Abstract: A method is proposed for identifying wheel tread defects by using a multi-sensor data fusion algorithm and an improved convolutional neural network (CNN) to solve difficulty in fully characterizing wheel information and locating and quantifying wheel damage in trackside signals. A vehicle-track dynamics coupling model is applied based on multi-body dynamics and finite element theory. Considering fewer sensor arrangements, multimodal features were extracted, and data fusion algorithms were optimized for parameters such as wheel geometry and vehicle speed. An improved CNN model is proposed based on 1D-CNN and 2D-CNN. At the same time, frequency domain features and image features are fused. And the CNN algorithm model considering fusion features is proposed. Defect feature extraction is performed on the reconstructed signal, and improved CNN is used to fuse data features to achieve wheel damage identification. The effectiveness of the proposed method is verified using simulation data and actual examples combined with the proportional vehicle test rig. The performance of proposed method is compared and analyzed with CNN, BP neural network (BPNN), and support vector machine (SVM) under different signal test sets and data features. The results are shown that the damage identification model based on multi-sensor data fusion algorithm and improved CNN can better identify wheel tread defects, and the results are consistent with the measured results. After fusing data features from different dimensions, defects under different degrees of damage can be characterized and recognition performance can be improved. It can solve the problem of incomplete realization of wheel status based on trackside data, and provide technical support for online damage identification of wheel defects.