论文标题:神经网络自适应控制理论及其应用
Neural Network Based Adaptive Control Theory and Its Application
论文作者 唐英干
论文导师 关新平,论文学位 硕士,论文专业 控制理论与控制工程
论文单位 燕山大学,点击次数 148,论文页数 103页File Size2865k
2001-09-01论文网 http://www.lw23.com/lunwen_44356332/ 神经网络;自适应控制;滑模控制;机器手;混沌;同步
neural network;adaptive control;sliding mode control;robotic manipulator;chaos;synchronization
非线性系统广泛存在于实际系统中，它的建模与控制一直是控制理论中的难点。由于经典控制理论只适用于线性系统；而非线性系统的相平面法和描述函数法只能分析简单的非线性系统；几何理论则需要系统的精确的数学模型，这给非线性系统的研究带来了很大的困难。因此，必须寻找新的工具和方法来研究非线性系统的控制问题。 本文利用神经网络的全局近似能力，研究了仿射非线性系统的自适应控制问题。基于Lyapunov稳定性理论，推导出了神经网络权值在线学习规律，保证了系统的全局稳定性，同时，为了消除或减小外部干扰以及神经网络逼近误差，在控制器设计时结合了鲁棒控制方法，使得系统具有一定的鲁棒性。最后考虑了这些控制方法在机器人控制和混沌系统同步中的应用问题。具体内容如下： 研究了一类单输入单输出仿射非线性系统的自适应跟踪问题。主要结果有：1)针对一类未知的单输入单输出仿射非线性系统，提出了一种基于神经网络的滑模自适应控制方法。该方法利用神经网络学习系统中的非线性函数，神经网络的权值由Lyapunov稳定性理论导出，并且在线调整；考虑到网络逼近误差和外部干扰的存在，文中利用滑动模态对参数和扰动不敏感的特点，实现了系统的鲁棒输出跟踪。2)针对一类未知的单输入单输出仿射非线性系统，提出了一种基于神经网络的H_∞自适应控制方法。该方法针对滑模控制方法由于控制律不连续容易引起系统的抖颤的不足，设汁了一种新的控制器，即H_∞控制器。所设计的控制律连续并且使得系统具有H_∞跟踪性能。3)针对一类不确定性非线性系统，利用神经网络来学习实际模型和系统标称模型之间的偏差，采用类似2)的方法，提出了一种H_∞自适应方法。该方法有效地克服了几何方法要求不确定性满足匹配条件的限制，同时利用系统已知的信息，减轻了神经网络的学习负担，加快了神经网络的训练速度，使得控制律更加简洁，便于工程实现。 研究了一类多输入多输出仿射非线性系统的自适应跟踪问题。这一部分的工作主要是单输入单输出情形的推广，在这里得到了类似单输入单输出情 蒸山大学工学硕士学位论文形的结果。 研究了n－自由度机械手的轨迹跟踪问题。针对计算力矩法对模型误差敏感，鲁棒性差的特点，文中将神经网络和计算力矩方法结合，首先根据标称模型设计计算力矩控制器，然后采用神经网络来学习系统中的不确定性，神经网络的输出作为补偿控制器：有效的克服了机械手由于工作环境和负载变化等不确定性引起的控制品质的破坏。 研究了扰动情况下混炖系统的同步问题。在同步的两个混饨系统都存在扰动情况下，利用神经网络学习系统中的不确定性和扰动，神经网络的输出作为补偿器，实现了两个系统的鲁棒自适应同步。有效地克服了扰动对同步的破坏。
In many practical engineering problems, the model of the controlled system vary with the change of time and working environment. If the parameters of the controlled system vary in large range, the controlling performance will be badly destroyed. Moreover, it leads to instability. So, it is important to make the system be adaptive, that is, the system can change the control parameter or control action according to the change of parameter or control index such that the system work at the best state. Proper speaking, many systems in control engineering are nonlinear and linear system is only a special case of it. Because of the complexity of nonlinear and the uncertain in the system, the study of nonlinear system is very difficult It is important to seek new tools and new approach to study nonlinear system." In mis thesis, the problem of adaptive control for affine nonlinear system is studied based on the global approximation of neural network. Based on Lyapunov stability theory, the on-line algorithm of network weights is derived and guarantees the global stability of system. The proposed approach overcome the disadvantages which the neural network need learning off-line before using as a controller and lack of real-time and can not guarantee stability of the system, hi order to eliminate the extern disturbance and approximation error of network, the controller design combine with robust control approach and the system possess robustness. At the same time, the application of the proposed approach is considered in this paper, that is, in robotic control and synchronization of chaotic system, the main results contain:Deep investigations on adaptive tracking problem for a class of single input single output affien nonlinear system. The main results contain: (1) An adaptive sliding mode control approach is proposed for a class of unknown SISO affine nonlinear system based on neural network. In this approach, the neural network is used to learning the nonlinear function of the system. The network weights are derived using Lyapunov-based design and are adapted on-line. Due to the existence of neural network approximation error and external disturbance, thesliding mode control which is insensitive to disturbance and parameter pertabation is used to achieve robust tracking for the system. (2) An adaptive Hx control approach is proposed for a class of unknown SISO affine nonlinear system based on neural network. This approach overcomes the disadvantages of chattering of the sliding mode control approach. The control law is continue and the system possess Hx tracking performance. (3) An adaptive Ha control approach is proposed for a class of uncertain nonlinear system. The neural network is used to learn the error between the real model and the nominal model. This approach relax the restriction of matching condition that geometric control approach need. Meanwhile, the learning burden is cut down and speed the learning process, so, the control law is simple is easy to apply to engineering.The adaptive tracking problem is studied for a class of MIMO affine nonlinear system. In this section, the main work is that some main results of SISO is extended.The application problem of neural network based adaptive control theory is studied for real model. First, n-degree of freedom robotic manipulator control problem is studied. On the basis of computed torque (CT), the uncertainty is learned by the neural network, and the output of the neural network is used as compensator. This approach overcome the deterioration of control performance due to the uncertainties including working circumstance and load change. Second, the synchronization of two chaotic system that exposure to disturbances is considered. Robust adaptive synchronization is achieved using neural network for two chaotic systems. The results show that the approach proposed in this section can overcome the effectively the disruption of perturbation.
|
