Generalized Zero-Shot Learning for Fault Diagnosis in High-Speed Train Bogies Based on Enhanced Diffusion Generative Models

Abstract

In the context of high-speed trains (HST) bogie fault diagnosis, most existing state-of-the-art approaches struggle to effectively identify engineering fault types that lack historical records, leading to issues such as insufficient feature learning and high misdiagnosis rates. To tackle the challenges, this article introduces a generalized zero-shot learning (GZSL) strategy and presents a fault diagnosis framework for HST bogies, referred to as “ResDDPM-GZSL.” The study initially designs and constructs a foundational attribute description matrix for HST bogies. Residual networks are utilized to extract data features, which facilitates the bidirectional mapping among data, attributes, and features. Furthermore, the structure of the diffusion model is enhanced and customized for better adaptation to low-dimensional data, thereby improving the capability of the model to efficiently learn and generate latent features of unknown data, while maintaining stability. Finally, the model is established based on known data feature extraction and unknown data feature generation. Experimental results demonstrate that the average diagnostic accuracy for known fault classes exceeds 95%, while the average diagnostic accuracy for unknown fault classes surpasses 70%, with a harmonic mean diagnostic accuracy exceeding 80%. The results obtained significantly surpass those of other generative method-based diagnostic approaches, indicating that the study offers an effective solution for zero-shot learning fault diagnosis.