A Domain Generalization Network with Multi-domain Feature Extractors for Running Rail Fault Localization under Unseen Working Conditions in Urban Rail Transit

The DC traction power supply systems (DC TPSS) typically employ running rails as the traction current return path. However, as urban rail transit lines operate for increasing years and the operating environment deteriorates, weak insulation in the running rails is prone to occur. In serious cases, even the running rails are directly grounded, resulting in an abnormal increase in rail potential (RP) and an unusual increase in stray current (SC), which leads to a series of safety hazards and equipment corrosion problems. Therefore, this study presents an insulation fault location approach based on a multi-domain feature extractor for domain generalization networks (MDFE), aiming to enable cross-domain fault diagnosis under unseen target conditions. To obtain domain-specific features from every source domain, the network is equipped with multiple domain feature extractors. After that, the original signal is mapped to a new feature space using an encoder-decoder in which the learned domain-specific features are removed. The finalized architecture implements a domain-invariant classifier to assimilate generalizable diagnostic knowledge across multiple source domains, thereby enabling task execution under unseen conditions. The findings show that the network achieves high fault location accuracy for unlabeled rail potential data under various operating conditions.