Imbalanced Fault Diagnosis of Key Components in Spacecraft ACS via Collaborative ResNet

In this paper, a fault diagnosis method based on Collaborative Residual Network (Collaborative ResNet) is proposed to address the problem of imbalanced fault diagnosis for key components of spacecraft attitude control system (ACS). The designed Collaborative ResNet consists of two branches, the conventional learning branch and rebalancing learning branch, which conduct representation learning and classifier learning, respectively. A new Dynamic Complementary Loss (DC Loss) is introduced in the rebalancing learning to dynamically alleviate the excessive gradient of negative samples on the few sample classes and compensate for the gradient of misdiagnosed samples, which enhances the network sensitivity to the class with few number samples. Then the features extracted by two branches are adaptively integrated through the collaborative learning strategy to combine the strengths of both approaches. The effectiveness of the proposed method was validated on the imbalanced data collected on a ground semi-physical platform.