Addressing the challenge of the current inability to qualitatively identify rail corrugation damage accurately from axle box acceleration data, this study proposes a novel approach. To indirectly identify rail corrugation from axle box acceleration, we introduce an improved successive variational mode decomposition (SVMD) algorithm, coupled with a deep learning model for corrugation recognition. First, a numerical model of the vehicle-rail-track slab system is established, considering rail corrugation. Indicator analysis is integrated into the SVMD. The improved SVMD is employed to decompose and reconstruct axle box acceleration, achieving noise reduction and extraction of useful components. Next, we apply the fast Continuous Wavelet Transform (fCWT) analysis method to effectively transform one-dimensional data into two-dimensional images. Finally, the You Only Look Once (YOLO) model serves as a classifier for the classification and recognition of corrugation with different wavelengths. The results demonstrate that the mechanism-driven improved SVMD effectively extracts corrugation components from axle box acceleration, while the YOLO model achieves rapid and efficient identification and classification of corrugation with different wavelengths. The results show that compared with other traditional models, the training time of the YOLO model is 60%-90% of the training time of the traditional algorithm, and the Recall rate is 1.15-1.45 times that of the traditional algorithm. In terms of wavelength identification, the YOLO model has a recognition rate of 98% for different wavelengths. The proposed approach offers an innovative and efficient solution for identifying rail corrugation damage in axle box acceleration data.