Pseudo-fault data enhanced relation network for fault detection and localization in train transmission systems

Abstract

Fault detection and localization is a critical technology for ensuring the safe and cefficient operation of train transmission systems (TTS). However, due to the frequent scarcity or absence of fault data, achieving precise detection and localization faces many challenges. To address this challenge, this paper proposes a novel pseudo-fault data enhanced relation network (PE-ERN) for fault detection and localization in TTS. The PE-ERN method generates pseudo-fault data by combining fault information from similar external equipment and out-of-distribution data. This strategy enriches the training dataset, enabling model to extract intrinsic heath state information from available health data. Additionally, a feature concatenation mechanism is developed to generate feature pairs by combining health state data with pseudo-fault data. This mechanism uncovers both health-unique and health-inherent attributes, which enhances the PE-ERN's ability to distinguish between fault and health states. Experimental results from subway TTS cases demonstrate that the proposed PE-ERN method outperforms existing fault detection techniques, achieving superior fault localization and detection accuracy across various fault modes, including single-fault, component-level compound-fault, and system-level compound-fault scenarios.