Traditional track geometry fault diagnosis methods, such as track geometry car, track detection trolley, and manual measurement, suffer from low detection frequency. Therefore, using vibration data from in-service metro vehicles to diagnose metro track geometry faults can improve the efficiency of metro track condition detection. Hence, a metro track geometry fault diagnosis model using car-body vibration was established in this study based on a convolutional neural network model, and the effects of the structural hyperparameters on the model performance were analysed. Firstly, the problem of car-body vibration data collection was addressed, and a mileage-locating model without the global navigation satellite system was developed. Then, the metro track geometry fault diagnosis model was derived using a convolutional neural network whose architecture is determined by four hyperparameters. Finally, a case study was performed using the Beijing Subway Line 1 track geometry car detection data and car-body vibration data. The results demonstrated the effectiveness of the proposed model.