A motor bearing fault diagnosis model based on an improved Bayesian network is proposed for the problem that the motor bearing vibration signal is affected by noise interference in feature extraction and low accuracy of traditional Bayesian network fault diagnosis. The adaptive ensemble modal decomposition of noise method is employed for noise reduction of data, which increases the model robustness; the Grasshopper Algorithm is optimized by using differential evolution and simulate anneal algorithm to enhance the global and local search ability of the Grasshopper Algorithm; the optimized Grasshopper Algorithm is applied to the Bayesian network structure and learning to construct the bearing fault diagnosis model; through comparing with other methods, it is proved that the method has stronger learning ability and higher accuracy rate for multi-fault classification of bearings, and the experiment fault diagnosis for some samples result reaches 97.15% and the average accuracy rate can reach 98.73%.