含未知动态与扰动的非线性系统神经网络嵌入学习控制  

Neural Network Embedded Learning Control for Nonlinear System With Unknown Dynamics and Disturbance

作　　者：马乐[1] 闫一鸣 徐东甫 李志伟 孙灵芳[1] MA Le;YAN Yi-Ming;XU Dong-Fu;LI Zhi-Wei;SUN Ling-Fang(School of Automation and Engineering,Northeast Electric Power University,Jilin 132012;Jilin Province International Research Center of Precision Drive and Intelligent Control,Jilin 132012)

机构地区：[1]东北电力大学自动化工程学院,吉林132012 [2]吉林省精密驱动智能控制国际联合研究中心,吉林132012

出　　处：《自动化学报》2021年第8期2016-2028,共13页Acta Automatica Sinica

基　　金：国家自然科学基金(61673101);吉林重点行业与产业科技创新计划人工智能专项(2019001090)资助。

摘　　要：针对带有不确定性与扰动的非线性系统的性能优化问题,提出一种基于神经网络嵌入的学习控制方法.对一类常见的Lyapunov函数导数形式,将神经网络控制器集成到某种对系统稳定的基准控制器中,其意义在于将原控制器改进为满足Lyapunov稳定的神经网络参数可调控制器,从而能够利用先进的神经网络学习技术实现控制器的在线优化.建立了跟踪误差的等效目标函数,避免了对系统输入–输出的辨识问题.建立了一种未知非线性与扰动等效值自适应方法,并依此方法设计基准控制器.以RBF(Radial basis function)反步自适应控制、基于卷积神经网络的滑模控制和深度强化学习控制为对比方法,对带有死区、饱和、三角函数等数值与物理非线性模型进行仿真分析以测试方法有效性,并针对上肢康复机器人控制问题进行虚拟实验以验证该方法的实用性.仿真与实验结果表明,该方法能在Lyapunov稳定条件下有效优化基础控制器性能,对比结果证实了该方法的实用性与先进性.To address the problem of controlling performance optimization for the nonlinear uncertain system with disturbance,a neural network embedded learning control scheme is proposed in this paper.This method works on a common formal derivative of Lyapunov function,in which a neural network controller is integrated with a benchmark controller that is stable for the system.The main contribution of our work lies in that the benchmark controller is improved to a new one with tunable parameters under Lyapunov stability condition,and the new controller can be online optimized by using frontier technology of neural network.Hence an equivalent objective function based on tracking errors is characterized in this paper,avoiding identification to the relations between inputs and outputs of system.We use a value adaptive method for estimating equivalent term composed of unknown nonlinear function and disturbance,and the benchmark controller is designed based on this method.Some baseline methods are employed for comparison with the proposed method,which contain adaptive control based on RBF(Radial basis function)-backstepping,sliding mode control based on convolutional neural network and deep reinforcement learning control.And for verifying the effectiveness of our method we test some numerical and physical nonlinear model simulations,which contain trigonometric function saturation and dead zone nonlinearities.And virtual experiments of robot arm controlling of upper limb rehabilitation to be tested to verify the practicability of our method.These results show that the method proposed is able to optimize control performance of benchmark controller with Lyapunov stability.And the results of comparisons of tests show our method is efficient and advanced.

关 键 词：神经网络嵌入 优化控制 深度学习技术 未知非线性动态 不确定与扰动 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP13[自动化与计算机技术—控制科学与工程]