Fault diagnosis method for high-speed train axlebox bearings based on the joint driving of bearing dynamics simulation and experimental data

Abstract

The performance of traditional deep learning depends on sufficient labeled samples and the training and test samples obey a uniform data distribution. However, this assumption is unrealistic in high-speed train engineering, leading to unsatisfactory results in practical deployments. To enhance the practical applicability of deep learning, we propose a novel fault diagnosis method based on the joint driving of bearing dynamics simulation and experimental data to address this challenge. First, labeled bearing data is generated based on an axlebox-vehicle-track coupled dynamics model with clear physical meaning, and is further expanded using amplitude scaling to cover more working conditions. Then, to extract generalized fault features, the joint maximum mean discrepancy is introduced as a loss term in the domain adversarial neural network structure to achieve domain alignment. In addition, to avoid the interference of abnormal samples in the source domain, each source domain sample is adaptively assigned a weight through the discrimination difficulty of the domain discriminator. Finally, the proposed method is tested using the constructed target domain. Specifically, a multisource domain/single-source domain cross-device transfer task is created to evaluate the proposed method. The results demonstrate that expanding the source domain through dynamic simulation enhances the effectiveness of transfer learning and helps mitigate the diagnostic limitations caused by a shortage of real-vehicle fault data.