Generalized Zero-Shot Approach Leveraging Attribute Space for High-Speed Train Bogie

Incomplete data are a prevalent phenomenon in mechanical health management, posing limitations on the development of traditional data-driven strategies for mechanical fault diagnosis applications. This article investigates the practical challenge of achieving accurate fault diagnosis using deep learning algorithms in the absence of historical data about composite faults in high-speed train (HST) bogies. To address this issue, we propose a signal processing-assisted deep learning method and an attribute description strategy for fault modes based on signal time-domain features. We integrate the extended network of DenseNet into a generalized zero-shot learning (GZSL) framework and combine it with an attribute description matrix, transforming the mapping process from samples to labels in traditional data-driven strategies into a step-by-step mapping from samples to attributes to labels. The feasibility of fault diagnosis based on attribute description is analyzed and theoretically explained. Furthermore, we design zero-shot fault diagnosis experiments for bogies to verify the effectiveness of our proposed method. The results demonstrate that our approach can diagnose composite faults without requiring training data.