In order to solve the problem that the single feature information extraction is insufficient in the field of switch machine fault diagnosis, and the single diagnosis method is difficult to avoid the classification error caused by the limitation of the method. At the same time, it has a certain degree of over-fitting, as well as insufficient generalization ability and robustness. A fault diagnosis method based on spatiotemporal feature combination model was proposed. Firstly, the time-frequency domain wavelet coefficients were extracted from the oil pressure curves corresponding to the eight fault modes and normal modes of the ZYJ7 electro-hydraulic switch machine as the original data set. The kernel principal component analysis (KPCA) and long short-term memory network (LSTM) were used to extract the spatial and temporal features, and then the spatiotemporal feature set was constructed based on the addition (ADD) idea. Secondly, after optimizing the key parameters of convolutional neural network (CNN) and LSTM, the fault diagnosis was carried out respectively, and the probability value and error coefficient of each fault type were obtained. Finally, the error reciprocal method was used to assign weights to the probability values of each fault type of the two classifiers, and the final output result was obtained. The simulation results show that the diagnostic accuracy of CNN-LSTM combined model is 98.14%, which is 7.40%, 5.55% and 1.85% higher than that of single multi-layer perceptron (MLP), CNN and LSTM models, respectively. It can be seen that this method effectively improves the accuracy of switch machine diagnosis, and provides an idea for the application of ensemble learning model in the field of switch machine fault diagnosis.