Abstract: A time-domain analysis of the motor current of metro sliding plug door is carried out in three states of normal, lower gear pin failure and pressure wheel faults. It is found that the motor currents in the three states have obvious differences in time-domain distribution after preprocessing. On this basis, the motor current in three states is extracted and appropriate time-domain feature parameters are screened, which are then combined with BP (back propagation) neural network. Thus a BP neural network model with Dropout optimization for fault detection of metro sliding plug doors is proposed, and the detection of lower gear pin and pressure wheel faults of the metro sliding plug door is implemented. Test results based on actual case data show that the model can reduce the occurrence of overfitting situation to certain extent and effectively detect the lower gear pin and pressure wheel faults of plug door with high accuracy.