Single-source domain adaptation fault diagnosis methods often suffer from domain mismatch problems, resulting in negative transfer and insufficient generalization capabilities. At the same time, actual industry often contains data from multiple source domains, and the information contained in the target domain varies greatly in different source domains. Therefore, the fault diagnosis method based on joint alignment of multiple source domain adaptation was proposed. First, in the face of multi-sensor signals, the average splicing fusion method was used to form the fusion signal. Second, the multi-scale feature extraction module with transferable residual module was proposed to ensure multi-scale feature extraction and enhance the non-extra parameterized transferability of the model. Finally, adaptive hyperparameters and multi-kernel maximum mean discrepancies were combined as constraints to eliminate the differences in data distribution in the network layer. The transferable residual module as a structural optimization strategy and multi-kernel maximum mean discrepancies as a statistical transformation strategy were jointly applied, which was called joint alignment. Experimental results show that the entire model can achieve high-accuracy fault diagnosis requirements in multi-source domains without introducing redundant hyperparameters.