采用注意力模型的多星交会序列优化方法  

Sequence Optimization Method for Multi-satellite Rendezvous Using Attention Model

作　　者：严冰 张进[1,2] 罗亚中 朱阅訸[1,2] YAN Bing;ZHANG Jin;LUO Yazhong;ZHU Yuehe(College of Aerospace Science and Engineering,National University of Defense Technology,Changsha 410073,China;Hunan Key Laboratory of Intelligent Planning and Simulation for Aerospace Missions,Changsha 410073,China)

机构地区：[1]国防科技大学空天科学学院,长沙410073 [2]空天任务智能规划与仿真湖南省重点实验室,长沙410073

出　　处：《宇航学报》2023年第11期1683-1692,共10页Journal of Astronautics

基　　金：湖南省自然科学基金(2022JJ10061);国家杰出青年科学基金(12125207);国家自然科学基金(11972044,12102460)。

摘　　要：针对强化学习方法能够实现组合优化问题从无序输入到有序输出的离散决策特点,提出一种基于注意力模型的单航天器对多目标交会的序列优化方法,以快速估计转移成本较低的交会序列和时间。通过拓展注意力的时间维度使得策略网络能够同时选择目标和时间,并设计启发因子引导注意力有侧重地分布于不同的交会时刻,然后基于带基线的REINFORCE算法对网络参数进行更新。将所提方法应用于轨道分布较为集中的10目标交会任务,所得序列成本相对于蚁群算法优化结果的平均误差为9.7%,且强化学习方法的计算时间极短,可用于多目标对多航天器的分组指派全局规划的底层评价。In view of the fact that reinforcement learning with the characteristic of discrete decision-making can output the optimal sequence of combinatorial optimization problems,a sequence optimization method based on attention model for multi-satellite rendezvous with one spacecraft is proposed to quickly estimate the sequence and time with low transfer cost.The temporal dimension of attention is expanded to make the policy network simultaneously choose the target and time,and the heuristic factor is designed to distribute attention at different times.Then the network parameters are updated based on the REINFORCE algorithm with baseline.For 10-target rendezvous with relatively concentrated orbital distribution,the proposed method is compared with the ant colony algorithm and the mean relative error of the sequence cost is about 9.7%.The calculation time of the learning method is very short,which can be taken as the fundamental support for multi-satellite rendezvous with multiple spacecraft.

关 键 词：空间交会 多星交会 序列优化 强化学习 注意力模型 

分 类 号：V448.2[航空宇航科学与技术—飞行器设计]