A novel fault diagnosis model based on deep feature fusion network under imbalanced data: Towards railway dual-switch machines traction occasion

Abstract Fault diagnosis of railway switch machines is crucial for ensuring safe and efficient train operations, as well as for the maintenance of intelligent Switching & Crossing systems. Current methods primarily focus on single-switch machine traction modes, often overlooking the challenges of effectively utilizing multi-source data and comprehensively representing fault information. This limitation results in restricted applicability and suboptimal recognition accuracy. To address these challenges, we propose a novel fault diagnosis model based on a deep feature fusion network (DFFN) specifically designed for railway dual-switch machines (RDSMs) in traction occasion, particularly under imbalanced data conditions. First, we introduce an improved synthetic minority oversampling technique (ISMOTE) that integrates clustering technology with neighbor-based strategies to balance the experimental data and mitigate training bias. Second, we incorporate a cross-branch convolutional collaborative self-attention mechanism network (CBCSAMN) and an adaptive weight learning network (AWLN) into the DFFN, facilitating the extraction of multi-scale fault feature correlations and promoting efficient fusion. Experimental results, based on multiple vibration sensing signals, demonstrate an average diagnostic accuracy of 96.66% and an F1-score of 96.85% in real railway environments. Comparative analyses with other state-of-the-art methods confirm that our approach achieves superior diagnostic performance.