The problem of long time and low accuracy of the generator bearing fault diagnosis model was considered, and a fault diagnosis method based on improved naive Bayesian classification was proposed. Taking a subway axial flow fan bearing vibration data as the foundation, the feature set of the original fault and the fault feature set after reducing the dimension were constructed separately by using the method of wavelet packet transform, rough set and principal component analysis. The original fault feature set and the dimensionality reduction feature set were input into the simple Bias classification model. On this basis, the fault diagnosis system of alternator bearing was designed. Finally, the improved Bayesian classification method, the neural network and least squares support vector machine method were compared and analyzed based on the vibration data of bearing of a metro vehicle. The results show the modeling time is shorter and the accuracy of fault diagnosis is higher based on the improved Bayesian classification method.