基于ASPPO的侧风着陆训练策略优化  

Optimization of Crosswind Landing Training Strategy Based on ASPPO

作　　者：李嘉伟 高振兴[1] 孙瑾[1] 张洋洋 孔维武 LI Jia-wei;GAO Zhen-xing;SUN Jin;ZHANG Yang-yang;KONG Wei-wu(Nanjing University of Aeronautics and Astronautics,Nanjing 210000,China)

机构地区：[1]南京航空航天大学,江苏南京211000

出　　处：《航空计算技术》2025年第1期76-81,共6页Aeronautical Computing Technique

基　　金：国家自然科学基金项目资助(52272351,U2333202);航空科学基金项目资助(2022Z066052002);民航安全能力建设资金项目资助(ASSA2023/22)。

摘　　要：传统侧风着陆训练依赖于经验式讲解,在复杂的侧风条件下训练质量较低。为提升民用飞机侧风着陆任务的安全性,提出一种改进侧风着陆训练的方法。该方法基于复杂飞行动作构建混合动作决策网络,结合飞行手册安全要求设计了奖励函数,提出使用动作分离近端策略优化算法训练策略网络生成侧风着陆下的最优决策动作序列;进一步,基于模式序列挖掘算法得到最大频繁动作序列模式,通过与飞行员训练实际操作动作序列对比指导飞行员侧风着陆训练。在不同侧风条件下实验结果表明训练的决策网络可以实现既定的着陆策略,同时通过挖掘得到的最大频繁动作序列模式与飞行员操作对比,能及时发现飞行员错漏、多余、偏时等问题,有利于提升飞行员训练效果。To enhance the safety of civil aircraft during crosswind landing tasks,this paper proposes an improved method for crosswind landing training,addressing the low training quality under complex crosswind conditions traditionally reliant on experiential explanations.This method constructs a hybrid action decision network based on complex flight maneuvers and designs a reward function incorporating safety requirements from flight manuals.It proposed an actor separation proximal policy optimization algorithm to train a strategic network that generates optimal decision action sequences for crosswind landings.Furthermore,it employed a pattern sequence mining algorithm to identify the most frequent action sequence patterns.By comparing these patterns with actual pilot training action sequences,it guides pilots in crosswind landing training.Experimental results under various crosswind conditions demonstrated that the trained decision network could implement the established landing strategy.Simultaneously,by comparing the mined maximum frequent action sequence patterns with pilot operations,it could promptly identify pilot errors,omissions,redundancies,and timing deviations,thereby contributing to improved pilot training effectiveness.

关 键 词：侧风着陆 仿真 强化学习 近端策略优化 

分 类 号：V323[航空宇航科学与技术—人机与环境工程]