中国空间科学技术

中国空间科学技术 ›› 2018, Vol. 38 ›› Issue (3): 8-14.doi: 10.16708/j.cnki.1000-758X.2018.0020

• 研究探讨 • 上一篇    下一篇

基于随机森林的月球表面软着陆实时最优控制

姜春生1,2，沈红新2，李恒年2，王永1,*   

  1. 1. 中国科学技术大学自动化系，合肥 230027
    2. 宇航动力学国家重点实验室，西安 71004
  • 收稿日期:2017-11-13 修回日期:2018-01-22 接受日期:2018-03-14 出版日期:2018-06-25 发布日期:2018-03-06
  • 通讯作者: 王永（1962-），男，教授，yongwang@ustc.edu.cn，研究方向为导航、制导与控制
  • 作者简介:姜春生（1990-），男，硕士研究生，csjiang@mail.ustc.edu.cn，研究方向为导航、制导与控制

Optimal real-time lunar soft landing using random forest

JIANG Chunsheng1,2, SHEN Hongxin2, LI Hengnian2, WANG Yong1,*   

  1. 1. Department of Automation, University of Science and Technology of China, Hefei 230027, China
    2. State Key Laboratory of Astronautic Dynamics, Xi'an 710043, China
  • Received:2017-11-13 Revised:2018-01-22 Accepted:2018-03-14 Published:2018-06-25 Online:2018-03-06

PDF (PC)

181

摘要： 针对传统月球表面软着陆在处理入轨偏差或降落过程轨迹偏离实时性差，文章提出一种通过监督学习离线训练随机森林结构，使得着陆器在降落过程中根据其状态，通过训练的随机森林结构在线计算其控制量，从而达到实时控制的目的。文章还提出一种基于随机森林的模型对月球着陆过程轨迹重规划技术，通过动力学建模将月球着陆过程分成制动段、接近段和着陆段共3个阶段，利用离线训练好的模型根据航天器状态在线计算其控制量，并通过三段下降过程逐渐降低航天器位置速度误差。仿真结果表明，针对入轨偏离500m的情况，通过第一阶段将位置误差缩短至50m，保证了航天器状态位于第二阶段训练集内，经过接近段后再将位置误差缩小至10m范围内，速度误差降至0.01m/s量级，满足着陆误差要求，且控制量计算时间短，达到了轨迹实时重规划的目的。

关键词: 月球表面软着陆, 实时控制, 监督学习, 随机森林, 动力学建模, 轨迹重规划

Abstract:

Traditional optimal trajectories for lunar soft landing can’t be calculated in real time if there’s error when spacecraft entering orbit or during the descending stage. A novel scheme was proposed to replan the trajectory in real time for the random forest model. The descending stage was modeled by three phases: deboost, descend and land. The modelcan predict the control by the state of spacecraft as the input, the lander descend slower and the error gets smaller. The simulation results show that with a 500m error in entering orbit the distances deceases to 50m after the deboost phase. This error is in the range of the training set of descend phases, so after that the position error is less than 10m and the velocity error is less than 0.01m/s, both in tolerance. Meanwhile this method is fast enough to satisfy the real time control.

Key words: lunar softlanding, real-time control, supervised learning, random forest, dynamic , modeling, trajectory replan