Transfer learning, as an effective technique to address distributional differences between domains, has received increasing interest in the field of fault diagnosis in recent years. However, the existing rotating machinery fault diagnosis methods usually fail to adequately consider the impact of different samples on the diagnostic results during the transfer learning process. In addition, the traditional edge distribution alignment method is not effective enough in reducing the distribution differences between source and target domains data, which largely limits the practical effectiveness of transfer learning methods. Aiming at the above problems, a rolling bearing fault diagnosis method based on dynamic calibration and joint distribution alignment is proposed. Firstly, the dynamically calibrated residual network (DCRN) is constructed as the feature extraction layer, which enhances the feature expression capability of the network by designing a dynamic calibration structure and adjusting the weights according to different samples. Secondly, the domain adaptive layer is designed and a new joint distribution alignment mechanism (JDAM) is proposed. This mechanism gives full consideration to the edge distribution differences and condition distribution differences between the data of the source and target domains during feature alignment, enabling the effective transfer of knowledge learned in the source domain to the target domain and significantly improving the performance of the target task. Finally, the I-Softmax function is used to optimize the classifier, allowing the network to better identify the faults in different states. Experimental validation was given using the Case Western Reserve University bearing dataset, the MFS Bearing dataset and the roller gear dataset. Under cross-domain and variable noise conditions, the proposed method achieved average accuracies of 96.50%, 96.87%, and 94.72%, respectively, demonstrating high fault diagnosis accuracy and good generalization capability.