【作者】 王良勇；

【导师】 柴天佑；

【作者基本信息】 东北大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 机械手系统在工业中广泛应用,是一类典型的非线性系统,其数学模型由运动学和动力学模型组成。机械手的任务一般在笛卡儿空间内给定,大部分机械手控制方法需要离线求解逆运动学方程,将机械手的任务分解到机械手各关节电机,然后进行底层的闭环控制。然而对于机械手的任务来说,这实际上是开环控制,使得机械手各关节的协调能力变差。虽然机械手的理论方法很多,但是很多算法都只是处于仿真研究阶段,尚未进行实验研究。另一方面当前的机械手实验平台可以分为封闭式结构和开放式结构两种,尽管开放式机械手实验平台解决了传统封闭式机械手平台不通用的问题,但是由于其算法开发环境采用的是Visual C++或其他面向对象编程语言,使得开发人员不仅要将主要精力放在算法开发上,还要在算法实现上花费大量的时间,因此需要开发一个适合高级控制算法实现的平台。针对上面提到的问题,本文依托东北大学“985工程”流程工业综合自动化科技创新平台,针对机械手系统中存在的一类非线性系统控制问题,设计了控制方法,并搭建了机械手快速原型控制实验平台,并将本文所提方法进行了实验研究。本文的主要研究工作归纳如下：(1)针对机械手系统中存在的非线性问题,提出三类神经网络控制方法：(a)考虑摩擦力的线性模型,机械手系统的动力学模型可看作为一类二阶严反馈非线性系统。针对此类系统提出了神经网络线性滑模控制器和神经网络终端滑模控制器,两种方法都具有PD控制加补偿器的结构,易于工业应用,而且采用神经网络终端滑模控制器还可以实现跟踪误差的有限时间收敛。(b)考虑摩擦力的线性模型,机械手系统的运动学模型和动力学模型可转化为严反馈多输入多输出非线性系统。针对此类系统,采用Backstepping的设计步骤,设计了神经网络串级控制器。该方法的优势在于：(i)与现有的严反馈非线性系统理论方法相比,该方法具有串级结构,这样在控制器投入使用时可以分级进行调试；另外控制器设计完成后,假如一个子系统发生变化,则只需重新设计相应的部分,无需重复整个控制器设计过程,改善了传统严反馈控制方法工业应用难的问题；(ii)与工业中广泛应用的串级控制器相比,该方法采用非线性的设计工具,提高了控制性能,同时可以严格证明系统的稳定性。(c)考虑摩擦力的非线性模型,机械手系统动力学模型可以转化为一类非线性输入输出离散系统。针对该类系统,本文提出了一种带有前馈和神经网络补偿的PD控制方法。该方法只需要系统的输入输出信号,不需要设计复杂的状态观测器,易于工业应用。(2)设计和开发了基于快速原型技术的机械手实验平台。该系统可以和Matlab/ Simulink实现无缝连接,控制算法开发、实时控制代码生成以及数据采集均可以借助于Matlab/Simulink实现。本实验平台的设计与开发为复杂算法的快速实现提供了一个有力的支撑。(3)将上面提出的三类神经网络控制方法在机械手实验平台上进行了实验研究,并且与机械手现有的控制方法进行了比较,实验结果表明本文方法取得了良好的控制性能。另外利用前面提出的严反馈多输入多输出非线性系统串级控制结构,解决笛卡尔空间机械手的轮廓跟随控制任务,提出了一种基于神经网络的轮廓跟随控制方法。该方法不仅提高了机械手各关节协调能力,而且还改善了现有机械手控制算法存在的“轨径缩减”问题,大大提高了产品的加工速度和精度。更多还原

【Abstract】 The robotic manipulator system which includes kinematic model and dynamics model is a class of nonlinear systems and widely used in industrial applications. The tasks of robotic manipulators are described in Cartesian space. Most of the reported control methods need to resolve the inverse kinematic model and transform the tasks into that of the motors of robotic manipulators. And then closed loop control in low level is performed. However, this means open loop control for tasks of robotic manipulators and would degrade the coordination ability of the joints of the robotic manipulator. There are many advanced control methods for robotic manipulators. However, most of them are only simulation cases and lack of experiment process on physics systems. On the other hand, the common control platform of robotics has two kinds:closed platform and open platform. Though the open platform has solved the problems of closed platform, the users have to spend much time to programme except for controller design since the development tools are VC++and other similar programme language. Therefore it is urgent to build a control platform to easily test complex controller.Supported by the "985" project process industry integrated automation innovation platform in Northeastern university, this thesis proposes control methods for robotics manipulators and builds a rapid control platform for robotic manipulators. Furthermore, the given methods are conducted on the presented platform.The main contributions can be summarized as:(1) Three kinds of neural network control methods are proposed to deal with the nonlinear problems in robotic manipulators.(a) The dynamics model of robotic manipulators with linear friction model can be transformed as a class of second order strict feedback nonlinear system. For such systems, the neural networks based linear sliding model control and the neural networks based terminal sliding model control methods are presented. The two methods possess a structure of PD control plus compensator, which can be easily applied in industrial applications. Moreover, finite time convergence of the tracking error can be realized for the neural networks based terminal sliding model control method.(b) The kinematic model and dynamics model of robotic manipulators with linear friction model can be transformed as strict feedback multi-input-multi-output nonlinear system. A neural networks based cascade controller is designed by a backstepping procedure for such system. The advantages of this method are listed as:(i) This method has a cascade structure. Hence, the controller can be put into practice step by step. Moreover, when a subsystem changes, only the related controller is redesigned without the need of redesigning the whole controller, which is easier to be applied than the reported methods for strict feedback systems. (ii) This method adopts nonlinear design tools. Therefore the control performance can be improved and the stability analysis can be critically addressed compared with the cascade control methods in industry.(c) The dynamics model of robotic manipulators with nonlinear friction model can be transformed as a class of nonlinear input output discrete system. PD controller including feedforward and neural network compensator is proposed for such system. If only the input and output signals of the system are known, the complex state observer will be not required, which is easy to be applied in industry applications.(2) A rapid prototype technology based experiment platform is designed and developed. The platform can be fully integrated with Matlab/Simulink. The development of control arithmetic, the real time control code, and data gathering can be performed by Matlab/ Simulink. Complex control methods can easily be carried out in the experiment platform.(3) Three kinds of neural network control methods proposed above are carried out in the experiment platform of the robotic manipulator. The detailed experiment results demonstrate the control performance of the proposed methods compared with other controllers. Moreover, the cascade structure is used to deal with contour following task of the robotic manipulator and a neural network based contour following controller is proposed. This controller can improve the coordination ability of the joints of the robotic manipulator and alleviate the "radius reduction" problem. Hence the machining speed and precision of production can be improved.更多还原