CNN-based Fault Diagnosis of High-speed Train with Imbalance Data: A Comparison Study

High-speed train bogie, the only component connecting the train body and track, its degradation and fault would directly threaten the safety of the vehicle. However, learning-based fault diagnosis methods are faced with imbalanced between normal samples and fault samples, which would lead to poor diagnosis performance. This paper provides a fault diagnosis architecture for high-speed train based on convolutional neural network, and critical comparison between three representative class balancing techniques, including weighted loss, focal loss, and synthetic minority over-sampling technique. The innovation of this study is concerning the judiciously chosen class balancing methods for neutral-network-based fault diagnosis of high-speed train. Based on the experiment results of this comparison study, it is found that class balancing method can significantly improve the performance of the developed diagnosis model, and synthetic minority over-sampling technique is more effective than two other approaches. This study is valuable for the further research and practical applications of fault diagnosis.