Generalized Zero-Shot Learning Fault Diagnosis for Bogies of High-speed Trains based on Improved Generative Adversarial Networks

In recent years, data-driven methods have developed rapidly in the realm of fault diagnosis, but in the face of engineering fault types without historical records, there are still problems such as inadequate learning of fault features and high misdiagnosis rate. To address the issue of limited generalization ability of fault diagnosis models owing to the lack of fault data types in high-speed train (HST) bogies, this paper proposes a generalized zero-shot learning (GZSL) fault diagnosis method for bogies based on generative adversarial networks. Firstly, the conditional Wassertein generative adversarial network using cosine distance (CCDWGAN) proposed in this paper is trained by known fault types, attributes and data. Then, all the possible unknown fault type data are generated by the trained CCDWGAN with attributes of unknown faults. Finally, the generated data of unknown fault types are trained together with known types fault data through a deep Residual Network (ResNet) to build a fault diagnosis model of bogies, and realize the GZSL fault diagnosis of unknown and known fault types. The experiments show that the GZSL based on adversarial generative network can effectively generate unknown fault types, and the fault diagnosis rate of the proposed framework can reach more than 95% for the known category and 70% for the unknown category, which provides a practical solution for the GZSL problem of train bogies fault.