【作者】 魏书宁；

【导师】 王耀南；

【作者基本信息】 湖南大学 ， 控制科学与工程， 2013， 博士

【摘要】 严重的高压输电线路覆冰会导致杆塔倾斜、倒塌、断线及绝缘子闪络，由此引起的线路跳闸、供电中断等事故给工农业生产和人民生活造成严重影响。采用机器人除冰具有无人员伤亡、无需停电和转移负载等优点，无需除冰作业时还可作巡检用途，其发展前景广阔。输电线除冰机器人工作在柔性输电导线上，在运行过程中需要翻越各类障碍物。受到环境风载等外部因素和机械振动等内部因素的影响，都可能造成越障过程中出现脱线情况，要实现除冰机器人自主抓线难度较大。常规的控制方法难以保证自主抓线控制精度，存在过于复杂、无法保证控制实时性等问题。所以，设计满足控制精度要求、简单可靠、实时性好、便于实现的机器人手臂抓线控制方法是除冰机器人关键技术之一。本论文围绕该技术开展了深入的研究，主要研究工作如下：1.论文对除冰机器人越障过程中抓线控制问题的难点进行分析，基于除冰机器人三关节手臂的结构特征，建立了三关节手臂的运动学和动力学模型，该模型在本文并且可在相关研究中得到应用。2.论文提出一类离散空间基于增强学习的抓线控制方法。根据经典增强学习控制方法可在线学习、易于实现的特点，论文提出基于Q学习、SARSA学习的抓线控制方法，并结合资格迹方法提出基于Q(λ)学习和基于SARSA(λ)学习的抓线控制算法。论文对所提算法进行了仿真实验和比较，实验表明基于经典增强学习的抓线控制算法是有效的，能够在多次迭代后找到“目标点”，能够解决外界恶劣环境干扰未知和手臂末端姿态的不确定性带来的控制问题。3.论文提出一类连续空间基于增强学习的抓线控制算法。针对经典增强学习算法对大规模和连续空间的优化决策问题难以保证算法收敛性以及存在学习效率不高的缺点，论文研究在输电线等效为蔓叶线模型的情况下，提出一类采用KNN算法结合资格迹的增强学习KNN-SARSA(λ)算法，实现连续状态-离散动作、连续状态-连续动作的抓线控制。仿真结果表明，基于KNN-SARSA(λ)的改进抓线控制算法能够解决二维空间内状态和动作输出的连续化表达问题，与传统增强学习控制方法相比，进一步提高了控制精度，具有良好的泛化能力和学习效率。4.论文提出一类基于迭代学习的除冰机器人轨迹跟踪控制方法。针对除冰机器人可以根据手臂末端与目标点的位置误差,采用抓线策略不断调整抓线手臂的动作，具有重复性的特点，提出一种鲁棒自适应迭代学习控制器，使之在PD控制器的基础上，随着作业任务的重复执行不断提高轨迹跟踪精度。该方法对处理器的计算和存储资源消耗低，可以实现干扰为不重复，包含线性化残差的不确定机器人动力学系统的鲁棒控制。仿真和实验表明该方法能够进一步提高轨迹跟踪精度。5.论文提出一类除冰机器人不确定项RBF神经网络逼近自适应控制方法。除冰机器人是一个非线性、强耦合复杂系统，控制难度较大。该方法采用计算力矩方法和神经网络补偿控制器相结合的控制方案，推导了神经网络权值自适应调整律，证明了系统的稳定性和误差的收敛性。补偿控制器的设计基于径向基神经网络，可以在线修正机器人模型误差，具有较好的适应性。仿真实验结果表明该方法有较好的轨迹跟踪性能和鲁棒性能。6.论文提出一类除冰机器人小波神经网络的鲁棒控制方法。该方法利用小波神经网络的强非线性学习性能来逼近除冰机器人系统的未知动力学部分，同时采用一个鲁棒控制器来补偿小波神经网络的逼近误差和外部干扰。该控制器能够有效降低模型不确定性和外部干扰的影响，减少了回归矩阵的计算，通过Lyapunov函数证明了控制系统的收敛性。仿真实验结果表明，该控制方法具有很强的抗干扰能力和很好的动态特性。更多还原

【Abstract】 Ice coating in high voltage power networks imposes heavy load upontransmission lines and could result in trip, disconnection, power-tower collapse andpower interruption, which has posed a serious of damages to economy. Adopting robotdeicing has advantages of avoiding casualties, temporal power failure and power loadtransferring. Furthermore, de-icing robot can be used for line inspection when there isno need for de-icing. For the above reasons, de-icing robot has good prospects.De-icing robot works on the flexible transmission line and need to cross variousobstacles on transmission line. Some external factors, e.g. strong wind, and someinternal factors, e.g. mechanical vibration could make the de-icing robot fail to grasptransmission line. Thus, it is of great difficulty for autonomous line-grasping ofde-icing robot. Conventional controlling methods, like PID control havedisadvantages due to low control precision, over-complexity and low real-timecapability. It is one of the key technologies for de-icing robot to design simple, robust,easy-realizable and real-time line-grasping control methods that satisfying controlaccuracy. This dissertation focuses on autonomous line-grasping control problem. Themain contributions are as follows:1. This dissertation analyses the obstacle crossing problem of de-icing robot andproposes kinematics and dynamics model of a three-link de-icing robot arm accordingto its structural characteristics. This kinematics model and dynamics model are usedin this dissertation and can be used in related research work.2. This dissertation proposes one type of discrete-space line-grasping controlmethods based on traditional reinforcement learning. Considering traditionalreinforcement learning methods can on-line study and is easy to implement, thisdissertation proposes line-grasping control methods based on Q-learning andSARSA-learning. Then by combing eligibity traces the line-grasping control methodsbased on Q(λ)-learning and SARSA(λ)-learning algorithm are proposed. Theproposed methods are evaluated and compared, experiment results show that thesemethods based on traditional reinforcement learning are effective and might adaptharsh environment, because the target point can be approximated in simulation aftersome times iterative computations.3. This dissertation proposes one type of continuous-space line-grasping control method based on reinforcement learning. Traditional reinforcement learning methodshave inevitable problems and learning efficiency is low for large and continuousspace. To overcome this limitation, after an equivalent cissoids model is deductedfrom transmission line model to facilitate the computation, one type of line-graspingcontrol KNN-SARSA(λ) methods of de-icing robot which combine the k-nearestneighbor algorithm and reinforcement learning are proposed. The proposed methodscan produce continuous-state-discrete-action and continuous-state-continuous-action.Simulations results show that these methods can solve the continuous representationproblem of state and action in two-dimensional space, with great generalizationability and learning efficiency.4. This dissertation proposes an adaptive learning control method for trajectorytracking of de-icing robot manipulator in an iterative operation mode. De-icing robotought to repetitively adjust line-grasping actions according to the position errors.Based on these characteristics, the proposed method consists of a classical PDfeedback structure and an additional robust adaptive updated term designed to copewith the non-repeated disturbances and unknown parameters. The controlimplementation is simple for the knowledge is not needed, and the only requirementon the PD and learning gains is the positive definiteness conditions. By usingLyapunov’s method, the asymptotic convergence of the closed-loop control systemcan be achieved. The simulation and experimental results of de-icing robotmanipulator are provided to verify the effectiveness of the proposed control method.5. This dissertation proposes a control method combining the well-knowncomputer torque method which is based on the known nominal robot dynamics, with acompensating controller which is based on the RBF neural network. This schemetakes advantages of the model based control approach and uses the neural networkcontroller to compensate for the robot modeling uncertainties, derives the adaptivelaw of the neural network. The neural network is trained on line which is based onLyapunov theory, thus its convergence is guaranteed. Simulation results are providedto demonstrate performance of the scheme.6. This dissertation proposes a WNN-based robust adaptive control method forde-icing robot. The bounds of the uncertainties are not necessarily known. A WNNsystem is used to approach the unknown controlled system, and a robust controller isdesigned to compensate for approximation error of neural network and externaldisturbances. It is shown that the proposed control scheme can guarantee estimationconvergence by Lyapunov function, reduce computation of regression matrix and the impact on model uncertainty and external disturbances. As demonstrated in theillustrated simulation, the control scheme proposed in this dissertation can achieve abetter model following tracking performance than the existing results.更多还原