A fault diagnosis of a train door system is carried out using the motor current signal that operates the door. A test rig is prepared, in which various fault modes are examined by applying extreme conditions, as well as the natural and artificial wears of critical components. Two approaches are undertaken toward the fault classification for comparative purposes: one is the traditional feature-based method that requires several steps for the processing features such as signal segmentation, the extraction of time-domain features, selection by Fisher’s discrimination, and K-nearest neighbor. The other is the deep learning approach by employing the convolutional neural network (CNN) to skip the hand-crafted features extraction process. In the traditional approach, good accuracy is found only after the current signal is segmented into the three velocity regimes, which enhances the discrimination capability. In the CNN, superior accuracy is obtained even by the original raw signal, which is more convenient in terms of implementation. However, in view of practical applications, the traditional approach is more useful in that the features processing can be easily applied to assess the health state of each fault and monitor the progression over time in the real operation, which is not enabled by the deep learning approach.