Detection of Railway Wheel Flat Based on CBAM-Enhanced ResNet for Imbalanced Data

Abstract

Wheel flat is a common fault during train operation, which seriously affects running safety. Deep learning flat detection methods can learn and identify flats automatically without relying on expert experience, which has attracted widespread attention. However, accurately detecting wheel flats remains challenging due to the strong interference signal components and the data imbalance from the lack of failure data. In this article, the convolutional block attention module-enhanced residual net (CBAM-enhanced ResNet) model is adopted for flat detection tasks to improve the robustness and the recognition ability of the model. To detect wheel flat for imbalanced data, a dataset expansion method based on wheel-rail dynamics simulation is proposed. In this method, the effects of wheel-flat lengths and the impact positions on flat signals were studied based on the developed vehicle-track coupled model. Then, new flat signals can be reconstructed by transforming the actual flat signal according to the obtained fitting relationships. Experiments were conducted to verify the effectiveness of the CBAM-enhanced ResNet model and the proposed dataset expansion method. The results show that the CBAM-enhanced ResNet model achieves better flat detection results than the ResNet model. After data expansion, the accuracy of both models was improved.