Deep imbalanced adversarial domain adaptation network for cross domain fault diagnosis

Abstract With the booming development of artificial intelligence and industrial big data technology, unsupervised domain adaptation methods have achieved remarkable results in the field of cross-domain fault diagnosis of rotating equipment. However, there is significant difficulty in obtaining fault data from the equipment, which leads to the model’s predictions being biased toward the majority class. To this end, this paper proposes the deep imbalance adversarial domain adaptation network to cope with the above challenges. A cost-sensitive weighting function is designed to suppress majority-class overfitting and rectify the source-domain decision boundary. With the source-domain decision boundary properly adjusted, the task classifier is reused as a domain discriminator through class-level distribution alignment, achieving simultaneous class discrimination and domain adaptation without additional network branches. Complementing this alignment strategy, a minimum class confusion mechanism is applied to the target domain to prevent prediction collapse toward majority classes. Together, these strategies form an integrated solution for imbalanced cross-domain fault diagnosis by sequentially addressing boundary correction, domain-invariant feature learning and target-domain bias mitigation. Extensive experiments on two imbalanced datasets demonstrate the effectiveness and superiority of the proposed method.