An Intelligent Assessment of Rail Surface Defects over the Life-Cycle Based on Improved Transformer Networks

Accurate assessment of the failure stage of rail rolling contact fatigue (RCF) is critical for guiding timely maintenance by track personnel, ensuring safe rail operations, and reducing maintenance costs. Although various methods have been developed to detect rail damage and classify surface defects, the rolling contact fatigue failure state of rails has not yet been comprehensively and objectively evaluated. This paper introduces the application of image processing and improved deep-learning network algorithms in rail failure evaluation and judgment. Based on Swin Transformer, a deep learning network is developed. By dividing the rail rolling contact fatigue failure process into five life-cycle stages, the proposed network can identify the current stage of the rail contact surface within its service life. Finally, compared with the commonly used neural network model, the recognition rate of the improved Transformer can reach 98.48%, which is far better than other network structures. The enhanced neural network forms a simple system for evaluating the life of the orbit. The system identifies potential failure hazards on rail surfaces. The results also provide early warning predictions for rolling contact fatigue failure.