The fault diagnosis system is an essential component of the satellite, which can ensure the normal operation of the satellite and extends its lifespan. With advancements in deep learning techniques and computational capabilities, data-driven fault diagnosis methods have become more precise and practical, making it less challenging to design fault diagnosis systems. However, existing data-driven fault diagnosis methods are infeasible for satellites lacking sufficient labeled data. Hence, a weighted conditional adversarial domain confusion network is introduced to address this issue. The proposed method employs the maximum mean discrepancy distance for constraining feature distributions of operational data collected by different satellites. Adversarial learning is also utilized for matching the joint distributions, which helps to transfer the fault diagnosis model of the source satellite to the target satellite. The proposed method is compared with other transfer learning algorithms, and simulation results and ablation studies are presented to further validate the effectiveness.