The main problems faced by deep migration learning in the field of fault diagnosis today are: inconsistent data distribution between the source and target domains and unbalanced fault data. The domain confusion method only uses the maximum mean difference (MMD) to measure the difference between the features of the source domain and the target domain, which cannot accurately measure the difference in features, and the problem of data imbalance leads to the risk of overfitting in training. In this research, the domain adversarial neural network (DANN), the open set domain adaptation by backpropagation algorithm (OSDABP), and the deep domain confusion (DDC) algorithm are studied. Based on the deep domain confusion neural network, we have made the following improvements: 1. The combination of the maximum mean difference and the correlation coefficient (CORAL) was used to measure the discrepancy in the feature distribution. 2. Use the early stop mechanism to monitor the change of the validation set accuracy, and trigger the early stop when the accuracy of the validation set is no longer improved to prevent fluctuations caused by overtraining. 3. Added label smoothing mechanism to reduce the risk of overfitting by smoothing hard labels. Based on the bearing dataset of Case Western Reserve University (CWRU), the proposed method is verified to have better performance and smoother accuracy curves through six cross-load diagnostic tasks, compared with other migration methods.