Hierarchical Physics-Constrained Machine Learning Pipeline for Long-Distance TDR Fault Localization

Time Domain Reflectometry (TDR) is a fundamentaltechnique for fault localization in cable networks. However,accuracy significantly degrades over long distances (>15 km)due to heterogeneous variations in propagation velocity, signalattenuation, and human error. This paper proposes a hybridapproach combining a physically realistic TDR simulator with ahierarchical machine learning pipeline. We generate a massivedataset of 570,000 samples, explicitly modeling velocity variationsusing weighted harmonic means and operator-induced biases.The pipeline utilizes zone-based classification followed by specialized XGBoost regressors. Results demonstrate a classificationaccuracy of 99.61% and mean absolute errors (MAE) rangingfrom 8.34 m to 40.40 m for distances up to 30 km, providing arobust solution for field deployments.