GEO Spacecraft Rendezvous Sequence Mission Planning Based on Data Augmented REINFORCE Algorithm

This paper proposes a fast multi-spacecraft rendezvous sequence mission planning algorithm based on neural combinatorial optimization (NCO) and deep reinforcement learning (DRL). The algorithm can address rendezvous requirements for various GEO spacecraft on-orbit servicing missions, such as on-orbit maintenance and on-orbit refueling. In order to meet the fast decision-making needs of multi-spacecraft rendezvous missions and ensure the successful completion of the missions, we take advantage of the fast calculation speed of the DRL algorithm and make some unique designs to address the challenges of missions, and finally propose the data augmented REINFORCE algorithm (DARA). The algorithm can rapidly plan the optimal rendezvous sequence for multiple GEO targets and minimize fuel consumption while completing all rendezvous missions. It is implemented based on the REINFORCE algorithm using a multi-head attention neural network. In order to solve the problem that DRL training is difficult due to the time-varying characteristics of the space targets' positions, a data augmentation method is designed to strengthen the learning weights of different parameters. In order to solve the problem of trivial objective function gradient and slow iteration caused by the dense distribution of orbital planes, a loss function is designed to improve the earning efficiency of the neural network. Through multiple comparative experiments with existing algorithms, the calculation speed of DARA significantly exceeds that of commonly used heuristic mission planning algorithms. The results accuracy is better than that of various meta-heuristic algorithms and typical reinforcement learning algorithm, which proves the effectiveness and superiority of the proposed algorithm.