Abstract: It is difficult to consider both efficiency and precision in traditional methods to detect wheel out-of-roundness (OOR) of high-speed train. An intelligent identification method based on the axle box vibration and dynamic model is proposed. Firstly, the original data of wheel OOR is collected by a static detection device, and a data enhancement technique is proposed to construct the enhanced data set of wheel OOR. Then, the enhanced data set serves as input to the vehicle-track coupled dynamics model of high-speed train to obtain the vibration sample set of axle box under the excitation of different types of wheel OOR. Finally, the features of axle box vibration signal are adaptively extracted to identify and classify wheel OOR types using a 1-dimensional convolutional neural network (1-DCNN) with appropriate structure and configuration parameters. The results show that the proposed method can intelligently classify five types of wheel OOR, including normal, polygonal, scratched, stochastic non-roundness and local defect wheels. The accuracy rate is 99.2% (standard deviation of 0.05) and the average identification time of a single sample is 0.4 ms. Combined with the field test, the proposed method has a good recognition ability for the measured axle box vibration, and the test accuracy is 95%. Compared with the Support Vector Machine (SVM) and Back Propagation (BP) neural network methods, the 1-DCNN model has higher accuracy.

Key words: wheel out-of-roundness, vehicle-track coupled dynamics model, axle box vibration, data enhancement, 1-DCNN