An improved convolutional neural network for convenient rail damage detection

The long-term operation of a railroad usually leads to defects in its rails, axles, fasteners, etc. These problems directly affect the safety of the rail system. Therefore, it is important to ensure the health of key railroad structures. In this paper, a deep learning-based rail damage identification method is established by analyzing the rail vibration signals collected with piezoelectric ceramic pads. The multiple features of vibration signals are combined and then reconstructed into grayscale maps as the inputs of the model. The key information of the grayscale maps is extracted using neural networks. The idea of pre-convolution is used to solve the problem that the model pays more attention to certain features due to the different input sizes and the implied weights of the features. Finally, the performance of the three convolutional neural networks (CNN) in rail damage detection is evaluated and compared. The results show that the CNN with pre-convolution and Residual structure has better recognition for the presence of rail damage than other methods.