A Novel Method for Subway Wheelset Tread Defect Detection with Improved Self-Attention and Loss Function
Wheelset tread defects in subway locomotives present critical safety hazards, yet manual inspection remains prevalent, suffering from inefficiency and human error. This study proposes an enhanced YOLOv5-based detection algorithm tailored for subway wheelset tread defects. A multi-head self-attention module is integrated to capture long-range dependencies within global feature maps, improving small-target detection. A weighted bidirectional feature pyramid network (BiFPN) enables balanced multi-scale feature fusion and efficient cross-scale integration. To mitigate limited labeled data and annotation inaccuracies, a novel loss function, W-MPDIoU, is introduced to accelerate convergence. Experimental validation using real defect data and simulated experimental data yields an average detection accuracy of 99.1%, a 4.29% improvement over the original YOLOv5, with a detection speed of 15 ms per image. The model also outperforms YOLOv12 in convergence speed, detection accuracy, and inference speed. Despite these gains, limitations persist in defect variety and dataset size, necessitating further refinement for broader generalization. The proposed method enables real-time tread defect detection, enhancing safety and operational efficiency in urban rail transit maintenance.