DAS-Based Vehicle Detection and Speed-Aware Signal Modeling for Railway Monitoring

Distributed Acoustic Sensing (DAS) enables large-scale railway train detection and tracking by capturing ground vibrations through standard roadside optical fibers. However, quasi-static signals generated by train axle loads often suffer from reduced spatial resolution due to factors such as finite gauge length and lateral distance between the fiber and the track. Additionally, while train speed significantly alters the temporal response, it leaves the spatial strain profile largely unchanged, posing challenges for accurate and real-time interpretation. To overcome these limitations, we propose a physics-informed learning framework that enhances both detection performance and processing efficiency. Temporal features sensitive to speed are extracted via Fast Fourier Transform (FFT), while a Long Short-Term Memory (LSTM) network captures speed-invariant spatial features for robust localization. The strain field is analytically modeled using the Flamant-Boussinesq formulation, and synthetic data illustrate the influence of physical parameters on signal response. Experimental validations demonstrate that the proposed approach improves localization accuracy and enables real-time monitoring across a wide range of train speeds, making it well-suited for intelligent railway infrastructure applications.