A Bearing Fault Diagnosis Method with Unsupervised Deep Adaptive Network

To improve the ability of the predictive model to generalize unlabeled data in fault diagnosis, an improved bearing fault diagnosis method with unsupervised deep adaptive network (UDAN) based on second-order statistics is presented. First, the motor vibration signal is transformed into a two-dimensional gray image to improve the extraction of transfer features. Then, the second-order statistics alignment of source domain and target domain is used to minimize the difference in domain distribution in the deep residual network. Finally, the combined loss function is constructed to realize the end to end adaptive fault diagnosis. Compared to other methods of unsupervised learning, experimental results show that UDAN fault diagnosis method has better generalization ability.