Distributed Sensor for Water Detection via Data-Driven TDR Inversion

Pipeline leakage remains a major challenge in public water supply systems, contributing to significant water losses, financial burdens, and environmental concerns. Traditional leak detection methods, such as visual inspections, electromagnetic sensing, and acoustic techniques, are often constrained by pipe material, network conditions, and reliance on expert interpretation. Time-domain reflectometry (TDR) presents a promising alternative by analyzing electromagnetic wave reflections along a sensing element placed near the pipeline. However, effective leak detection requires solving the TDR inversion problem, i.e., estimating the spatial distribution of dielectric properties from the measured reflectogram. This paper presents preliminary results on a data-driven approach to TDR inversion using a transformer-based neural network trained on simulated reflectograms under diverse conditions. Unlike conventional methods that rely on iterative optimization, the proposed approach provides one-shot predictions in a fully automated manner, assuming the training dataset is representative of the target scenario. The method is validated through simulations and demonstrates strong potential for accurate and continuous pipeline monitoring.