Door control system is one of the most important subsystems in the subway vehicle. Due to the complex mechatronic structure, frequent open and close movement, and crowded passenger flow environment, high failure rate of door system persists. To accurately detect the incipient fault, a bigdatadriven optimal feature selection algorithm and a random forests (RF) based incipient fault diagnosis method are proposed in this paper. Firstly, multiphase timedomain fault features are extracted from door′s position, drivenmotor′s speed and current signals. Secondly, the irrelevant and redundant features are removed and the optimal fault features are retained by using distance evaluation technology. The selected optimal fault features are adopted as the input of RF classifier. The fault labels are utilized to formulate an intelligent fault diagnosis model. Finally, the fault diagnosis model can realize the online automatic recognition of different incipient faults in the door subsystem. Experiments are conducted on the bench testing door system of Hangzhou line 4. Results show that the proposed method can extract the early features of incipient faults. Compared with several existing methods, the diagnostic accuracy and robustness of the proposed method are greatly improved after optimal feature selection.