Accurate and Fast Quantitative Detection of Rail Corrugation Driven by Deep Learning Algorithms Using Adaptive Targeted Loss and Directed Metric Functions

Rail corrugation is a common rail surface defect in wheel–rail system and can significantly reduce the running quality of vehicles. Accurate and fast detection of corrugation is of great importance for uncovering its mechanism, predicting the evolution and planning maintenance. Traditional physical model-driven detection approaches rely on high-fidelity simulations and are poorly automated. The popular data-driven methods have good availability. However, they ignore the power of expert experience, resulting in limited accuracy. This paper presents a quantitative model for detecting rail corrugation based on measurement data and expert knowledge using deep learning (DL) algorithms. First, a data quality improvement framework is proposed to overcome the problem that conventional DL models may produce incorrect results when trained on low-quality datasets. Then, the characteristics of rail corrugation are deeply investigated and adaptive targeted loss and directed metric functions are developed according to the individual samples to build a detection model with better global feature regression and local detail convergence. Finally, the effectiveness of the proposed method is verified using on-site measurement data from a commercial metro line. Comparative analyses and ablation studies demonstrate the superiority of targeted loss and directed metric in terms of more accurate detection and scientific evaluation of rail corrugation roughness, respectively.