Abstract: Because of the insufficient use of physical monitor-ing information, the fault diagnosis of EMU (electric multiple unit) axle box bearing has a low accuracy rate problem. First, highspeed EMU axle box bearing test bench is used to obtain ample data, integrate temperature and vibration feature data, and the PCA (principal component analysis) method is used for fusion and dimension reduction. Then, the axle box bearing fault diagnosis model based on temperature-vibration fusion and DAE (deep autoencoder) is established, and the model was trained by the DAE. Finally, the model is verified by the test set data obtained from the high-speed EMU axle box bearing test bench. Verification results show that, compared with other comparative models, the axle box bearing fault diagnosis model based on temperature-vibration fusion and DAE has higher accuracy rate.