Multimodal bearing fault diagnosis based on semantic-aware and enhanced cross-attention

Abstract Bearing fault diagnosis is a critical issue in rotating machinery. However, existing multimodal diagnostic methods often struggle to effectively extract key information in environments of strong noise interference. Large language models possess the capability to understand the internal physical mechanisms of bearings and can assist in feature learning, thereby enhancing diagnostic performance. Therefore, a semantic-aware and enhanced cross-attention method is proposed for multimodal bearing fault diagnosis. First, a semantic-guided feature mapping mechanism enhances noise robustness by purifying sensor signals within a unified semantic space. Subsequently, a time-event synchronization strategy achieves precise cross-modal alignment. Finally, the derived domain-invariant semantic representations overcome the limitations of cross-operating-condition generalization in multimodal data fusion. To address sufficient separation and utilization of shared and private features in multimodal diagnosis, a cross-modal cross-attention mechanism is developed to build correlations between modalities, enabling information complementarity and enhancement. Furthermore, an signal-to-noise ratio-based gating mechanism is introduced to dynamically suppress noise interference in modal features. Experimental results on two publicly available datasets show that the proposed method achieves high fault classification accuracy even in various strong noise environments and maintains robust performance under complex operating conditions, thereby fully validating its effectiveness and superiority.