基于卷积神经网络的高铁车轮损伤识别方法研究

In this paper, the measured vibration data during the train operations were analyzed, a specific Convolutional Neural Network (CNN) structure for wheel damage detection was designed, network training was carried out based on vibration data with different labels for damage feature extraction, then the well trained CNN was used to realize long-term and efficient damage detection for high-speed rail wheel. Due to the black box characteristics of CNN, it is difficult to understand its mechanism for damage detection. In this paper, the convolution kernel of the CNN was visualized by using the gradient ascent method, and the response of the convolution kernel was analyzed by using the cross-power spectrum density to seek the physical explanation of feature extraction. Finally, the CNN damage detection results were compared with the results of Deep Neural Network and Recurrent Neural Network. It is concluded that CNN has a high accuracy for damage detection, with an average accuracy of 99.40%, and this accuracy is about 6% higher than other methods. In addition, the number of training parameters for CNN is smaller, so CNN is a light and efficient damage detection model for high-speed rail wheels.