关键词: 在轨服务, Lambert转移, 重规划, 优化, 集束搜索, 注意力模型

Abstract: To meet the requirements of large-scale on-orbit service mission planning, existing research is usually confined to static scenarios, making it difficult to tackle the invalidation of initial plans caused by unexpected events such as spacecraft failures and task changes. This calls for on-the-fly replanning with flexible Lambert transfers. Meanwhile, this mission is a mixed-integer programming problem involving three nested optimization layers: trajectory planning, sequence planning, and task assignment. Current strategies involve extensive iterative computations, especially for Lambert transfers that require numerical iterative solutions, which severely limits their practical engineering applications. To address this challenge, this paper proposes an efficient algorithmic framework: first, for the underlying trajectory planning and sequence planning problems, deep neural networks and attention models are adopted respectively to replace the online optimization iteration process, enabling fast solutions. Second, for the upper-layer task assignment problem, a Beam Pre-search and Neighborhood Optimization Algorithm (BPNOA) based on historical pre-plans is proposed, which enables efficient solutions through mechanisms including target assignment state labeling, coarse screening of suitable spacecraft based on combined orbit transfer strategies, pre-plan adjustment using beam search and auction strategies, and local optimization, thus realizing an efficient solution process characterized by “fast initial optimization followed by fine-tuning”. Finally, scenarios constructed using real orbital data from the CelesTrak website are used to compare the performance of the proposed method with that of genetic algorithms. In small-scale scenario tests, the proposed method exhibits an average convergence efficiency over 150 times higher; in 99 Monte Carlo tests of large-scale scenarios, with the same calculations, the average optimization result of the proposed method is im-proved by 71%, and the standard deviation of multiple solutions is reduced by 84.1%. The convergence is better and more robust. The results verify the effectiveness of the new method, which can significantly improve the efficiency of task replanning and meet the real-time requirements of future applications.

Key words: on-orbit service, Lambert transfer, re-planning, optimization, beam search, attention model