor faults (about 10%) [1-4].
Bearings are essential mechanical elements designed to support loads while minimizing friction and wear through rolling elements such as balls or rollers. Despite their importance, bearing defects including outer race, inner race, ball, and cage faults remain a major source of vibration, noise, overheating, and mechanical instability. If undetected, these issues can escalate to severe damage, system downtime, and significant economic losses. Consequently, the timely detection and diagnosis of bearing faults have become a cornerstone of predictive maintenance strategies [4-9].
Traditional approaches to fault detection, such as vibration analysis, thermography, and ultrasonic testing, provide valuable diagnostic information but often rely on human interpretation and predefined fault signatures [10-14]. These limitations have motivated the growing adoption of data-driven methods, particularly machine learning algorithms, for automated fault detection and classification. Neural networks, in particular, have shown strong potential in modeling nonlinear patterns in vibration signals and improving diagnostic accuracy [15-23].
This study investigates the application of Multilayer Perceptron (MLP) neural networks for the detection and classification of bearing faults in rotating machinery. Vibration signals are analyzed, and statistical features such as maximum, minimum, and median values are extracted as network inputs. Several architectures are evaluated using benchmark datasets, including the Case Western Reserve University (CWRU) dataset.
The contributions of this paper are threefold:
•To highlight the significance of bearing fault detection in industrial rotating machinery.
•To evaluate the effectiveness of MLP neural networks for fault classification using vibration features.
•To propose an automated framework that improves accuracy and reliability in predictive maintenance.
The remainder of this paper is organized as follows: Section 2 briefly presents the architecture of neural networks; Section 3 describes the CWRU dataset; Section 4 introduces the application of neural networks for bearing fault detection; Section 5 discusses the classification results; and Section 6 concludes the study.
Artificial Neural Networks (ANNs), illustrated in Figure 1, are computational models inspired by the structure and functioning of biological neurons in the human brain. They are composed of interconnected layers of processing units (neurons) organized into an input layer, one or more hidden layers, and an output layer [10, 22-26].
Figure 1. Schematic diagram of the neural network
In the context of classification tasks, the input layer receives raw data features, which are then processed through hidden layers using weighted connections and nonlinear activation functions. These transformations enable the network to capture complex relationships and extract relevant patterns from the data. The output layer generates the final decision, typically represented as a probability distribution over the possible classes, from which the most likely class is selected.
Due to their capacity to model nonlinear relationships and handle high-dimensional data, ANNs are particularly effective for pattern recognition, classification, and prediction problems. In this study, they are of significant importance for analyzing complex datasets and enhancing the reliability of decision-making processes.
In the context of bearing fault diagnosis, ANNs can learn the complex nonlinear relationships between the measured signals (such as vibration or current) and the fault types. Once trained, the ANN can classify new unseen data with high accuracy [27-30].
The dataset utilized in this study is sourced from the CWRU bearing data center, which is widely regarded as one of the most comprehensive and frequently cited benchmarks in the field of bearing fault diagnosis [27, 28]. The experiments were carried out using a 2-horsepower reliance electric motor, in which faults were artificially induced using electro-discharge machining as shown in Figure 2.
Figure 2. Experimental setup of the CWRU bearing fault test stand
The dataset provides vibration signal data collected under both healthy and faulty bearing conditions, encompassing four distinct states: healthy or faulty bearings as illustrated in Figure 3.
Figure 3. Cross-sectional diagrams of a radial ball bearing under different conditions
Figure 4. Vibration signals for healthy and faulty bearings under 3 hp load condition. Healthy (green), ORF (red), IRF (orange), BF (magenta)
The dataset includes fault sizes ranging from 0.007 to 0.028 inches. Vibration signals were captured using accelerometers mounted both near and far from the drive-end bearing. Measurements were taken under varying motor loads (from 0 to 3 horsepower) and speeds ranging from 1720 to 1797 RPM. Figure 4 illustrates the vibration signals acquired under a 3 hp load condition for different bearing states. It can be observed that the signal amplitude differs noticeably among the cases.
These acceleration signals serve as the primary input for training and validating the machine learning models developed in this study, particularly the MLP, to facilitate accurate detection and classification of bearing fault types [31-34].
Table 1 presents the parameters of the CWRU bearing dataset, including fault types, fault sizes, load conditions, rotational speeds, and the corresponding data format.
Table 1. CWRU bearing dataset parameters
|
Fault Type |
Fault Size (in) |
Load (hp) |
Speed (RPM) |
Data Format |
|
Healthy |
0 |
0, 1, 2, 3 |
1720–1797 |
.mat |
|
ORF |
0.007, 0.014, 0.021 |
0, 1, 2, 3 |
1720–1797 |
.mat |
|
IRF |
0.007, 0.014, 0.021, 0.028 |
0, 1, 2, 3 |
1720–1797 |
.mat |
|
BF |
0.007, 0.014, 0.021, 0.028 |
0, 1, 2, 3 |
1720–1797 |
.mat |
To assess model performance, five predefined feedforward ANN architectures of varying depth and neuron count were trained alongside one automatically ONN. This setup allowed for a comparison of standard network configurations with an optimized architecture leveraging hyperparameter tuning for improved accuracy and generalization.
Figure 5 outlines the bearing fault diagnosis pipeline, from raw vibration signal acquisition to final classification.
Figure 5. Workflow of the bearing fault diagnosis methodology using neural networks
4.1 Data pre-processing and feature extraction
In this study, vibration signals were extracted from the CWRU dataset under four operating conditions: healthy bearing, BF, IRF, and ORF. For each condition, the full-length signal was segmented into 10 equal-length sub signals to ensure a balanced and diverse dataset.
Figure 6 shows the segmented vibration signals for the healthy bearing and the three faulty conditions (ORF, IRF, BF) under a 3 hp load. The signals are divided into ten subsignals.
Figure 6. Segmented vibration signals for healthy and faulty bearings under a 3 hp load condition. Healthy (green), ORF (red), IRF (orange), and BF (magenta)
Each signal has a duration of 10 s and was sampled at a frequency of 12 kHz, resulting in 120000 data points per signal. The signals are divided into 10 sub signals.
This segmentation process not only highlights the amplitude differences among bearing conditions but also serves to augment the dataset, thereby providing a larger and more representative input set for training the neural network.
Table 2 summarizes the structure of the final processed dataset used for training the neural network models. For each bearing class, the table shows the number of original signals, their composition in terms of sizes and loads, the total number of sub signals obtained after segmentation, the number of features extracted per sub signal, and the resulting total feature samples. As shown, segmentation and feature extraction expand the dataset to 4800 samples, ensuring sufficient representation of all bearing conditions for accurate ANN training and evaluation.
Table 2. Summary of common ANN models and their use cases
|
Class |
No. of Original Signals |
Comp. |
No. of Sub-signs |
Features Per Sub-signal |
Total Feature Samples |
|
Healthy |
4 |
4 loads |
10 × 4 = 40 |
10 |
400 |
|
ORF |
12 |
3 sizes × 4 loads |
10 × 12 = 120 |
10 |
1200 |
|
IRF |
16 |
4 sizes × 4 loads |
10 × 16 = 160 |
10 |
1600 |
|
BF |
16 |
4 sizes × 4 loads |
10 × 16 = 160 |
10 |
1600 |
|
Total |
48 |
— |
480 |
10 |
4800 |
The extracted statistical features and their corresponding mathematical expressions are summarized in Table 3. All ten of these features were utilized for training the classifiers without applying any further feature selection. Prior to training, each feature was normalized to a [0, 1] range using min-max scaling to ensure equal contribution during the learning process (Appendix A).
Table 3. Extracted statistical features and their mathematical expressions
|
Feature Name |
Symbol |
Mathematical Expression |
|
Minimum |
min(x) |
Min(x1, x2, …, xN) |
|
Maximum |
max(x) |
Max(x1, x2, …, xN) |
|
Mean |
µ |
µ = (1/N) Σ xi |
|
Median |
Med |
Middle value of ordered set |
|
Root Mean Square |
RMS |
RMS = sqrt((1/N) Σ xi²) |
|
Standard Deviation |
Σ |
σ = sqrt((1/N) Σ (xi – µ)²) |
|
Kurtosis |
Kur |
kur = (1/N) Σ ((xi – µ)/σ)⁴ |
|
Peak Value |
Peak |
Max(|xi|) |
|
Crest Factor |
CF |
CF = Peak / RMS |
|
Impulse Factor |
IF |
IF = Peak / [(1/N) Σ |xi|] |
4.2 Training and validation
The dataset was randomly divided into two subsets: 70% for training and 30% for testing. To enhance generalization and prevent overfitting, a 5-fold cross-validation strategy was applied during training [35-40].
4.3 Neural network models
Six ANN architectures were evaluated for bearing fault classification, including five predefined feedforward networks of varying depth and neuron count, along with one optimized network model. The optimized network was automatically tuned to achieve an optimal balance between classification accuracy and computational complexity.
Hyperparameters are predefined configuration parameters that govern the learning process of a neural network and are set prior to training. They determine the structure and behavior of the model, including the number of hidden layers, the number of neurons per layer, and the type of activation function (ReLU, sigmoid, or tanh). All models, including the optimized network, employed cross-validation to improve generalization and ensure robust performance (Appendix B). The optimized network further leverages automatic hyperparameter tuning, utilizing techniques such as Bayesian optimization to enhance predictive reliability (Appendix C).
The evaluated models include the Narrow Neural Network (NNN), Medium Neural Network (MNN), Wide Neural Network (WNN), Bilayered Neural Network (BNN), Trilayered Neural Network (TNN), and the Optimizable Neural Network (ONN). All networks were trained using the backpropagation algorithm.
Table 4 presents a summary of common ANN models, detailing their structural characteristics and typical use cases.
Table 4. Summary of common ANN models and their use cases
|
Model |
Definition |
Use Case |
|
NNN |
Few neurons per hidden layer |
Simple problems or limited data |
|
MNN |
Moderate number of neurons per layer, between narrow and wide |
General applications balancing performance and cost |
|
WNN |
Large number of neurons per hidden layer |
Complex problems with large datasets |
|
BNN |
Two hidden layers |
Moderate complexity tasks requiring additional depth |
|
TNN |
Three hidden layers |
Complex problems such as signal processing or image recognition |
|
ONN |
Architecture or hyperparameters optimized automatically (Bayesian optimization) |
Tasks requiring high accuracy and adaptive tuning |
4.4 Performance evaluation metrics
To comprehensively assess the performance of classification models, several standard metrics are considered, including accuracy, recall (Rec), Precision (Pre), and F1-score. These metrics, computed from the confusion matrix components—True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN)—provide complementary insights into model performance. Using them together ensures a balanced evaluation, capturing both the model’s overall correctness and its ability to correctly identify positive and negative instances, which is particularly important in cases of imbalanced datasets. Eqs. (1)-(4) present the formulas for accuracy, precision, recall, and the F1-score, respectively [26, 41-43].
$Accuracy =\frac{T P+T N}{T P+T N+F P+F N}$ (1)
$Recall=\frac{T P}{T P+F N}$ (2)
$Precision=\frac{T P}{T P+F P}$ (3)
$F 1- Score =2 \times \frac{{ Precision }+ { Recall }}{ { Precision }+ { Recall }}$ (4)
In addition to the aforementioned metrics, the total misclassification cost is reported in the results. This cost was calculated using a uniform cost matrix where any prediction error is assigned a unit penalty. Specifically, for our four-class problem, a cost of 0 was applied for correct classifications and a cost of 1 was applied for any misclassification. Consequently, the reported 'Cost' is equivalent to the total number of misclassified samples in the dataset.
The performance of the six ANN architectures (NNN, MNN, WNN, BNN, TNN, and ONN) was assessed during training and testing. Table 5 summarizes classification accuracy and total misclassification cost, while Table 6 reports recall, precision, and F1-score. The hyperparameter settings of each model are provided in Table 7.
Table 5. Classification accuracy and total cost of ANN models
|
Model |
Training Phase |
Testing Phase |
||
|
Accuracy |
Cost |
Accuracy |
Cost |
|
|
NNN |
94.3% |
19 |
96.5% |
5 |
|
MNN |
95.8% |
14 |
94.4% |
8 |
|
WNN |
94.3% |
19 |
95.8% |
6 |
|
BNN |
95.8% |
14 |
96.5% |
5 |
|
TNN |
93.8% |
21 |
97.9% |
3 |
|
ONN |
97.6% |
8 |
98.6% |
2 |
Table 6. Performance evaluation of ANN models based on recall, precision, and F1-score
|
Model |
Training Phase |
Testing Phase |
||||
|
Rec |
Pre |
F1-Score |
Rec |
Pre |
F1-Score |
|
|
NNN |
0.95 |
0.95 |
0.95 |
0.99 |
0.98 |
0.97 |
|
MNN |
0.95 |
0.96 |
0.96 |
0.94 |
0.96 |
0.95 |
|
WNN |
0.94 |
0.93 |
0.94 |
0.96 |
0.97 |
0.96 |
|
BNN |
0.95 |
0.96 |
0.96 |
0.97 |
0.96 |
0.96 |
|
TNN |
0.94 |
0.92 |
0.93 |
0.98 |
0.99 |
0.98 |
|
ONN |
0.97 |
0.98 |
0.97 |
0.99 |
0.99 |
0.99 |
Table 7. Hyperparameter settings of ANN models
|
Model |
No. of Layers |
First Layer |
Second Layer |
Third Layer |
Activation Function |
|
NNN |
1 |
10 |
– |
– |
ReLU |
|
MNN |
1 |
25 |
– |
– |
ReLU |
|
WNN |
1 |
100 |
– |
– |
ReLU |
|
BNN |
2 |
10 |
10 |
– |
ReLU |
|
TNN |
3 |
10 |
10 |
10 |
ReLU |
|
ONN |
1 |
12 |
– |
– |
Sigmoid |
These results form the basis for the following discussion.
5.1 Discussion of results
The results indicate that the ONN consistently outperformed all other models, achieving the highest testing accuracy (98.6%), an F1-score of 0.99, and the lowest misclassification cost (2). This confirms the advantage of automatic hyperparameter tuning, such as Bayesian optimization and cross-validation, in enhancing generalization and predictive reliability.
The TNN provided the next best performance (97.9% accuracy, F1-score 0.98), but with greater computational cost, highlighting those deeper architectures improve accuracy at the expense of complex ity. The MNN and WNN achieved moderate results, reflecting a balance between simplicity and performance, while NNN and BNN remained less effective compared with deeper or optimized architectures.
Most models obtained recall, precision, and F1-scores around 0.95 in training, suggesting stable learning. However, ONNs consistently superior testing performance demonstrates stronger robustness and reduced overfitting.
In summary, the ONN emerges as the most reliable architecture for bearing fault diagnosis, combining high accuracy, low cost, and optimized parameters. These findings emphasize the importance of architectural and hyperparameter optimization in ANN-based fault detection.
It is worth emphasizing that, unlike many previous studies that selectively use data subsets to artificially boost classification accuracy, this study employed the entire CWRU dataset without any selective filtering. This comprehensive approach provides a more realistic evaluation of the models’ performance across all fault types and operating conditions, thereby enhancing the validity and generalizability of our findings.
At the same time, it should be acknowledged that this study was conducted solely using the CWRU dataset. While the results are encouraging, the generalizability of the models to other bearing types or to real-world industrial environments with different noise levels and operational conditions remains to be validated. Future work will focus on evaluating these architectures on additional datasets to assess their robustness and practical applicability.
This study presented a systematic approach for bearing fault diagnosis using vibration signals from the CWRU dataset. Several predefined ANN architectures were evaluated alongside an automatically optimized neural network (ONN) to assess the impact of network design and hyperparameter tuning on classification performance.
The experimental results demonstrated that the ONN consistently outperformed all predefined models, achieving the highest testing accuracy of 98.6%, an F1-score of 0.99, and the lowest misclassification cost. These findings confirm the synergistic effect of combining systematic feature extraction with automatic hyperparameter optimization. The statistical features provide a compact, discriminative representation of the raw vibration signals, which simplifies the learning task. Bayesian optimization then tailors the network architecture specifically to this feature set, finding the optimal depth, width, and training parameters to maximize information extraction and minimize overfitting. This synergy is evident in the ONN's superior generalization and robustness compared to models where architecture and features were not co-optimized.
Deeper architectures, such as the TNN, also achieved high accuracy but incurred greater computational costs, highlighting the trade-off between complexity and performance. Predefined models with moderate depth or width yielded reasonable results, yet they were surpassed by the ONN in both accuracy and robustness.
Overall, this work emphasizes the critical role of integrating network architecture selection and hyperparameter tuning with feature engineering for ANN-based fault diagnosis. The proposed methodology provides a robust and efficient framework for automated bearing fault detection, supporting predictive maintenance strategies and enhancing operational reliability in rotating machinery. These results confirm that ONN design, when driven by informative features, offers significant and reliable improvements for industrial diagnostic applications.
The researchers wish to extend their sincere gratitude to the Deanship of Scientific Research at the Islamic University of Madinah (KSA) for the support provided to the Post-Publishing Program.
The authors also acknowledge Zarqa University (Jordan) for their support and contributions to this research.
|
CWRU |
Case Western Reserve University |
|
MLP |
Multilayer Perceptron |
|
ANN |
Artificial Neural Network |
|
IRF |
Inner Race Fault |
|
ORF |
Outer Race Fault |
|
BF |
Ball Fault |
|
NNN |
Narrow Neural Network |
|
MNN |
Medium Neural Network |
|
WNN |
Wide Neural Network |
|
BNN |
Bilayered Neural Network |
|
TNN |
Trilayered Neural Network |
|
ONN |
Optimizable Neural Network |
|
TP |
True Positive |
|
TN |
True Negative |
|
FP |
False Positive |
|
FN |
False Negative |
Appendix A. Feature extraction code and data splitting
The following MATLAB code was used to extract the ten statistical features from each signal segment. The same logic was repeated for every bearing condition in the dataset.
|
% --- Load Data and Initialize Parameters --- load('97.mat'); % Load the specific data file signal = X097_DE_time(1:120000,:); N = length(signal); num_segments = 10; segment_length = floor(N / num_segments);
% Preallocate results matrix results = zeros(num_segments, 10);
% --- Calculate Features for Each Segment --- for i = 1:num_segments % Define segment indices start_idx = (i-1) * segment_length + 1; if i == num_segments end_idx = N; % Ensure last segment captures all remaining points else end_idx = i * segment_length; end
% Extract the signal segment sub_signal = signal(start_idx:end_idx);
% Calculate the 10 statistical features min_val = min(sub_signal); max_val = max(sub_signal); mean_val = mean(sub_signal); median_val = median(sub_signal); % Using built-in function for robustness rms_val = sqrt(mean(sub_signal.^2)); std_val = std(sub_signal); kurtosis_val = kurtosis(sub_signal); peak_val = max(abs(sub_signal)); crest_factor = peak_val / rms_val; impulse_factor = peak_val / mean(abs(sub_signal));
% Store results for the segment results(i, :) = [min_val, max_val, mean_val, median_val, rms_val, ... std_val, kurtosis_val, peak_val, crest_factor, impulse_factor]; end
% --- Save Features to File --- feature_names = {'Min', 'Max', 'Mean', 'Median', 'RMS', 'StdDev', ... 'Kurtosis', 'Peak', 'CrestFactor', 'ImpulseFactor'}; T = array2table(results, 'VariableNames', feature_names); writetable(T, 'features_DE.xlsx'); disp('Features successfully saved to "features_DE.xlsx"'); |
Appendix B.
This appendix contains the core MATLAB code template used to train and evaluate all ANN models presented in the paper. The specific hyperparameters for each model (NNN, MNN, WNN, BNN, TNN, ONN)—including the number of layers, neurons per layer, and activation function—were varied according to the configurations specified in Table 7.
|
function [resultsTable, metricsTable] = NN_model(train_data, test_data) % Neural Network classifier with performance metrics % Extract predictors and response predictorNames = {'Max', 'Min', 'Median', 'Std', 'RMS', 'Kurtosis', 'Skewness', 'Range', 'CF', 'IF'}; predictors = train_data(:, predictorNames); response = train_data.Target; % Train neural network classificationNeuralNetwork = fitcnet(predictors, response, ... 'LayerSizes', [85 80], 'Activations', 'relu', ... 'Lambda', 2.230905500034876e-08, 'IterationLimit', 1000, ... 'Standardize', True); % Cross-validation (Training metrics) partitionedModel = crossval(classificationNeuralNetwork, 'KFold', 5); validationPredictions = kfoldPredict(partitionedModel); train_accuracy = (1 - kfoldLoss(partitionedModel)) * 100; % Testing metrics testPredictions = predict(classificationNeuralNetwork, test_data(:, predictorNames)); test_accuracy = sum(testPredictions == test_data.Target) / numel(test_data.Target) * 100; % Calculate costs costMatrix = ones(4) - eye(4); % 0 diagonal, 1 elsewhere confMat_train = confusionmat(response, validationPredictions); confMat_test = confusionmat(test_data.Target, testPredictions); train_cost = sum(sum(confMat_train .* costMatrix)); test_cost = sum(sum(confMat_test .* costMatrix)); % Calculate metrics [train_recall, train_precision, train_f1] = calculate_metrics(confMat_train); [test_recall, test_precision, test_f1] = calculate_metrics(confMat_test); % Create results tables resultsTable = table(round(train_accuracy,1), train_cost, round(test_accuracy,1), test_cost, ... 'VariableNames', {'Train_Acc', 'Train_Cost', 'Test_Acc', 'Test_Cost'}, 'RowNames', {'NN'}); metricsTable = table(round(train_recall,3), round(train_precision,3), round(train_f1,3), ... round(test_recall,3), round(test_precision,3), round(test_f1,3), ... 'VariableNames', {'Train_Rec', 'Train_Pre', 'Train_F1', 'Test_Rec', 'Test_Pre', 'Test_F1'}, ... 'RowNames', {'NN'}); end function [recall, precision, f1] = calculate_metrics(confMat) % Calculate macro-average metrics from confusion matrix n = size(confMat,1); recall_vec = zeros(n,1); precision_vec = zeros(n,1); for i = 1:n recall_vec(i) = confMat(i,i) / sum(confMat(i,:)); precision_vec(i) = confMat(i,i) / sum(confMat(:,i)); end |
Appendix C. Bayesian optimization framework for the ONN
This appendix describes the methodology and hyperparameter search space used for the Bayesian optimization of the ONN. The process was carried out to automatically identify the most effective neural architecture and training parameters for the bearing fault diagnosis task, going beyond manually predefined model structures.
C.1 Optimization setup
|
Objective |
Minimize 5-fold cross-validation classification error |
|
Algorithm |
Bayesian optimization with Gaussian process |
|
Iterations |
30 iterations |
|
Implementation |
MATLAB Classification Learner toolbox |
C.2 Bayesian optimization search space configuration
|
Hyperparameter |
Search Space |
Description |
|
Number of Layers |
1, 2, 3 |
Number of fully connected layers |
|
First Layer Size |
[1, 300] |
Number of neurons in first hidden layer |
|
Second Layer Size |
[1, 300] |
Number of neurons in second hidden layer |
|
Third Layer Size |
[1, 300] |
Number of neurons in third hidden layer |
|
Activation Function |
ReLU, Tanh, |
Activation Function |
|
Regularization Strength (Lambda) |
[1e-5/n, 1e5/n] |
Regularization Strength (Lambda) |
|
Standardize Data |
True, False |
Standardize input predictors before training |
[1] Thomson, W.T., Fenger, M. (2001). Current signature analysis to detect induction motor faults. IEEE Industry Applications Magazine, 7(4): 26-34. https://doi.org/10.1109/2943.930988
[2] Benbouzid, M.E.H., Kliman, G.B. (2003). What stator current processing-based technique to use for induction motor rotor faults diagnosis? IEEE Transactions on Energy Conversion, 18(2): 238-244. https://doi.org/10.1109/TEC.2003.811741
[3] Zhang, P., Du, Y., Habetler, T.G., Lu, B. (2010). A survey of condition monitoring and protection methods for medium-voltage induction motors. IEEE Transactions on Industry Applications, 47(1): 34-46. https://doi.org/10.1109/TIA.2010.2090839
[4] Riera-Guasp, M., Antonino-Daviu, J.A., Capolino, G.A. (2014). Advances in electrical machine, power electronic, and drive condition monitoring and fault detection: State of the art. IEEE Transactions on Industrial Electronics, 62(3): 1746-1759. https://doi.org/10.1109/TIE.2014.2375853
[5] Li, X., Yang, Y., Pan, H., Cheng, J., Cheng, J. (2019). A novel deep stacking least squares support vector machine for rolling bearing fault diagnosis. Computers in Industry, 110: 36-47. https://doi.org/10.1016/j.compind.2019.05.005
[6] Smith, W.A., Randall, R.B. (2015). Rolling element bearing diagnostics using the Case Western Reserve University data: A benchmark study. Mechanical Systems and Signal Processing, 64: 100-131. https://doi.org/10.1016/j.ymssp.2015.04.021
[7] Peng, B., Bi, Y., Xue, B., Zhang, M., Wan, S. (2022). A survey on fault diagnosis of rolling bearings. Algorithms, 15(10): 347. https://doi.org/10.3390/a15100347
[8] Wu, G., Yan, T., Yang, G., Chai, H., Cao, C. (2022). A review on rolling bearing fault signal detection methods based on different sensors. Sensors, 22(21): 8330. https://doi.org/10.3390/s22218330
[9] Gangsar, P., Tiwari, R. (2020). Signal based condition monitoring techniques for fault detection and diagnosis of induction motors: A state-of-the-art review. Mechanical Systems and Signal Processing, 144: 106908. https://doi.org/10.1016/j.ymssp.2020.106908
[10] Bensaoucha, S., Bessedik, S.A., Ameur, A., Teta, A. (2019). Induction motors broken rotor bars detection using RPVM and neural network. COMPEL-The International Journal for Computation and Mathematics in Electrical and Electronic Engineering, 38(2): 596-615. https://doi.org/10.1108/COMPEL-06-2018-0256
[11] Elbouchikhi, E., Choqueuse, V., Amirat, Y., Benbouzid, M.E.H., Turri, S. (2017). An efficient Hilbert–Huang transform-based bearing faults detection in induction machines. IEEE Transactions on Energy Conversion, 32(2): 401-413. https://doi.org/10.1109/TEC.2017.2661541
[12] Puche-Panadero, R., Pineda-Sanchez, M., Riera-Guasp, M., Roger-Folch, J., Hurtado-Perez, E., Perez-Cruz, J. (2009). Improved resolution of the MCSA method via Hilbert transform, enabling the diagnosis of rotor asymmetries at very low slip. IEEE Transactions on Energy Conversion, 24(1): 52-59. https://doi.org/10.1109/TEC.2008.2003207
[13] Georgoulas, G., Loutas, T., Stylios, C.D., Kostopoulos, V. (2013). Bearing fault detection based on hybrid ensemble detector and empirical mode decomposition. Mechanical Systems and Signal Processing, 41(1-2): 510-525. https://doi.org/10.1016/j.ymssp.2013.02.020
[14] Ghabar, I., Burqan, A., Gharib, G. (2024). The optical model absorption term in the frame of fractional derivatives. Atoms, 12(7): 37. https://doi.org/10.3390/atoms12070037
[15] Chennana, A., Megherbi, A.C., Bessous, N., Sbaa, S., et al. (2025). Vibration signal analysis for rolling bearings faults diagnosis based on deep-shallow features fusion. Scientific Reports, 15(1): 9270. https://doi.org/10.1038/s41598-025-93133-y
[16] Chennana, A., Bessous, N., Megherbi, A.C., Saidi, L., Sbaa, S., Sayadi, M., Teta, A. (2025). Deep-shallow features fusion for induction motor fault diagnosis through infrared thermography images. In 2025 International Conference on Control, Automation and Diagnosis (ICCAD), Barcelona, Spain, pp. 1-6. https://doi.org/10.1109/ICCAD64771.2025.11099458
[17] Bensaoucha, S., Bessedik, S.A., Ameur, A., Moreau, S., Teta, A. (2020). A comparative study for broken rotor bars fault detection in induction machine using DWT and MUSIC techniques. In 2020 1st International Conference on Communications, Control Systems and Signal Processing (CCSSP), El Oued, Algeria, pp. 523-528. https://doi.org/10.1109/CCSSP49278.2020.9151772
[18] Shendi, A.F., Nahhas, A.M. (2023). Power generation by using Photovoltaic systems for Yanbu and Rabigh regions in Saudi Arabia: A cost-effective study. The Islamic University Journal of Applied Sciences (IUJAS), 2023(7): 10-44. https://doi.org/10.63070/jesc.2023.001
[19] Alrehaily, A. (2024). Advances in bioinformatics techniques to predict neoantigen: Exploring tumor immune microenvironment and transforming data into therapeutic insights. The Islamic University Journal of Applied Sciences (IUJAS), 2024(12): 81-108. https://doi.org/10.63070/jesc.2024.016
[20] Yang, T., Huang, S. (2017). Fault diagnosis based on improved deep belief network. In 2017 5th International Conference on Enterprise Systems (ES), Beijing, China, pp. 305-310. https://doi.org/10.1109/ES.2017.57
[21] Xia, M., Li, T., Xu, L., Liu, L., De Silva, C.W. (2017). Fault diagnosis for rotating machinery using multiple sensors and convolutional neural networks. IEEE/ASME transactions on mechatronics, 23(1): 101-110. https://doi.org/10.1109/TMECH.2017.2728371
[22] Konar, P., Chattopadhyay, P. (2011). Bearing fault detection of induction motor using wavelet and Support Vector Machines (SVMs). Applied Soft Computing, 11(6): 4203-4211. https://doi.org/10.1016/j.asoc.2011.03.014
[23] Alsaawy, Y. (2024). Machine learning pipeline: Feature selection and adaptive training for DDoS detection to improve cloud security. The Islamic University Journal of Applied Sciences (IUJAS), 5(2): 242-272. https://doi.org/10.63070/jesc.2024.023
[24] Siddique, M.F., Zaman, W., Ullah, S., Umar, M., et al. (2024). Advanced bearing-fault diagnosis and classification using mel-scalograms and FOX-optimized ANN. Sensors, 24(22): 7303. https://doi.org/10.3390/s24227303
[25] Chen, X., Li, J., Yu, A., Cai, B., Wu, Q., Xia, M. (2025). Ultra-low latency ANN-SNN conversion for bearing fault diagnosis. IEEE Transactions on Instrumentation and Measurement, 74: 1-10. https://doi.org/10.1109/TIM.2025.3548254
[26] Zhang, S., Zhang, S., Wang, B., Habetler, T.G. (2020). Deep learning algorithms for bearing fault diagnostics—A comprehensive review. IEEE Access, 8: 29857-29881. https://doi.org/10.1109/ACCESS.2020.2972859
[27] Bakria, D., Beladel, A., Korich, B., Teta, A., et al. (2025). A novel enhanced Grey Wolf Optimizer for global optimization problems: Application to photovoltaic parameter extraction. Energy Reports, 14: 2782-2796. https://doi.org/10.1016/j.egyr.2025.09.027
[28] Tibermacine, I.E., Russo, S., Scarano, G., Tedesco, G., et al. (2025). Conditional VAE for personalized neurofeedback in cognitive training. PLoS One, 20(10): e0335364. https://doi.org/10.1371/journal.pone.0335364
[29] Teta, A., Elbar, M., Beladel, A., Bakria, D., Korich, B., Chennana, A., Rabehi, A. (2025). Robust and fast fault diagnosis framework for grid-connected photovoltaic systems via adaptive channel-wise representations and transfer learning. Measurement, 258: 119290. https://doi.org/10.1016/j.measurement.2025.119290
[30] Mohamed, A., Nacera, Y., Ahcene, B., Teta, A., et al. (2025). Optimized YOLO based model for photovoltaic defect detection in electroluminescence images. Scientific Reports, 15(1): 32955. https://doi.org/10.1038/s41598-025-13956-7
[31] Ferkous, K., Menakh, S., Guermoui, M., Bellaour, A., et al. (2025). Optimized solar power forecasting: A multi-decomposition framework using VMD and swarm techniques. AIP Advances, 15(9). https://doi.org/10.1063/5.0282210
[32] Taibi, A., Ikhlef, N., Aomar, L., Touati, S., et al. (2025). Diagnosis of misalignment faults using the DTCWT-RCMFDE and LSSVM algorithms. Scientific Reports, 15(1): 32128. https://doi.org/10.1038/s41598-025-12407-7
[33] Ali, M., Souahlia, A., Rabehi, A., Guermoui, M., et al. (2025). A robust deep learning approach for photovoltaic power forecasting based on feature selection and variational mode decomposition. Journal of the Nigerian Society of Physical Sciences, 7(3): 2795. https://doi.org/10.46481/jnsps.2025.2795
[34] Ali, M., Rabehi, A., Souahlia, A., Guermoui, M., et al. (2025). Enhancing PV power forecasting through feature selection and artificial neural networks: A case study. Scientific Reports, 15(1): 22574. https://doi.org/10.1038/s41598-025-07038-x
[35] Ouahabi, M.S., Benyounes, A., Barkat, S., Ihammouchen, S., et al. (2025). Real-time sensor fault tolerant control of DC-DC converters in DC microgrids using a switching unknown input observer. IEEE Access, 13: 95837-95850. https://doi.org/10.1109/ACCESS.2025.3571650
[36] Ladjal, B., Nadour, M., Bechouat, M., Hadroug, N., et al. (2025). Hybrid deep learning CNN-LSTM model for forecasting direct normal irradiance: A study on solar potential in Ghardaia, Algeria. Scientific Reports, 15(1): 15404. https://doi.org/10.1038/s41598-025-94239-z
[37] Bentegri, H., Rabehi, M., Kherfane, S., Nahool, T.A., et al. (2025). Assessment of compressive strength of eco-concrete reinforced using machine learning tools. Scientific Reports, 15(1): 5017. https://doi.org/10.1038/s41598-025-89530-y
[38] Mehallou, A., M’hamdi, B., Amari, A., Teguar, M., et al. (2025). Optimal multiobjective design of an autonomous hybrid renewable energy system in the Adrar Region, Algeria. Scientific Reports, 15(1): 4173. https://doi.org/10.1038/s41598-025-88438-x
[39] Ladjal, B., Tibermacine, I.E., Bechouat, M., Sedraoui, M., et al. (2024). Hybrid models for direct normal irradiance forecasting: A case study of Ghardaia zone (Algeria). Natural Hazards, 120(15): 14703-14725. https://doi.org/10.1007/s11069-024-06837-1
[40] Baitiche, O., Bendelala, F., Cheknane, A., Rabehi, A., Comini, E. (2024). Numerical modeling of hybrid solar/thermal conversion efficiency enhanced by metamaterial light scattering for ultrathin PbS QDs-STPV cell. Crystals, 14(7): 668. https://doi.org/10.3390/cryst14070668
[41] Hamadneh, T., Batiha, B., Gharib, G.M., Aribowo, W. (2025). Application of Orangutan Optimization Algorithm for feature selection problems. INASS Express, 1(1): 1-9. https://doi.org/10.22266/inassexpress.2025.001
[42] Neupane, D., Seok, J. (2020). Bearing fault detection and diagnosis using case western reserve university dataset with deep learning approaches: A review. IEEE Access, 8: 93155-93178. https://doi.org/10.1109/ACCESS.2020.2990528
[43] Alsaoudi, M., Gharib, G.M., Al-Husban, A., Abudayyeh, J.A. (2026). A unified framework for solving abel's and linear volterra integral equations and their neutrosophic generalizations using the GALM transform. International Journal of Neutrosophic Science (IJNS), 27(1): 19-35. https://doi.org/10.54216/IJNS.270103