基于DQN的智能农机路径跟踪控制研究  

DQN-based Path Tracking Control for Intelligent Agricultural Machinery

作　　者：杨琰 张瑞瑞[2] 张林焕 陈立平[2] 伊铜川 吴明齐 岳晓龙 Yang Yan;Zhang Ruirui;Zhang Linhuan;Chen Liping;Yi Tongchuan;Wu Mingqi;Yue Xiaolong(School of Agricultural Engineering,Jiangsu University,Zhenjiang 212013,China;Beijing Research Center of Intelligent Equipment for Agriculture,Beijing Academy of Agricultural and Forestry Sciences,Beijing 100097,China;Nongxin(Nanjing)Smart Agriculture Research Institute Co.,Ltd,Nanjing 211800,China)

机构地区：[1]江苏大学农业工程学院,江苏镇江212013 [2]北京市农林科学院智能装备技术研究中心,北京100097 [3]农芯(南京)智慧农业研究院有限公司,南京211800

出　　处：《农机化研究》2025年第3期28-34,共7页Journal of Agricultural Mechanization Research

基　　金：国家重点研发计划项目(2022YFD2202102-02);南京市企业院士工作站关键核心技术攻关项目(2021—2023);北京市农林科学院改革与发展项目(2023);青年北京学者专项(2023)。

摘　　要：针对复杂路面条件下无人化农业作业车辆路径跟踪控制精度低、控制器参数整定困难的问题,设计了一种基于深度强化学习的路径跟踪控制算法。基于五层BP神经网络构建了DQN(Deep Q-Network)路径跟踪控制器,实现了网络的轻量化和高度的可移植性。控制器网络的输入状态在采用车辆当前舵角、车辆与目标路径上控制点间横向距离偏差的基础上,引入车辆前方设定距离内的平均路径曲率,实现了车辆转向性能的提高。分别开展了仿真和田间试验,验证了所设计深度强化学习网络算法的收敛性,并对比了有路径曲率输入和无路径曲率输入两种网络的路径跟踪控制性能。仿真试验中,基于正弦曲线对两种路径跟踪控制方法进行训练,两种路径跟踪控制网络收敛后的平均跟踪距离误差为0.008 4、0.017 7 m。在间隔6 m的U型路径上进行田间测试试验,结果显示:有路径曲率输入的模型性能显著优于无路径曲率输入的模型,两种路径跟踪控制方法在路径上平均跟踪距离误差为0.038 9、0.068 4 m。验证了该路径跟踪控制方法有效性,可满足农业作业车辆路径跟踪控制需求。Aiming at problems of low control precision and difficulty of controller parameter tuning for unmanned agricultural vehicles under complex agricultural unstructured field conditions,a deep reinforcement learning based path tracking control algorithm was designed.The DQN(Deep Q-Network)based path tracking controller was established based on five-layer BP neural networks,which realized lightweight and high portability of the network.As input states of the controller network,average path curvature of path segment in front of the vehicle was introduced along with steering angle and lateral distance deviation between the vehicle and path control point,which benefit to improve tracking accuracy on curved path.Simulation and field experiments were carried out to verify the convergence of the designed path tracking algorithm,and further compared the path tracking performance of two types of networks with path curvature input and without path curvature input.For the simulation experiment,the two types of path tracking controller were trained based on a sinusoidal curve,and average distance tracking error of the two type of controller were 0.0084 m and 0.0177 m,respectively.Field experiment was conducted on U-shaped paths with interval of 6 m,and the results showed that,the tracking accuracy of the controller with path curvature input was higher than that of the model without path curvature input.The average distance tracking errors for the two types of path tracking controller were 0.0389 m and 0.0684 m,respectively.The effectiveness of the path tracking control method proposed in this paper was verified,and satisfied the requirements of agricultural application.

关 键 词：路径跟踪 深度强化学习 DQN 路径曲率 智能农机 

分 类 号：S126[农业科学—农业基础科学]