CASIA OpenIR  > 毕业生  > 博士学位论文
基于新型神经网络与触发机制的机械臂自适应控制研究
高洁
2022-05
Pages185
Subtype博士
Abstract

实现精准的轨迹跟踪和与环境的可靠交互一直是机械臂控制研究的重点。随着机械臂操作任务和环境的日益复杂化,为保证操作性能,机械臂的动力学结构逐渐具备高耦合、高冗余和强非线性等特点,而复杂冗余机械臂系统所包含的扰动和参数不确定性等进一步增加了精准控制的难度。近年来,基于反步法和神经网络估计的自适应控制技术在解决高阶不确定非线性系统的控制问题上具有很好的效果。然而,自适应控制的精度和稳定性往往受网络估计性能的约束;其次,自适应神经网络控制中的信息传递和参数更新是由时间驱动的,由于系统内部各节点之间的通信带宽和计算资源的有限性,参数的实时更新和传输往往难以保证,且频繁的传输与更新也进一步增加了通信与计算负担。针对上述问题,本论文分别针对机械臂的可靠交互和精准跟踪两个任务场景,通过融合新型神经网络估计和触发式传输调控等技术,开展面向机械臂的自适应控制新方法研究。一方面,针对机械臂在有限空间下的可靠交互,本课题构建了基于新结构动态网络估计的自适应阻抗控制方法,提高控制系统对复杂动力学参数和不确定接触扰动的估计能力;另一方面,针对信息交换有限和参数不确定下的机械臂稳定跟踪任务,研究构建了不同通信信道下的自适应触发控制,在确保控制稳定的前提下,提高通信资源的有效利用率并实现多性能均衡。鉴于上述内容,本论文的研究主要分为以下四个内容:

  1. 针对不确定动力学、接触扰动以及输入饱和下的机械臂可靠交互问题,本章研究基于反步法和阻抗模型,构建了带有新型动态网络的自适应状态反馈阻抗控制器,通过该动态网络实现对动力学参数以及接触扰动的估计补偿。为进一步提高自适应控制器中动态网络的估计性能,研究分别在网络构型和自适应律两方面做了新的设计。在网络构型的设计方面,研究提出了新型时滞回声状态神经网络结构,一方面,在输入层与隐含层之间引入感受野的关联映射机制,以提高网络对不同输入信号的区分性和局部泛化性;另一方面,在隐含层构建了带有随机时滞神经元反馈的蓄水池计算框架,用于增强网络的非线性动态特性和多分辨率时序记忆能力。在自适应律的设计方面,研究从渐近稳定的角度出发,基于泰勒展开和矩阵映射算法,为网络权值构造了带有 σ 修正项的自适应律,以保证权值的收敛性和有界性,并降低权值收敛对持续激励的需求。进一步地,研究通过引入自适应辅助变量补偿饱和误差,以降低其对精准控制的影响。基于李雅普诺夫稳定性理论,研究证明了所提控制方案能够使误差变量满足一致最终有界,并给出了具体上界函数的表达形式。最后,研究在一个三自由度机械臂的交互仿真中验证了所提控制方法的有效性。
  2. 在上一章关于动态网络估计控制的研究基础上,本章研究针对动力学参数不确定、未知干扰和控制-执行通道有限传输下的机械臂稳定跟踪问题,提出了一种基于扰动观测器和动态网络辨识的自适应事件触发控制器,在保证跟踪误差渐近收敛的情况下,降低控制-执行通道的传输负担。为抑制未知扰动对触发控制性能的影响,研究根据系统的动力学结构设计了扰动观测器,并将扰动观测值作为状态反馈控制器的补偿项。另外,研究构建了带有输出反馈的回声状态动态网络以实现对动力学参数的辨识。为降低控制-执行通道的传输负担,研究建立了事件调控的控制信号传输机制,并设计了带有正则项的事件触发条件以满足最小触发间隔的正定性,即无芝诺行为。最后,通过李雅普诺夫稳定性理论,研究给出了系统误差变量满足一致最终有界的参数条件,并在二自由度机械臂上对所提方法的有效性进行了仿真验证。
  3. 本章研究针对速度状态不可测、动力学和干扰参数不确定以及传感-控制通道的传输受限下机械臂稳定跟踪控制问题,设计了一种基于模型的自适应动态事件触发输出反馈控制器,其中,未知的速度状态由连续网络观测器进行估计。在所提的控制方案中,被控对象的显式模型被用来拟合非触发期间系统状态的演化以泛化零阶保持器 (zero-order holder,ZOH) 的作用,达到延长相邻触发时间间隔的效果,且该模型状态值被用作控制器的实际反馈项。同时,研究构建了具有非周期更新特性的径向基神经网络用于对不确定的动力学和干扰进行估计,且自适应律构造为关于触发偏差的函数保证权值在有限更新下的收敛性。另外,研究在事件触发条件中构造了离散动态阈值,使触发阈值随系统性能的变化产生不同程度地收缩调整,以达到触发效率和控制精度的均衡。在稳定性分析方面,研究将该触发式非连续反馈控制系统描述为脉冲系统,并基于脉冲系统的李雅普诺夫稳定性理论,给出了所有误差变量满足半全局一致最终有界的参数条件,并证明相邻执行间隔始终存在正定下界。最后,研究在三自由度机械臂上的跟踪仿真实验上,验证了该方法在提高触发控制精度和降低传输频率的有效性。
  4. 本章研究继续针对传感-控制通道信息传输受限下的观测器输出反馈跟踪控制问题,构建了基于一阶滤波技术的自适应事件触发和自触发输出反馈控制方案。与上一章研究工作不同的是,本章研究考虑了观测器和控制器同时存在反馈状态的间歇性传输情况。为促进间歇性反馈下观测和控制误差的收敛,研究首先提出了基于模型自适应估计的事件触发控制方法,该方法通过构建自适应模型,预测了系统状态在非触发阶段的演化,并将其用于开环阶段的状态反馈补偿以保证观测器和控制器的收敛。另外,研究还考虑了触发不连续性对控制设计和性能分析的影响。一方面,为消除非连续反馈下虚拟变量的求导奇异问题,研究引入了带有滤波补偿的一阶滤波控制方法,避免了对虚拟变量的直接求导。另一方面,研究新设计了步长可调的“软”触发更新机制,以降低触发时刻下状态跳变引发的控制抖振。基于上述设计,本章构造了带有相对阈值和可调死区容限的触发条件,以保证系统误差的一致最终有界和最小执行间隔的正定性。除此之外,研究还从理论上给出了自触发的形式,实现了触发时刻的预判。最后,研究通过软件仿真和硬件实验,验证了所提方法在提高触发控制可靠性和灵活性方面的优势。
Other Abstract

Accurate trajectory tracking and reliable interaction with the environment have always been the focus of manipulator control research. With the increasing complexity of manipulator operation tasks and environment, the dynamic structure of manipulator is gradually characterized by high coupling, high redundancy and strong nonlinearity in order to achieve the desired performance, Whereas, the disturbance and parameter uncertainty contained in the system with complicated configuration further increases the difficulty of the precise control. In recent years, adaptive control techniques based on backstepping and neural network estimation have achieved good results in solving the stable control problem of high-order and uncertain nonlinear systems. However, the accuracy and stability of adaptive control are often constrained by the performance of network’s estimation. Besides that, the information transmission and parameter updating in the adaptive neural network control are often driven by time, and due to the limitation of communication bandwidth and computing resources among nodes within the system, the real-time updating and transmission are difficult to be guaranteed, which may cause the high computation and transmission burden. In view of the above problems, this paper has carried out new adaptive backstepping control methods by integrating novel network estimation and triggered mechanism, for the manipulator under two task scenarios, namely, reliable interaction and trajectory tracking, respectively. On the one hand, an adaptive impedance control method based on dynamic network estimation with the new structure has been proposed to improve the estimation accuracy of complex parameters and uncertain contact disturbances of the control system during the reliable interaction. On the other hand, aiming at the stable tracking with limited information exchange and uncertain parameters, the adaptive triggered control under different communication channels has been constructed to improve the effective utilization of communication resources and achieve multi-performance balance under the premise of stability. In view of the above contents, the research of this paper is mainly divided into the following four contents:

  1. In view of realizing the accurate tracking and reliable interaction of the manipulator with dynamics uncertainty, contact disturbance, and input saturation, an adaptive state feedback impedance controller with novel dynamic neural network estimation has been constructed based on backstepping method and impedance model, where the dynamic uncertainty and contact disturbance were compensated by neural network. In order to improve the estimation performance of the adaptive controller for further, a novel time-delay echo state dynamic neural network with global online learning mechanism was proposed, and the new design was made on the architecture and adaptive law of the network. In the aspect of architecture design, a receptive field association mapping mechanism was introduced between the input and the hidden layer to improve the local discrimination and generalization of different input signals; Besides that, a reservoir computing framework with stochastic time-delay feedback was constructed at the hidden layer to enhance the nonlinear dynamics and the multi-resolution memory of timing sequence. In the aspect of adaptive laws design, we have constructed adaptive laws with σ modification based on Taylor expansion and projection algorithm for weights updating, so as to ensure the convergence and boundedness of weights without the requirement of persistent excitation conditions. Furthermore, the negative effect of saturation bias on the precise control was further reduced by introducing an adaptive auxiliary variable. From the perspective of Lyapunov stability, it has been proved that the proposed control scheme could make all variables uniformly ultimately bounded, and the upper bound was given. Finally, the effectiveness of the proposed control method was verified by an interactive simulation of a 3-DOF manipulator.
  2. Based on the research of dynamic network estimation control in the last chapter, this study has paid close at the stable tracking of manipulator with uncertain dynamics, uncomprehending disturbance and limited transmission in the control-to-actuator channel, and developed an adaptive event-triggered controller based on disturbance observer and dynamic network identification to reduce the transmission burden of control-to-actuator channel while ensuring the convergence of tracking error. To reduce the effect of unknown disturbance on triggered control performance, a disturbance observer was designed according to the dynamic structure of the system, and the observation value was used as the compensation term of the state feedback controller. In addition, an echo state network with output feedback was constructed to identify uncertain dynamic parameters. To reduce the transmission burden of control-to-actuator channel, the event-regulated control signal transmission mechanism was established, and within the above design, the event condition with a regular term was designed to guarantee the positivity of minimum triggering interval. Finally, the parameter condition of uniformly bounded error was given based on the Lyapunov theory, and the effectiveness of the proposed method in keeping small control.  transmission was verified by simulation on a 2-DOF manipulator.
  3. To solve the stable tracking control problem of manipulator under the effect of undetectable velocity state, the dynamics uncertainty and disturbance, as well as limited transmission in the sensor-to-control channel, an adaptive model-based event-triggered output feedback controller has been designed, where the unknown velocity was estimated by the observer. In the proposed control scheme, an explicit model of the plant was used to estimate the evolution of the system state during the non-triggering period, such that the effect of the zero-order holder could be generalized, and this model state was used as the actual feedback item of the controller to further extend the time interval of adjacent triggers. A radial basis function neural network with updating was designed to estimate the uncertain dynamics and disturbance of the controller, and the adaptive law was constructed as a function about triggered deviation to realize the convergence under finite updating. In addition, a dead-zone event condition with discrete dynamic threshold was designed to make the triggered threshold shrink with the change of system performance, so as to make a trade-off between the transmission efficiency and control accuracy. In terms of stability analysis, the discontinuous triggered control system was described as an impulse system, and based on the Lyapunov stability theory of the impulse system, the parameter condition that all error variables satisfied the semi-global uniformly boundedness was given, and the fact that there is no occurrence of Zeno behavior has been proved. Finally, simulation experiments on a 3-DOF manipulator verified the effectiveness of the proposed method in improving the triggered control accuracy with low transmission frequency.
  4. To tackle the problem of observer output feedback tracking control under communication constraints in the sensor-to-control channel, two triggered control mechanisms, namely, first-order filter based adaptive event triggered control and self-triggered control, have been proposed. Different from the research in the last chapter, the intermittent transmission of feedback signals in both observer and controller was considered in this study. To promote the convergence of observation and control errors under the intermittent feedback, an event-triggered control method based on adaptive model estimation was proposed. Through constructing an adaptive model, the evolution of system state during the time flow was predicted, such that the state feedback compensation in the open-loop stage was formed to guarantee the convergence of the observer and controller. Besides that, the effect of triggered discontinuity on control design and analysis was also considered. In terms of eliminating the derivation singularity problem of the virtual variable with discontinuous feedback, a first-order filter with error compensation was designed into the control, which avoided the direct derivation of virtual variable. Meanwhile, a “soft” triggered updating mechanism was designed to realize the state updating with the tunable step size at the triggering instants, to reduce the discontinuous jumping-induced control chattering and instability. Based on the above design, the triggered condition with relative threshold and adjustable tolerance was further constructed to ensure the local uniformly boundedness of system and positivity of minimum execution interval. In addition, a self-triggered mechanism is presented theoretically to predict the triggering instant. Finally, simulation and hardware experiments verified the advantage of the proposed method in improving the reliability and flexibility of triggered control.
Keyword自适应神经网络控制 触发控制 机械臂运动控制 反步法 观测器估计 滤波控制
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/48850
Collection毕业生_博士学位论文
复杂系统管理与控制国家重点实验室_机器人理论与应用
Corresponding Author高洁
Recommended Citation
GB/T 7714
高洁. 基于新型神经网络与触发机制的机械臂自适应控制研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2022.
Files in This Item:
File Name/Size DocType Version Access License
gaojieThesisnew.pdf(12996KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[高洁]'s Articles
Baidu academic
Similar articles in Baidu academic
[高洁]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[高洁]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.