FV-BiCA-Net: A Bearing Fault Diagnosis Method Based on FFT-VMD Hybrid Feature Enhancement and Bidirectional Cross-Attention

In rotating machinery monitoring, bearing faults can cause severe safety risks and economic losses. Traditional diagnostic methods struggle with non-stationary vibration signals and show poor generalization in real-world applications. To address these issues, we propose a hybrid model combining frequency-domain guided decomposition with time-domain feature enhancement. The model consists of three key modules: (1) An FFT-VMD module that optimizes decomposition parameters using spectral energy to enhance SNR; (2) An improved CNN-BiTCN module featuring two-layer convolution stacking and bidirectional temporal convolutions; (3) An EnhancedCrossAttention module for precise cross-domain feature alignment. On the CWRU bearing dataset, our method achieves $99.74 \%$ accuracy in a 10-class classification task and maintains over $99.2 \%$ accuracy with just 600 training samples. Compared to existing methods, our approach improves SNR by $28.8 \%$ (vs. EMD) and $41.7 \%$ (vs. wavelet transform). Industrial deployment tests confirm a false alarm rate reduction to $0.26 \%$, significantly outperforming traditional methods. This work provides an effective solution for enhancing the operational safety and maintenance efficiency of rotating machinery.