Automatic identification of high-speed railway wheelset defects by integrating PointNet++ and Swin Transformer

In order to address the technical challenges of detecting defects in high-speed railway wheelsets under complex conditions such as dynamic lighting, foreign object occlusion, and microscale anomalies, this paper proposes a dual-mode deep learning framework that integrates PointNet++ and Swin Transformer. This paper enhances defect recognition through cross modal feature collaboration, and combines cross modal attention (CMA) mechanism for dynamic feature alignment and geometric guidance suppression strategy for reducing occlusion noise. The experimental results showed an accuracy of 0.985, an F1 value of 0.982, and a recognition rate of 0.938 for defects smaller than 1 millimeter. Research has shown that the model maintains robust accuracy under different lighting conditions (strong/weak/reflective) and up to 40% occlusion, while optimized deployment on edge devices can achieve 23FPS with only 12M parameters. This work significantly improves the intelligence and reliability of the high-speed railway wheelset detection system.