PDC-SWM-LPF: A novel feature extraction method for on-board detection of wheel polygonal wear

Abstract

Wheel polygonal wear often leads to a decrease in vehicle ride comfort, causes fractures in vehicle components, and can result in serious accidents. Therefore, monitoring the fault state of polygonal wheel in real time is particularly important. Currently, quantitative analysis of polygonal wheels mainly adopts a big data-driven approach. However, in practical engineering scenarios, there is not always an abundance of data, posing challenges for the amplitude recognition of polygonal wheels. This paper employs an Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) to determine the order of polygonal wheel wear. On this basis, a feature extraction method (SWM-LPF) for the probability density curve (PDC) vibration acceleration of axle box was proposed, enabling the generalized regression neural network (GRNN) to predict the amplitude of polygonal wheel with limited sample support. In ablation experiments, Residual Neural Network (ResNet) and 1D-CNN were used as the baseline models. The experiments show that using only 16.7% of the training samples required by the baseline models, SWM-LPF enabled GRNN to achieve comparable accuracy in estimating the amplitude of polygonal wheel. With the assistance of SWM-LPF, the iteration rounds of ResNet and 1D-CNN were reduced by 80%, significantly improving the accuracy of amplitude detection for polygonal wheel.