Generalized zero-shot diagnosis strategy for high-speed train bogie faced with a scarcity of specific fault samples

Abstract

In the field of high-speed train bogie fault diagnosis, the quality and quantity of monitoring data play pivotal roles in influencing the accuracy of diagnosis. It is a prevalent issue that there is often a scarcity of target fault samples during actual operation. In this paper, a challenging fault diagnosis task is addressed wherein none of the target domain fault samples are involved in the training process of the diagnosis model. Specifically, an attribute description strategy for faults is integrated into the Convolutional Recurrent Neural Network (CRNN). First, an attribute matrix of the bogie failure mode is defined manually. It is directly related to the semantic description of the fault modes involved. Subsequently, the attribute classifier is trained in parallel for each attribute to extract the attribute features corresponding to single faults and normal state. Following this, a GZSL-CRNN model based on generalized zero sample learning (GZSL) is constructed using the obtainable attribute matrix of unknown composite faults. Finally, the efficiency of the method is validated on the SIMPACK dynamic model. The results demonstrate that the method not only ensures the high accuracy of diagnosing known single faults but also efficiently identifies unknown composite faults.