Adaptive time-frequency signal enhancement and deep spatial-temporal fusion for noise-resistant bearing fault...

In industrial environments, substantial background noise often obscures the crucial impact and periodic features within bearing vibration signals, hindering the efficacy of conventional diagnostic methods. To address this challenge, this paper proposes a novel fault diagnosis framework for rolling bearings, initiated by an advanced signal enhancement stage using a time-frequency extension of adaptive filtering (TFEAF) method. This method effectively enhances the signal's time-frequency representation, achieving robust denoising and revealing the underlying fault characteristics previously masked by noise. Subsequently, we introduce a novel Huge Kernel Convolutional Bidirectional Long Short-Term Memory Network with a Cross-Attention mechanism (HCBM-CANet). This network is specifically designed to extract deep spatial and temporal features from the denoised signal. The integrated cross-attention mechanism synergistically fuses these features, achieving significant boosts in diagnostic precision under complex operating conditions. Experimental results on the CWRU and Paderborn University datasets demonstrate the superiority of our approach. Under high-noise scenarios, the proposed TFEAF-HCBM-CANet framework achieves an average diagnostic accuracy exceeding 99 %, outperforming seven state-of-the-art network models. These findings validate the exceptional robustness and effectiveness of our method for processing signals heavily corrupted by noise.