【作者】 张昌凡；

【导师】 王耀南；

【作者基本信息】 湖南大学 ， 控制理论与控制工程， 2001， 博士

【摘要】 对交流传动等快速变化的非线性复杂系统，人们已提出了各种控制方案，这些方案解决了一些问题，但还存在一些不足之处。因此对快速变化的复杂工业系统，研究一种工程实用的有效控制方法是有待解决的问题。 滑模变结构响应快，对系统参数和外部干扰呈不变性，可保证系统是渐近稳定的。尤其可贵的是其算法简单，易于工程实现。其缺点是存在抖动和需知不确定参数上下界等问题。模糊控制无需建立数学模型，控制的机理符合人们对过程控制作用的直观描述和逻辑思维。其缺点是设计缺乏系统性，控制规则的选择多采用试凑法。神经网络具有较强的自学习能力，可以充分逼近任意复杂的非线性。其缺点是学习速度较慢，难以控制较复杂的对象。 由此可见，变结构控制、模糊控制和神经网络控制相互之间具有很强的互补性。而研究这种互补性，实现滑模变结构的智能控制就是本文研究的目的。 本论文的研究内容包括以下方面： 1、研究了模糊系统、神经网络的特点，对滑模变结构的发展现状进行了较为详细的评述，分析了交流传动的控制策略。 2、对电流源逆变器(CSIM)构成的交流变频调速系统，首先分析了CSIM在低速存在的问题，利用滑模变结构对参数变化的不灵敏性，推导出CSIM的简化数学模型，并以常规滑模变结构理论为指导，设计了CSIM控制器。最后利用微型计算机实现了该控制系统，并给出了实验和仿真结果。 3、提出了一种基于模糊神经网络的滑模变结构控制。首先通过模糊神经网络在线调控符号函数项的幅值，利用Lyapunov稳定性定理推出保证控制系统稳定的模糊神经网络学习率取值范围；对于大范围参数摄动，研究了一种全参数调节的模糊神经网络滑模控制；提出了一种神经网络滑模鲁棒控制器设计方法，以使系统在任意初始条件下都处于滑动状态，实现运动过程中的全程滑模控制。 4、针对神经网络学习速度收敛较慢的问题，重点研究了提高神经网络收敛速度的方法。首先采用一种快速的变尺度优化学习算法，对不确定系统进行逼近，提出了一种基于系统辨识的滑模变结构控制方法；结合单层Adaline 划曰大学烬士学馒诀文网络，研究了一种具有滑模变结构的权值训练法，提出了一种神经网络自适应滑模变结构控制，对它的稳定性和收敛性进行了深人探讨；研究了一种基于神经网络的自学习滑模变结构设计方法，对三层自适应神经网络结构、算法进行了分析，最后给出了仿真实验结果。 5、主要研究基于观测器的滑模变结构鲁棒智能控制。针对离散多变量系统，对系统确定性部分设计了一个Luenberger观测器，然后用一个神经网络来动态补偿系统的不确定性。为了进一步提高观测器的观测全局性，设计了一个在线神经网络滑模参数调节器：针对一类仿射非线性系统，本文还提出了一种智能滑模变结构状态观测器的设计方法，并对系统的鲁棒特性进行了研究。更多还原

【Abstract】 Many control programs have been proposed for the fast-changing nonlinear complex systems like the AC drives which have solved some problems of these systems, but there are still some shortcomings in those programs. Therefore, It is necessary to develop a practical and efficient controller for the fast-changing complicated industrial systems. The sliding mode variable structure is able to response quickly, invariant to systemic parameters and external disturbance, and able to keep the system asymptotic a stable state. What is more precious is that its algorithm is simple and it is easy to be realized engineering. But it has such shortcomings as chattering arid the requirement of the knowledge of the upper and lower bounds of uncertain parameters. The fuzzy control requires no mathematical model arid its mechanics are agreeable to people’s logical thinking and people’s direct description of the process control, but, it is not systematic in design, its control rule selection is conducted with the trial approximation method. The neural network is powerful in its self-learning abilities. It is able to approximate fully any complicated nonlinear state, but it is difficult to be applied to the control of comparatively complicated objects because of its slow learning speed. It can be concluded that the variable structure control, fuzzy control, and neural network control are greatly complementary. The aim of this research is to go into the complementarities of the three and realize the intelligent control of sliding mode variable structures. The content of this research falls into the following five parts: 1. The characteristics of the fuzzy systems and the neural network are studied. It gives a detailed comment on the present development of the sliding mode variable structure and analyzes the controlling strategies in AC drive. 2. Directed to the alternating speed regulation controller composed of the current source inverter motor(CSIM), this paper first analyzes the problems of CSIM at a low speed, infers the simplified mathematical model of CSIM on the basis of the insensitivity of the sliding mode variable structure to parameter variations, and then designs a CSIM controller in light of the routine sliding mode variable structure theory. Finally, it realizes the control of the system on the microcomputer and shows the experimental and simulation results. 3. A sliding mode variable structure control is introduced on the basis of the fuzzy neural network. The essay first adjusts on-line the value of the symbolic function terms with the fuzzy neural network, then infers the learning rate range of the fuzzy neural network in which the stability of the controller is warranted by using the Lyapunov stability theorem, and then develops a fuzzy neural network sliding mode controller capable of all-parameter adjustment for the great parameter change, and introduces a design method for a neural network sliding mode robust controller which keeps the system in a sliding state in any initial conditions and realizes the omnidistance sliding mode control over the movement of the system. 4. In view of the slow convergence of learning speed of the neural network, this research lays its emphasis on the development of a method to raise the converging speed of the neural network. It first uses a fast scale-change optimal learning algorithm to approximate the uncertain system and propose a new sliding mode variable structure control method based on systemic identification. Then, in combination with更多还原