A feature extraction framework for metro rail corrugation detection using onboard vibration and noise monitoring data

Abstract With the rapid expansion of subway operational mileage, rail corrugation—a periodic wave-like wear on rail surfaces—has become a critical issue, adversely affecting train safety, passenger comfort, and maintenance efficiency. Rapid and accurate detection of rail corrugation, combined with data-driven maintenance strategies, is essential for ensuring the safety and reliability of urban rail transit systems. This study introduces a novel feature extraction framework for rail corrugation detection, leveraging onboard vibration and acoustic monitoring data. Using Micro-Electro-Mechanical Systems sensors installed within train carriages, vibration and noise signals are synchronously collected, while field measurements of rail surface roughness provide ground truth data. A mapping relationship between rail corrugation and onboard noise is established, enabling the extraction of sensitive features through time-frequency domain analysis and wavelet decomposition. These features are integrated into machine learning models to classify rail corrugation states. The results demonstrate exceptional classification performance, with the Back Propagation Neural Network model achieving the highest accuracy of 95.31 per cent. This cost-effective and efficient approach offers a promising solution for real-time rail condition monitoring, contributing to the optimization of urban rail maintenance practices and ensuring safer subway operations.