基于强化学习的多空间碎片清除快速任务 规划方法

To address the problems of low efficiency and strong resource coupling in mission planning for multiple space debris removal，a rapid task planning method based on reinforcement learning is proposed by taking optimal fuel consumption and debris removal priorities into comprehensive consideration. An environment model for space debris removal mission planning is established，which provides support for satellite orbital dynamics analysis，state transition and reward calculation. A feature fusion and encoding mechanism called“global deconstruction-pairing encoding-deep fusion” is designed to convert high-dimensional and strongly coupled global task states into service satellite-debris pairing units. It realizes efficient fusion and encoding of service satellite states，debris characteristics and relative features，and reduces the computational complexity of high-dimensional features. Combined with multi-head attention and context fusion mechanisms，a four-layer progressive hierarchical solution architecture is developed to achieve fast processing of complex coupling constraints，and the debris removal sequence and corresponding operation time of each service satellite are obtained. Simulation results show that the proposed method can stably generate effective planning schemes with excellent generalization ability. Compared with classical meta-heuristic algorithms and mainstream reinforcement learning algorithms，it performs better in key indicators including planning efficiency，cumulative reward and fuel consumption， which can provide important technical reference for the engineering application of large-scale space debris removal.