Generalizable domain-invariant feature learning method for rail insulation fault localization under unseen load conditions in DC metro systems
Abstract
The running rails typically carry traction current back to the substation in DC metro systems (DC MS). However, with the extension of service years, the operating environment of DC metro lines deteriorates, degradation of running rail insulation is prone to occur. Direct grounding of the running rails under severe conditions can cause abnormal increases in rail potential (RP) and stray current (SC), thereby posing significant safety risks. Therefore, this paper proposes a generalizable domain-invariant feature learning (GDIFL) method for rail insulation fault localization under unseen load conditions in DC metro systems, aiming to enable the model to perform accurate insulation fault localization in unseen target domains. For each source domain, a domain feature extractor is utilized to characterize its domain-related information. The input signals are subsequently processed by an encoder–decoder network, which maps them into a feature representation while suppressing domain specific components. Domain-invariant classifier is then constructed to integrate generalizable knowledge from the source domains and support fault localization in unseen load conditions. Experimental findings demonstrate that the proposed approach effectively localizes insulation faults in unlabeled RP data across different load conditions.