关键词: 多目标, 协同观测, 任务规划, 强化学习, 自注意力机制, 近端策略优化

Abstract: With the increasing number of space targets, the problem of orbit determination of the targets is becoming increasingly important for space security. Due to the large number and dynamic feature of the space targets that need to be observed, coupled with limited observation resources, it is necessary to dynamically adjust the collaborative observation scheme to efficiently utilize constellation observation resources and ensure that each target has better positioning accuracy. Thus, it is required to solve the mission planning problem of multiple targets using multiple observation satellites. This paper first establishes the orbit dynamic model of the flying targets, as well as the Kalman filter model of the collaborative positioning algorithm using the multiple line of sight information of different observation satellites. Then, a collaborative positioning accuracy estimation model and an observation priority model of the targets based on the Geometric Dilution of Precision (GDOP) is proposed. Based on the above models, a mission planning framework for collaborative observation based on reinforcement learning (RL) is developed. A policy network based on multihead selfattention mechanism is designed accordingly to calculate the planning results. The proximal policy optimization (PPO) algorithm is adopted to train the policy network in a training environment. Compared with the heuristic algorithm based on tracking priority, simulation results shows that the proposed RL method can effectively improve the overall tracking accuracy as well as the total tracking time of all the targets, and can provide faster computation speed compared to genetic algorithms.

Key words: multiple targets, collaborative observing, mission planning, reinforcement learning, selfattention mechanism, proximal policy optimization algorithm