基于线性滤波器的四旋翼无人机强化学习控制策略  

Reinforcement Learning Control Strategy of Quadrotor Unmanned Aerial Vehicles Based on Linear Filter

作　　者：华和安 方勇纯 钱辰 张雪涛[2] HUA He'an;FANG Yongchun;QIAN Chen;ZHANG Xuetao(College of Artificial Intelligence,Nankaz University,Tianjm 300350,China;Intelligent Robotic Laboratory,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]南开大学人工智能学院,天津300350 [2]大连理工大学智能机器人实验室,大连116024

出　　处：《电子与信息学报》2021年第12期3407-3417,共11页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61873132,61633012)。

摘　　要：针对四旋翼无人机(UAVs)系统,该文提出一种基于线性降阶滤波器的深度强化学习(RL)策略,进而设计了一种新型的智能控制方法,有效地提高了旋翼无人机对外界干扰和未建模动态的鲁棒性。首先,基于线性降阶滤波技术,设计了维数更少的滤波器变量作为深度网络的输入,减小了策略的探索空间,提高了策略的探索效率。在此基础上,为了增强策略对稳态误差的感知,该文结合滤波器变量和积分项,设计集总误差作为策略的新输入,提高了旋翼无人机的定位精度。该文的新颖之处在于,首次提出一种基于线性滤波器的深度强化学习策略,有效地消除了未知干扰和未建模动态对四旋翼无人机控制系统的影响,提高了系统的定位精度。对比实验结果表明,该方法能显著地提升旋翼无人机的定位精度和对干扰的鲁棒性。In this paper, based on linear filter, a deep Reinforcement Learning(RL) strategy is proposed, then a novel intelligent control method is put forward for quadrotor Unmanned Aerial Vehicles(UAVs), which improves effectively the robustness against disturbance and unmodeled dynamics. First of all, based on linear reduced-order filtering technology, filter variables with fewer dimensions are designed as the input of the deep network, which reduces the exploration space of the strategy and improves the exploration efficiency. On this basis, to enhance strategy perception of steady-state errors, the filter variables and integration terms are combined to design the lumped error as the new network input, which improves the positioning accuracy of quadrotor UAVs. The novelty of this paper lies in that it is the first intelligent approach based on linear filtering technology, to eliminate successfully the influence of unknown disturbance and unmodeled dynamics of quadrotor UAVs, which improves the positioning accuracy. The results of comparative experiments show the effectiveness of the proposed method in terms of improving positioning accuracy and enhancing robustness.

关 键 词：四旋翼无人机 智能控制 强化学习 未知干扰 

分 类 号：V279[航空宇航科学与技术—飞行器设计] TP273[自动化与计算机技术—检测技术与自动化装置]