Rolling element bearings are critical components in rotating machinery, and their sudden failure can result in significant economic and operational consequences in industrial systems. In recent years, the remarkable performance of machine learning models across a wide range of tasks has increased interest in applying these approaches to condition monitoring and fault diagnosis. However, challenges such as the complexity of vibration signal patterns, varying operating conditions, and differences in fault severity and location—which lead to distributional differences in data—have emphasized the need for intelligent methods in diagnostic systems. This study focuses on designing and eva‎luating learning-based models for classifying faults in rolling element bearings. Three benchmark datasets—CWRU, PU, and MFPT—were analyzed. Envelope spectrum analysis of their vibration signals revealed that while some data exhibited classic bearing fault features, others did not show such features or contained patterns inconsistent with typical bearing faults. Two types of neural networks, a multilayer perceptron (MLP) and a one-dimensional convolutional neural network (1D-CNN), were used for fault classification. The models were trained using two data-splitting strategies: random splitting and fault-size-based splitting. In the random split, models achieved nearly perfect accuracy, primarily due to data leakage, where parts of the same signal were present in both training and test sets. This resulted in overestimated performance. In contrast, fault-size-based splitting eliminated this issue and provided a more realistic view of the diagnostic task, though with reduced accuracy due to increased task complexity. To address this performance drop, various input representations were explored, including raw signals, signal spectra, envelope spectra, and cepstrum pre-whitened envelope spectra, across multiple input lengths. Unlike many previous studies that limited inputs to raw or frequency-domain signals, this research utilized richer, physics-based representations that better preserve fault-related information. Results showed that these enhanced inputs substantially improved model performance under more realistic and challenging diagnostic conditions. After conducting extensive hyperparameter tuning, the 1D-CNN consistently outperformed the MLP across both datasets. A simplified 1D-CNN model using envelope spectrum input was ultimately selected. This model achieved accuracies of 75.00% on the CWRU dataset and 73.93% on the PU dataset—surpassing the more complex models developed in earlier experiments. Further investigations into data quality revealed that removing signals that did not show classic bearing fault features significantly enhanced model performance. After removing the data samples that did not show classic bearing fault features, the model achieved 99.77% accuracy on the CWRU dataset and 92.03% on the PU dataset. Moreover, adjusting the input length and applying Gaussian noise for data augmentation raised PU accuracy to 98.6%. eva‎luations indicated that models trained on data containing classic fault features produced outputs that were well aligned with signal processing-based fault analysis, suggesting consistency with domain-specific understanding of fault behavior. Transfer learning was also employed to improve performance on data that lacked distinct indications of classic bearing faults. When PU was used as the source domain for transfer learning to CWRU, an accuracy improvement of 4.89% was observed. However, using CWRU or MFPT as the source for PU not only failed to improve results but led to decreased performance. These findings highlight the importance of source domain selection and the risk of negative transfer when structural or statistical discrepancies exist between domains.

علي لقماني , رضا تيكني

سعيد ضيائي راد , جلال ذهبي

بهزاد نظري , محمدرضا نيرومند