The bearing monitoring data under different working conditions have different distributions, which affect the accuracy of bearing intelligent diagnosis methods. Aiming at this issue, a multi-domain deep adversarial transfer network is proposed for intelligent diagnosis of bearings based on deep learning and transfer learning. The proposed method treats the datasets under different working conditions as belonging to different domains. In the method, firstly, the deep residual network is used to extract the features from source training data and target testing data. Then, multiple domain adversarial modules are designed to capture multi-mode structures from the source and target domain data, which enable fine-grained alignment of the distribution of the source domain and target domain data. Finally, the fault classifier is trained to transfer fault diagnosis knowledge of source domain to the target domain and ensure the diagnosis accuracy under variable working conditions. The proposed method is verified by using the two bearings dataset under variable working conditions. The results show that the proposed method can capture the multi-mode structures of the bearing data and extract domain invariant features, thus improving the fault diagnosis accuracy under variable working conditions compared with other related methods.