Currently, classification methods based on deep learning are developing rapidly in the field of fault diagnosis. The excellent performance is due to the superior feature extraction capability and extremely high classification accuracy, but this requires a mass of training data. However, in practical industrial applications, bearings are mostly operating normally, the real fault data are difficult to obtain. The unbalanced distribution of data will lead to a serious decline in the performance of the deep model. Considering the actual application situation, a rolling bearing fault diagnosis based on digital twin method is proposed in this paper. Firstly, a parametric dynamic twin model of the bearing is constructed based on the generation principle of the bearing vibration signal. Secondly, combined with classical parameter optimization and deep learning methods, multiple corrections of simulated signals are realized at the model aspect and signal aspect, which improves the matching degree with real signals. The training requirement of existing fault diagnosis method is meet. Finally, a fault diagnosis method based on transfer learning is proposed to reduce the difference of feature distribution between the twin samples and the real samples. The proposed method is verified by using the public data set and compared with other classical methods. The results show that the proposed method can alleviate the problem of unbalanced data and perform well in the classification problems.