Unknown Compound Faults Diagnosis of High Speed Train Based on Capsule Network

Convolutional neural networks (CNN) have the ability of self-adaptive learning features, which provides new ideas for fault diagnosis and analysis in the field of high-speed trains(HST). Combined with deep learning and wavelet transform, a diagnostic model for unknown compound faults based on capsule network is proposed. It is used to solve the problems of nonlinear of vibration signals and the difficulty of diagnosing unknown compound faults. Firstly, the collected vibration signal is converted into a spectrum map suitable for the network size and directly input into the convolution network layer for feature learning, which avoids the shortage of information loss caused by manual extraction of features. Secondly, the basic features detected by the convolutional layer are input into the capsule layer for combination and packaging of features. Finally, the fault condition is identified by the trained classifier. Experiments on different data sets collected in the laboratory simulation show that the diagnostic rate of this method for unknown compound faults is 90.31%, increasing by 7.94% to compared with the existing methods. Experiments were carried out using different types of unknown compound faults, and the generalization ability and robustness of the proposed model were verified.