High precision RailFusion-DAS for railway track defect detection using existing communication fibers

This paper proposes RailFusion-DAS, a novel method for detecting railway track defects using existing communication fibers in air-filled trenches via distributed acoustic sensing (DAS). Air-coupled fibers exhibit severe vibration attenuation, generating extremely weak signals during train passages. To overcome this challenge, our framework derives multi-kinematic representations (displacement, velocity, acceleration) from raw DAS data. We employ dual-domain processing: CNN-LSTM networks capture temporal defect signatures while variational mode decomposition (VMD) extracts frequency-domain features. Furthermore, an attentional feature fusion mechanism then dynamically integrates these complementary representations, prioritizing diagnostically critical patterns in low-SNR signals. Validated on Zhangjiakou heavy-haul railway trials, RailFusion-DAS consistently achieves an average accuracy of 97.11% in classifying four typical defects of corrugation, unsupported sleepers, rolling contact fatigue, and normal states during random tests. This transforms communication infrastructure into a sensitive monitoring network, enabling cost-effective, long-distance railway maintenance.