Denoised Vibration Modeling for Early Fault Identification in Metro Track Monitoring

Abstract

Metro track monitoring systems continuously collect vibration time series from rail fasteners, track beds, wheel-rail contact points, tunnel structures, and onboard inspection devices. These signals are strongly affected by train speed, passenger load, tunnel echo, wheel condition, and environmental vibration, making early fault detection difficult under noisy operating conditions. This study proposes a denoised vibration sequence modeling method for early fault detection in metro track monitoring systems. The method uses a diffusion-based denoising module to recover stable vibration patterns from noisy temporal signals. A disentangled representation layer separates train-operation noise, structural vibration variation, and fault-related residual components. Anomaly scores are then calculated from denoised residual energy and component-specific deviation boundaries. Experiments are conducted on a metro vibration dataset covering 11 subway lines, 286 track sections, 740 vibration sensors, and 5-second records collected over 14 months. The dataset contains 218 million vibration records and 1,540 maintenance-confirmed abnormal events, including rail corrugation, fastener loosening, track-bed settlement, wheel-rail impact, and tunnel segment vibration abnormality. The proposed method reduces median early-warning delay from 6.8 days to 2.1 days compared with a wavelet-denoising autoencoder baseline. False maintenance alerts are controlled at 1.9 cases per track section per quarter. Diffusion denoising improves the average vibration signal-to-noise ratio by 6.2 dB, and event-boundary deviation decreases by 11.6 hours. The model completes daily line-level assessment in 8.4 minutes. The results show that diffusion-driven denoising and component disentanglement can improve robust anomaly detection in noisy metro vibration time series.