The Recognition of Rail Surface State Based on Improved ResNet-50 Deep Learning Network

The traction/braking performance of rail transit vehicles depends on the wheel rail contact condition. However the running performance of vehicles is quite different under different rail surface conditions, such as dry, wet and greasy. In view of the problems of large subjectivity and serious lag in traditional artificial experience based rail surface state recognition, a rail surface state recognition method based on an improved ResNet-50 deep learning network is proposed. Firstly, the ResNet-50 network is used to build the rail surface state recognition model. Secondly, transfer learning is introduced to improve the structure and parameters of the ResNet-50 network, and the improved ResNet-50 network is used to train the image data of the rail surface state. Finally, a model that can be used for classification is obtained to recognize the rail surface state. The results show that the optimized ResNet-50 model is more effective in identifying the rail surface state compared with the traditional ResNet-50 model. The model performance is better, and the identification accuracy can reach 92.75%.