A Two-Stage Multi-Source Unilateral Alignment Method for Rail Insulation Fault Localization Under Target-Domain Missing-Class Conditions

To address the decline in insulation fault localization accuracy of urban rail DC traction return systems under varying operating conditions and missing target-domain classes, this study proposes a two-stage multi-source unilateral alignment network (MUAN). In practice, different loads and operating conditions lead to different rail-potential patterns, while the target-domain training set often lacks some fault classes. Conventional domain adaptation methods usually assume identical label spaces across domains, which may cause source-private classes to be incorrectly aligned and thus induce negative transfer. To mitigate this issue, a multi-source partial domain adaptation framework is developed. In the first stage, labeled source data from multiple conditions are used to learn discriminative features, and source anchor features are extracted for subsequent transfer. In the second stage, source-alignment and cross-domain alignment modules are introduced to guide the target features toward the shared class space while preserving the source class structure. Domain-adversarial learning is further employed to reduce cross-condition distribution gaps and improve generalization to unlabeled target data with missing classes. Experiments on both a dynamic rail-potential simulation platform and a return-system hardware platform show that the proposed method achieves superior fault localization performance, especially under target-domain missing-class settings.