An infrared-vibration-based DMCF-Net for axlebox bearing fault diagnosis

The utilization of multimodal information fusion to leverage information complementarity and comprehensively perceive the operating state of faulty equipment represents a significant trend in future fault diagnosis development. However, challenges such as inadequate coordination of modal differences and inefficient integration of heterogeneous information persist. To address these issues, this study proposes an infrared-vibration-based dual-branch multimodal cooperative fusion network (DMCF-Net) for axlebox bearing fault diagnosis. This method constructs a DMCF-Net and adopts a shallow-to-deep two-stage heterogeneous feature fusion strategy. Firstly, parallel dual-branch modules are utilized to extract differentiated features from infrared and vibration modal data. Secondly, a feature dynamic fusion module (FDFM) is embedded to highlight key spatial channel features, enabling shallow-layer fusion of heterogeneous features. Finally, the Mamba cooperative feature interaction enhancement method is employed to accomplish deep-layer fusion of multimodal features through dynamic weighted interaction. Experimental results from the QPZZ-II and the train bearing comprehensive test bench demonstrate that the proposed method achieves fault diagnosis accuracies of 99.75% and 99.67%, respectively, significantly outperforming traditional single-modal and multimodal analysis approaches. This work provides an effective paradigm for multimodal heterogeneous feature fusion in rotating machinery fault diagnosis.