深度神经网络辅助的垂直回收火箭在线轨迹优化方法

A deep neural network assisted trajectory optimization algorithm for vertical landing vehicles

针对垂直回收火箭在线轨迹规划的计算效率和初始敏感问题,提出一种深度神经网络辅助的在线轨迹优化算法。考虑火箭动力下降段的气动阻力,使用变分法和庞德里亚金极小值原理推导最优性条件,首次证明最优推力矢量幅值存在Bang-Bang特征。在此基础上,设计离线训练和在线优化两步求解框架。一是离线训练深度神经网络,在初值大范围波动条件下,有监督学习Bang-Bang特征的结构参数;二是在线规划最优轨迹,将训练好的深度神经网络作为辅助求解器,生成伪谱离散法的分段点,嵌入序列凸优化算法求解。该算法将最优推力的与伪谱法的分段特性有机结合,提高了有限离散点下的求解精度。仿真结果表明,该算法能有效提升在线轨迹规划的求解效率和初值适应性。

It is challenging to solve the powered descent guidance problem online for its computational cost and uncertain initial conditions.An Hp-pseudospectral convex optimization algorithm assisted by deep neural network is presented.For the highly nonlinear dynamics in atmosphere,it is proved for the first time that the thrust magnitude profile has the Bang-Bang feature based on variational method and Pontryagin′s maximum principle.In the wide range of initial states,the deep neural network is applied to learn the segment feature of optimal thrust offline.Then the trained neural net is embedded in the online successive convex optimization algorithm,which combines the Hp-pseudospectral discretization with Bang-Bang feature.This learning assisted strategy leads to more accurate results with the same number of discretized nodes.Numerical simulations show that the proposed algorithm shows better computational efficiency and adaptability to initial conditions.