On-board identification of wheel polygonization of metro trains based on convolutional neural network regression analysis and angular-domain synchronous averaging
Wenjing Sun, Xuan Geng, David Thompson, Tengfei Wang, Jinsong Zhou, Jin Zhang
Abstract
Wheel polygonization, a form of wheel out-of-roundness, has become a common problem on trains of urban rail transit systems in recent years. It results in a significant increase of the dynamic responses of both the vehicle and the track, high vibration and noise levels, and structural fatigue. This paper proposes an innovative method for identifying wheel polygonization orders and their effective values using convolutional neural network (CNN) regression analysis. First, the acceleration signal measured on the axle box has been processed with the angular-domain synchronous averaging (ADSA) method, effectively separating the characteristic information associated with wheel polygonization within the signal. To extract comprehensive wheel polygonization information, a feature fusion method is employed, integrating features from both the time and frequency domain. Then, a CNN regression model is established and trained, with validation conducted using measured data of vehicle vibration and the wheel polygonization measured during field tests. Comparative analysis with different identification methods is performed, including a comparison of different preprocessing methods and machine learning models, which demonstrates the effectiveness of the proposed method in this study. The verification results show that the proposed method achieves high identification accuracy for wheel polygonization up to the 25th order. The overall average root mean square error value is 2.0 dB. Finally, the influence of wheel polygonization conditions, track stiffness, and speed fluctuation on the identification accuracy is discussed. The results show the proposed method exhibits robust identification capacity under varying conditions, which indicates its wide application and accuracy in complex situations during train service. This research contributes to advancing the field of wheel polygonization detection, offering a reliable and effective solution for application in railway systems.