Transformer-Based Bearing Fault Classification with VMD-Based Noise Suppression and rCCA-Enhanced Correlation Modeling

Abstract

Early detection of bearing faults in rotating machinery is essential for ensuring system reliability and effective maintenance planning. Vibration signals inherently contain characteristic fault-related frequency components, providing rich information for both physically interpretable and data-driven analyses. In this study, a multi-representation and correlation-aware feature extraction framework is proposed for automatic classification of bearing faults from vibration signals. Experimental evaluations are conducted using the Case Western Reserve University (CWRU) Bearing Dataset. The dataset includes vibration recordings corresponding to inner race, outer race, ball faults, and healthy conditions under different damage severities. The proposed approach first applies Variational Mode Decomposition (VMD) to suppress noise and enhance frequency-related characteristics. Three different feature representations are then constructed: analytical spectral descriptors, raw Transformer-based deep representations, and a hybrid feature vector obtained by combining these two representations. The hybrid structure is further enhanced through regularized Canonical Correlation Analysis (rCCA), which models the relationship between Transformer representations and spectral descriptors, enabling correlation-aware feature fusion. Spectral, raw Transformer, and rCCA-enhanced hybrid feature vectors are evaluated separately using SVM, Random Forest, and XGBoost classifiers. The results demonstrate that both spectral and Transformer-based representations provide strong performance individually; however, integrating these complementary information sources while modeling their correlations leads to superior and more balanced classification performance. In particular, the rCCA-enhanced hybrid feature vector achieves the best results across all performance metrics. The findings indicate that combining physically meaningful frequency-domain information with data-driven deep representations yields a more robust and generalizable solution for bearing fault diagnosis.