Abstract
Ensuring the stable operation of space power systems is a critical challenge for long-duration missions. Remote operation and communication latency hinder timely fault detection and make human expert intervention more difficult, while limited historical data restricts the generalizability of machine learning models. This study proposes a human-in-the-loop machine learning diagnostic framework based on a Transformer backbone, with fusion mechanisms at both the feature and decision levels to integrate expert knowledge and data-driven learning. To further characterize the interaction between human and machine intelligence, a multi-stage game-theoretic method is introduced to quantify their respective contributions in the decision-making process. Experiments on the TOPAZ-II power system, involving 11205 noisy samples across 13 fault types, show that the proposed framework improves diagnostic accuracy and robustness compared to both data-driven and human-only approaches. Under 10% data noise and 30% expert error, it achieves 98.19% precision and 98.22% recall. Additional evaluation on the MVTec benchmark demonstrates its superiority over existing models. The framework provides a scalable, interpretable, and effective solution for fault diagnosis in data-limited, high-risk environments.