Bearing Fault Diagnosis via Adaptive Feature Fusion Networks

Accurate bearing fault diagnosis remains challenging due to complex vibration patterns and noise interference. This paper presents an Adaptive Feature Fusion Network (AFFNet) that innovatively integrates the db4 wavelet transform with variable mode decomposition (VMD) through a dual attention mechanism framework. The proposed method establishes a new paradigm for fault feature extraction by simultaneously leveraging db4 wavelet’s precise time-frequency localization and VMD’s adaptive noise suppression capability. A novel hybrid attention architecture, combining SENet’s channel-wise feature recalibration with GAM’s global context modeling, enables intelligent fusion of complementary feature representations. The cross-attention-based feature integration mechanism further enhances the model’s ability to discern subtle fault characteristics under varying operational conditions. Experimental validation on the CWRU 10-class bearing dataset demonstrates the superior performance of the framework, achieving perfect 100% classification accuracy while significantly outperforming existing approaches including CNN-1D (95.54%) and LSTM (98.66%), etc. The system exhibits remarkable robustness, maintaining 99.55% accuracy under challenging 10dB noise conditions, confirming its practical viability for industrial applications. The technical advances include the synergistic combination of time-frequency analysis techniques with advanced attention mechanisms, establishing a new state-of-the-art in bearing fault diagnosis.