Abstract: To address the challenges of fully characterizing wheel information and accurately locating and quantifying wheel damage using trackside signals, this paper proposes a multi-sensor data fusion algorithm combined with an improved convolutional neural network (CNN) for wheel tread defect identification. A vehicle-track dynamics coupling model is established based on multi-body dynamics and finite element theory. By strategically arranging fewer sensors, multimodal features are extracted, and data fusion algorithms are optimized for parameters like wheel geometry and vehicle speed. An improved CNN model is then proposed, building upon both 1D-CNN and 2D-CNN architectures. Simutaneously, frequency domain features and image features are fused, leading to a new CNN algorithm model that incorporates these fusion features. Defect feature extraction is performed on the reconstructed signal, and the improved CNN, leveraging the fused data features, is used to achieve wheel damage identification. The effectiveness of the proposed method is validated using both simulation data and actual case studies, in conjunction with a proportional vehicle test rig. The damage identification performance of the proposed model is compared against CNN, BP neural network (BPNN), and support vector machine (SVM) under various signal test sets and data features. Results indicate that the proposed damage identification model can more effectively identify wheel tread defects, showing good consistency with measured results. Fusing data features from different dimensions can characterize defects under varying degrees of damage and significantly improve identification performance. This approach successfully addresses issues where trackside data alone cannot fully reconstruct wheel status, thereby providing crucial technical support for the online damage identification of wheel defects.