The onboard high-speed detection is a critical development direction for the operation and maintenance of railway networks. Rail corrugation is the most common track irregularity in metro systems, causing a series of environmental vibration and noise issues. Current onboard detection of rail corrugation primarily relies on axle box acceleration data and depends on data labeling and supervised learning frameworks. To facilitate data acquisition and reduce labeling costs, an unsupervised learning framework based on train body vertical vibration acceleration signals is proposed. First, the train body vertical vibration acceleration signals were transformed into time-frequency spectrogram datasets using Synchrosqueezed Wave Packet Transform (SSWPT). Then, an unsupervised learning framework was established based on momentum contrastive learning and trained on the unlabeled time-frequency spectrogram datasets. Finally, the pretrained model was fine-tuned for downstream tasks, including corrugation wavelength classification and amplitude assessment. Results indicate that the model achieved optimal performance after 430,000 iterations. For the binary classification of corrugation presence, the accuracy of the wavelength classification task ranged between 95% and 100%, while the amplitude assessment task achieved over 95% accuracy. Field application results demonstrate that the proposed model possesses strong generalization capabilities.