To solve the problem that the fault data and sample labels from the actual working conditions of rolling bearings are scarce, and the distribution of different fields is very different. A fault diagnosis method of rolling bearing based on multi-source domain transfer learning based on wavelet information initialization and double-path convolution is proposed. In this method, labeled samples under various working conditions are selected as multi-source domains. Firstly, the optimized wavelet weights are used to initialize the first-layer CNN weights of the general feature extraction module to enhance the reliability and robustness of the data-driven model from the perspective of signal processing. Then, two-path convolution is used as a specific feature extraction module to fully learn the feature information of the data in multi-source and target domains. At the same time, domain counter loss, multi-core maximum mean difference (MK-MMD) and improved local maximum mean difference (ILMMD) were set at different nodes of the migration model, and a novel multi-source domain adaptive strategy was constructed to further reduce the feature distribution differences between multiple source domains and between source domains and target domains, enhance domain alignment, and facilitate the network to extract domain invariant features. Finally, the proposed method is validated by analyzing bearing fault transfer learning tasks under different speeds and loads on actual rolling bearing data sets. The results show that the proposed method is superior to the comparison method in multiple tasks, which indicates the effectiveness and superiority of the proposed method.