Transformer-based Time Domain Reflectometry Inversion for Distributed Sensing Applications

This paper presents a novel approach to Time-Domain Reflectometry (TDR) inversion based on a Transformer neural architecture (T5), originally developed for natural language processing and here adapted for numerical sequence analysis. The inversion is formulated as a sequence-to-sequence problem, where the network maps the TDR samples to the corresponding spatial distribution of the dielectric constant along a transmission line acting as a distributed sensor. The method is developed and thoroughly characterized through a case study on water-region localization, selected because it can be easily reproduced and controlled in laboratory conditions, thereby providing a rigorous framework for testing and validation. Although the study focuses on this specific application, the proposed approach is general and applicable to a broad class of TDR inversion problems involving distributed sensing and dielectric profiling. The Transformer model is trained on a synthetic dataset of reflectogram–capacitance profile pairs generated by a physical model of the measurement setup. Its performance is evaluated on both synthetic and experimental data, showing high accuracy in reconstructing the distributed capacitance profile and in localizing the water region. These results demonstrate that attention-based architectures can effectively perform TDR inversion and open new perspectives for data-driven distributed sensing.