A machine learning method combined multi-sensors data fusion with improved convolutional neural network for wheel defect identification

A novel method for identifying wheel tread defects is proposed, combining multi-sensor data fusion with an improved convolutional neural network (CNN). This approach addresses challenges in fully characterizing wheel information and precisely quantifying damage from trackside signals. A vehicle-track dynamics model is established using multi-body dynamics and finite element theory. The method optimized sensor arrangements, extracted multi-modal features and refined by data fusion algorithms. An improved CNN model, incorporating 1D and 2D features, enhances defect identification precision. The fused features enable comprehensive damage characterization. The effectiveness of this method has been verified through simulation and scaled test bench experiments, and it outperforms traditional CNN, backpropagation neural network (BPNN), and support vector machine (SVM) models. By integrating data features from different dimensions, the identification performance is enhanced, enabling the characterization of defects with varying degrees of damage. This method effectively solves the limitation of incomplete realization of wheel status from trackside data and provides strong technical support for online identification of wheel defects.