基于深度神经网络的无限时域型航天器追逃策略求解  

Solution of Infinite Time Domain Spacecraft Pursuit Strategy Based on Deep Neural Network

作　　者：吴其昌 李彬 李君[2] 张洪波[1] Wu Qichang;Li Bin;Li Jun;Zhang Hongbo(College of Aerospace Science and Engineering,National University of Defense Technology,Changsha 410073,China;China Academy of Launch Vehicle Technology,Beijing 100076,China)

机构地区：[1]国防科技大学空天科学学院,长沙410073 [2]中国运载火箭技术研究院,北京100076

出　　处：《航天控制》2019年第6期13-18,58,共7页Aerospace Control

基　　金：装备预研航天科技联合基金(6141B060907)

摘　　要：航天器追逃博弈是当前航天领域的一个研究热点,传统上多采用微分对策来获取追逃双方的最优控制策略,但是方法求解复杂、计算量大,难以满足复杂任务和对抗类任务的实时性要求。随着机器学习技术的发展,利用深度神经网络结构实现全部或部分的在线决策成为可能,因此研究了基于深度神经网络生成无限时域型追逃博弈最优控制策略问题。首先基于CW方程建立追逃博弈相对运动模型,采用微分对策理论得到追逃最优控制策略,得到训练数据集和测试数据集;基于TensorFlow环境搭建了4层神经网络,采用Adam优化算法对网络进行训练。仿真结果表明,经过训练的深度神经网络生成的控制策略与传统方法的策略基本一致,虽然长时间追逃的控制差异逐渐增大,但变化趋势相同,说明利用深度神经网络生成航天器追逃博弈的机动策略是有效的。The spacecraft pursuit game is a research hotspot in the field of aerospace.Traditionally,differential countermeasures are used to obtain the optimal control strategy for both sides.However,the method is complex and computationally intensive,and it is difficult to meet the real-time requirements of complex tasks and confrontation tasks.With the development of machine learning technology,it is possible to realize all or part of online decision making by using deep neural network structure.Therefore,the problem of generating optimal control strategy for infinite time-domain pursuit game based on deep neural network is studied.Firstly,based on the CW equation,the relative motion model of the pursuit game is established. The optimalgame control strategy is obtained by using the differential game theory to obtain the training data setand the test data set. Based on the TensorFlow environment,a 4-layer neural network is built,and the Adamoptimization algorithm is used to train the network. The simulation results show that the control strategygenerated by the trained deep neural network is basically consistent with the traditional method,althoughthe control difference of long-term pursuit is gradually increased. However,the trend of change is the same,and it shows that the mobile strategy of using the deep neural network to generate the spacecraft pursuitgame is effective.

关 键 词：深度神经网络 微分对策 追逃博弈 

分 类 号：V412.[航空宇航科学与技术—航空宇航推进理论与工程]