无人机栖落机动的基于学习的可达域算法研究  

Research on learning-based reachable region algorithm for UAV perching maneuver

作　　者：侯青杉 彭余萧 何真[1] 张鹏生 汤张帆 HOU Qingshan;PENG Yuxiao;HE Zhen;ZHANG Pengsheng;TANG Zhangfan(College of Automation Engineering,NUAA,Nanjing 210016,China)

机构地区：[1]南京航空航天大学自动化学院,江苏南京210016

出　　处：《飞行力学》2025年第1期48-55,共8页Flight Dynamics

基　　金：国家自然科学基金资助(61873126)。

摘　　要：无人机栖落机动过程具有高度非线性的动力学特性,控制律较为复杂,可达域计算则是验证其控制系统能否完成栖落的一个重要技术手段。首先,基于Hamilton-Jacobi可达性分析理论设计一种时间折扣贝尔曼方程,验证基于该方程的可达域算法的可行性;接着拓展到可达避免问题,在时间折扣贝尔曼方程中引入目标裕度函数,使无人机在遵守安全性约束的同时能被引导向目标区域;然后,根据栖落机动过程中无人机的状态量具有时间序列的特点,采用递归神经网络对算法网络进行了改进;最后,开展了仿真分析。研究结果表明,该算法在计算无人机栖落机动控制系统的可达域中表现优异,具有高正确率和低迭代次数。The perching maneuver of unmanned aerial vehicle(UAV)exhibits highly nonlinear dy-namic characteristics,and its control laws are rather complex.Reachability analysis serves as a cru-cial technical means to verify whether the control system can successfully accomplish the perching maneuver.Firstly,a time-discounted Bellman equation is designed based on the Hamilton-Jacobi reachability analysis theory,verifying the feasibility of the reachable region algorithm derived from this equation.Subsequently,extending to the reach-avoid problem,a target margin function is intro-duced into the time-discounted Bellman equation.This allows the UAV to be guided to the target ar-ea while adhering to safety constraints.Then,based on the time-series characteristics of the state variables of the UAV during the perching maneuver,the algorithm network is improved using a re-current neural network.Finally,simulation analysis is conducted.The research results show that the algorithm works excellently in computing the reachable set of the UAV perching maneuver control system,with high accuracy and low numbers of iterations.

关 键 词：栖落机动 可达域 可达避免 递归神经网络 

分 类 号：V279[航空宇航科学与技术—飞行器设计] V249