MiDAN: A framework for cross-domain intelligent fault diagnosis with imbalanced datasets

Abstract The success of deep learning techniques for intelligent fault diagnosis relies on two explicit assumptions: (1) the training and testing data are drawn from the same distribution; (2) the data are class-wise balanced. However, those assumptions often prove to be a significant obstacle for industrial applications. This paper presents a framework named deep mixup domain adaptation network (MiDAN) to tackle the distribution mismatches and data imbalance simultaneously. Specifically, the rebalanced mixup training (ReMix) is proposed associated with domain adversarial training, which guarantees a more reliable distribution matching based on a data-agnostic manner. Furthermore, a strong–weak learning framework is adopted to automatically learn representative features and directly explore the hidden information shared by the source and the target domains. Validation experiments are performed based on the cross-domain rolling bearing fault diagnosis tasks. As reported, the proposed method achieves superior diagnosis performance comparing to the state-of-the-art deep learning methods.