A spatiotemporal dual-channel bearing fault diagnosis method integrating variational mode decomposition and cross-attention

Rolling bearings are key components of rotating machinery and their faults are a primary cause of industrial equipment shutdowns, with particularly significant impact in high-value fields such as wind power and rail transportation. Traditional methods relying on manual features lead to poor robustness, while deep learning methods exhibit weak generalisation under noise and complex operating conditions. Existing fusion approaches still suffer from feature redundancy and insufficient fusion due to inadequate modal feature concatenation and cross-modal interaction. This paper proposes a hybrid model integrating variational mode decomposition (VMD)-based multi-scale decomposition and cross-modal attention mechanisms (cross-attention). Using a Transformer-bidirectional long short-term memory (Transformer-BiLSTM) dual-channel architecture to extract spatiotemporal features, the cross-attention dynamically aligns and fuses modal and deep features, significantly improving the robustness and adaptability of bearing fault diagnosis under noisy and varying conditions. The experimental results show that, on the Case Western Reserve University (CWRU) public bearing dataset, this model demonstrates excellent diagnostic performance in both strong noise backgrounds and complex variable load conditions, with an average recognition accuracy of over 99%. Under extreme noise, it exhibits outstanding stability and generalisation; ablation experiments confirm that removing VMD or cross-attention leads to a 1.27% to 2.28% drop in performance. In multi-condition tests (0 HP to 3 HP loads), it maintains an accuracy of 99.02% to 99.15%, with precision, recall and F1 scores all stable above 0.99, confirming strong adaptability to complex conditions and providing a robustness solution for intelligent operation and maintenance of industrial equipment.