A small sample diagnosis method driven by simulation and test data: applied to axle box bearings of high-speed train

Abstract

Abstract Fault diagnosis of axle box bearings is an important technology to improve the service safety and economy of high-speed trains. However, it is difficult to obtain sufficient fault samples in actual train operation, which limits the application of deep learning methods in the field of high-speed train fault diagnosis. Therefore, in this paper, a simulation and test data fusion-driven fault diagnosis method for axle box bearings with few samples is proposed. In this method, a bearing dynamics model is constructed to obtain simulation data for fault bearings, and the accuracy of the constructed dynamics model is verified by experimental data. Data fusion methods are proposed which use massive simulated fault samples and only normal test samples to construct new samples for the pre-training of deep learning diagnostic models. The parameter transfer strategy is used to fine-tune the diagnostic model with a small amount of test fault samples. A real train experiment is carried out for validating the proposed methods which use the real fault bearings that fail during operation. The results show that the proposed method via the fusion of simulation and test data has higher generalization ability and diagnostic accuracy in small sample situations.