<p>The real-time monitoring data of bearings presents non-stationarity, non-linearity, and low-value density characteristics, and the state data collected under different operating conditions have different distribution characteristics. However, simulation data is difficult to accurately simulate the type and degree of motor failure, resulting in a serious imbalance between normal data and fault data samples, which can lead to problems such as overfitting, low accuracy, and poor generalization ability in model training for fault diagnosis. A variable operating condition fault diagnosis method based on Condition Domain Adversarial Neural Networks—Joint Maximum Mean Discrepancy (CDANN-JMMD) is proposed. First, CDANN is used to map the data of the source domain and target domain under varying conditions into the same feature space. The adversarial strategy is used to confuse the source domain and target domain, and to pull the edge distributions of different domains closer together; then, JMMD is used to further constrain the boundary distribution and conditional distribution of different categories within the domain, improving the stability of the accuracy of variable-condition fault diagnosis and achieving variable-condition migration fault diagnosis; finally, using the CWRU dataset and the Jiangnan University dataset, we conducted verification experiments for variable operating conditions and variable bearing fault diagnosis, respectively. The experimental results show that the proposed method effectively reduces the feature distribution difference between the source domain and the target domain, and the classification accuracy of fault diagnosis for different transfer task models is above 99.5%, which verifies the model’s good generalization and robustness in variable operating condition fault diagnosis.</p>