航空学报 > 2021, Vol. 42 Issue (2): 324149-324149   doi: 10.7527/S1000-6893.2020.24149

基于机器学习的空间翻滚目标实时运动预测

余敏1,2, 罗建军1,2, 王明明1,2   

  1. 1. 西北工业大学 深圳研究院, 深圳 518057;
    2. 西北工业大学 航天动力学国家重点实验室, 西安 710072
  • 收稿日期:2020-04-27 修回日期:2020-05-20 发布日期:2020-07-06
  • 通讯作者: 王明明 E-mail:mwang@nwpu.edu.cn
  • 基金资助:
    深圳市科技研发资金(JCYJ20190806154412671);国家自然科学基金(12072269,61973256,61690211);西北工业大学博士论文创新基金(CX202019)

Real-time motion prediction of space tumbling targets based on machine learning

YU Min1,2, LUO Jianjun1,2, WANG Mingming1,2   

  1. 1. Research & Development Institute of Northwestern Polytechnical University in Shenzhen, Shenzhen 518057, China;
    2. School of Astronautics, National Key Laboratory of Aerospace Flight Dynamics, Northwestern Polytechnical University, Xi'an 710072, China
  • Received:2020-04-27 Revised:2020-05-20 Published:2020-07-06
  • Supported by:
    Science, Technology and Innovation Commission of Shenzhen Municipality (JCYJ20190806154412671); National Natural Science Foundation of China (12072269, 61973256, 61690211); Innovation Foundation for Doctor Dissertation of Northwestern Polytechnical University (CX202019)

摘要: 借助监督式机器学习(ML)方法,对空间翻滚目标的运动状态预测问题进行研究,为空间机器人抓捕空间翻滚目标提供可靠的数据依据。基于物理模型的运动预测方法依赖理想的建模假设,需要连续的视觉反馈信息,解决目标预测问题的能力有限。因此,本文采用机器学习中纯数据驱动方式的稀疏伪输入高斯过程(SPGP)回归方法进行空间翻滚目标的运动预测。给定空间翻滚目标运动状态的历史观测数据,通过连续优化真实观测数据,得到稀疏的伪训练数据集,进而在线快速预测目标的运动状态,预测的计算效率达到毫秒级。此外,利用马尔科夫链蒙特卡洛(MCMC)法处理连续优化过程,克服由于随机初始值造成的优化过程陷入局部极小值问题。利用Snelson数据验证了所提稀疏伪输入高斯过程回归方法的正确性,并通过4组仿真算例验证了所提方法对于空间翻滚目标运动预测的有效性和鲁棒性。

关键词: 监督式机器学习, 数据驱动, 稀疏伪输入高斯过程(SPGP), 马尔科夫链蒙特卡罗(MCMC), 空间翻滚目标, 运动预测

Abstract: Based on supervised machine learning (ML), this paper addresses the motion prediction issue of space tumbling targets to provide reliable data of the target motion for space robots when they capture the target. Physics-based motion prediction methods find it hard to solve this problem due to their ideal modelling assumptions and constant requests for vision feedbacks. Hence, a purely data-driven learning-based method, named Sparse Pseudo-input Gaussian Process (SPGP), is employed. Given observed data for the motion state of the space tumbling target, this method continuously optimizes the real data to obtain a sparse pseudo training dataset, making it feasible for a fast online motion prediction implementation with the computational time of prediction within milliseconds. Moreover, the Markov Chain Monte Carlo(MCMC) method is adopted for the continuous optimization, overcoming the local minima problem resulted from the random initial guess during the optimization process. Snelson's data is employed to validate the correctness of the proposed SPGP regression method, and several simulation cases are conducted to demonstrate its effectiveness and robustness.

Key words: supervised machine learning, data-driven, Sparse Pseudo-input Gaussian Process (SPGP), Markov Chain Monte Carlo(MCMC), space tumbling targets, motion prediction

中图分类号: