Aim at the emergency collision avoidance scenarios caused by the close-range space debris, a learning-based spacecraft rapid trajectory planning method, which can adapt to complex constraints and satisfy the requirements of the business service and real-time replanning, is proposed in this paper. First, the emergency collision avoidance scenarios are initialized and the optimal multi-constraint avoidance trajectories are generated based on the Gauss pseudo-spectral method with corresponding feasibility checks. Then, taking the generated collocation points as the initial guess, the new trajectories are regenerated by finetuning scenarios. When the trajectory data is collected to some extent, the scenarios will be reset. The “state-action” data set can be established and extended by the above “plan-check-finetune-reset” loops. On this basis, two types of “state-action” neural networks and the corresponding supervised training method are specially designed based on ideas of multi-layer and bidirectional long short-term memory networks by considering continuous-discrete hybrid control characteristics of spacecraft actuators and interdependent temporal logical relationships in the data generated by dynamic differential equations. The designed networks are trained based on the established data set. Finally, spacecraft attitude-orbit maneuvering instructions can be resolved in real-time by trained networks according to the perceptive information for space debris. Simulation results show that the runtime of the proposed method in each step can be maintained within 10.4 ms, and the overall avoidance success rate can reach 87.6 % in Monte Carlo test conditions.

中文翻译：

---

基于学习的航天器紧急避碰多约束快速轨迹规划

针对近距离空间碎片引起的紧急避碰场景，提出一种基于学习的航天器快速轨迹规划方法，能够适应复杂的约束条件，满足业务服务和实时重规划的要求。纸。首先，初始化紧急避碰场景，并基于高斯伪谱方法生成最优的多约束避碰轨迹，并进行相应的可行性检查。然后，以生成的搭配点作为初始猜测，通过微调场景重新生成新的轨迹。当轨迹数据收集到一定程度时，场景将会被重置。通过上述“计划-检查-微调-重置”循环可以建立和扩展“状态-动作”数据集。在此基础上，考虑航天器作动器的连续-离散混合控制特性，基于多层双向长短期记忆网络的思想，专门设计了两类“状态-动作”神经网络及相应的监督训练方法。由动态微分方程生成的数据中相互依赖的时间逻辑关系。设计的网络根据已建立的数据集进行训练。最后，经过训练的网络可以根据空间碎片的感知信息实时解析航天器姿态轨道机动指令。仿真结果表明，该方法每一步的运行时间均能保持在10.4 ms以内，在蒙特卡罗测试条件下总体避障成功率可达87.6%。