Abstract: To improve the maintenance efficiency of ZDJ9 type switch machine in urban rail transit, this paper proposed a fault detection method for switch machines based on Back Propagation (BP) neural networks. The paper deeply analyzed the principle of collecting action current of the switch machine and the characteristics of current curves at different stages during the on-site switch machine conversion process, determined the type of fault current curve, performed wavelet decomposition and reconstruction on the action current curve during the switch machine conversion process, extracted key feature values from the reconstructed curve, and used them as training data for the fault detection model based on BP neural network. Finally, after 8 000 iterations of training, the fault detection accuracy of the model reaches 96%, it indicates that this method can effectively detect switch machine faults and their types.