Fault diagnosis of aerospace electromechanical systems in orbit faces the dual challenges of data scarcity and domain discrepancy. To address this problem, a contrastive learning enhanced transfer framework is proposed. Large-scale general bearing data are employed for self-supervised pretraining to obtain transferable representations, which are then adapted with a limited number of labeled CMG samples. The model is subsequently validated on in-orbit CMG data under real operating conditions. Experimental results demonstrate that the proposed method substantially improves diagnostic performance compared with traditional approaches, achieving a relative gain of about 24.1%. These findings highlight the effectiveness of integrating contrastive learning with transfer learning to overcome small-sample and cross-domain challenges, and provide a feasible pathway for in-orbit health management of aerospace electromechanical systems.

contrastive learning,transfer learning,fault diagnosis,aerospace electromechanical systems