Diffusion-based unsupervised distributed acoustic sensing for rail condition monitoring and degradation assessment

Railway infrastructure deteriorates continuously under repeated loading and environmental effects, while maintenance decisions still rely on periodic inspections that cannot capture degradation evolution. This study proposes a distributed acoustic sensing (DAS) and diffusion-based reconstruction framework to quantify rail condition from long-term vibration data without requiring dense labels. Using a 2 km metro deployment monitored over one year, reconstruction-based indicators are shown to form a continuous degradation curve, distinguishing different stages of rail corrugation development. The results are consistent with vehicle–track dynamic simulations, indicating that the method reflects underlying degradation behavior. A model trained on a short replaced-rail segment generalizes to network-scale monitoring and identifies early-stage defects confirmed by maintenance records. The proposed approach supports continuous condition assessment and enables maintenance decisions to shift from fixed inspection cycles to data-driven, trend-based interventions.