Wavelet Scattering Transformer: A Robust Rail Corrugation Detection Method Under Varying Speed Condition Based on Multiscale Time–Frequency Feature Representation

Abstract

Rail corrugation detection, which identifies regular undulation wear on rail surfaces, is crucial for improving metro safety and ensuring system reliability. Current detection methods are usually limited to predetermined corrugation types and fixed speeds, which is not applicable to actual metro operations in an open environment. This article proposes a robust and reliable corrugation detection model, wavelet scattering transformer (WSTrans for short), by incorporating wavelet scattering transform and tensor decomposition into a deep transformer architecture. The original vibration signals are first fed into the tensor Tucker decomposition to get the core feature representation, enhancing the model’s reliability by filtering out irregular noise. Then, the obtained features are input into a wavelet scattering attention (WSA) mechanism that uses wavelet scattering transform as filters to capture multiscale information across the full-frequency band and uses channel attention to select critical frequency channels. The variation characteristics of passing frequency at varying speeds can be obtained to reliably identify whether a corrugation occurs through the transformer’s feedforward layer. An alternating training algorithm is designed to seek the optimal corrugation feature representation and the best robustness to noise interference. Experiments with real-world data from Beijing Subway in 2023 validate WSTrans’s effectiveness and reliability in adaptively detecting rail corrugation across varying speeds.