A Few-Shot Fault Diagnosis Method for Train Transmission Systems Based on Spatial Median Prototypical Networks

Intelligent fault diagnosis methods based on few-shot learning are able to diagnose train transmission systems with scarce training samples accurately. However, the training samples may be mislabeled in engineering scenes, and the existing few-shot fault diagnosis methods lack robust learning strategies for partially mislabeled training samples, resulting in degradation of diagnosis accuracy. To address this issue, this paper proposes a few-shot fault diagnosis method for train transmission systems based on spatial median prototypical networks. Firstly, according to prototypical networks, a prototype aggregation with mislabeling tolerance strategy is adopted, where the prototype deviation caused by the mislabeled support samples is alleviated by calculating the spatial median of the fault features instead of the mean of fault features in the prototype aggregation process. After that, a mislabeling suppression loss function is introduced to reduce the interference of mislabeled query samples on parameter updating by adopting a novel definition of the loss function. Finally, fault simulation experiments of subway train transmission systems are designed and conducted, and the experimental results show that even if the mislabeled samples are mixed in the limited training datasets, the proposed method is still able to diagnose the faults of train transmission systems accurately, and the diagnosis performance is better than that of the existing methods.