Fault diagnosis of suspension system of high-speed train based on model-agnostic meta-learning

Aim at the problem of lack of samples in machine learning-based fault diagnosis of suspension system of high-speed train, this study introduces the model-agnostic meta-learning (MAML) algorithm to train the two dimension (2D) convolutional neural network (CNN). A sample reconstruction method is proposed to convert the raw vibration signals of the suspension system into feature matrices containing more fault information, and the feature matrices are used as the training samples of 2D CNN. The results show that the 2D CNN achieve the fault diagnosis accuracy of exceeding 90.41% with one training sample. It means that this study has important potential for real-time fault diagnosis of suspension system under few-shot condition.