【作者】 陈晓雷；

【导师】 李国勇；

【作者基本信息】 太原理工大学 ， 控制理论与控制工程， 2007， 硕士

【摘要】 工业生产过程往往具有非线性、不确定性，难以建立精确的数学模型，应用常规的PID控制器难以达到理想的控制效果。作为智能控制的重要分支，人工神经网络具有良好的非线性映射能力和高度的并行信息处理能力，已成为非线性系统建模、辨识和控制中极具魅力的理论和方法。本文在查阅大量国内外文献的基础上，研究了神经网络辨识的模型，基于单神经元PID自适应控制器，基于BP神经网络的PID自适应控制器以及神经网络自适应控制器。针对PID控制的不足，研究了神经网络与自适应控制的结合，以探讨收敛速度快、辨识精度高、实时性能好的神经网络控制方法。主要以研究各种算法的模型、实现及性能为基础，具体的内容如下：(1)以神经网络逆模型的原理、结构与算法的研究为基础，针对一类状态不可直接测量的非线性时变系统，给出了一种基于BP神经网络逆模型的状态观测器，可以对系统的状态进行实时观测，理论分析和仿真结果表明，这种状态观测器可以很好地观测系统的状态。(2)单神经元自适应控制有几个问题值得研究：①权系数初值的选择问题。权系数的初值对控制性能的好坏有很大影响②权系数的限幅问题。当权系数饱和时，系统将失去应有的学习能力。③神经元控制系统的上升时间比较长，受到扰动以后动态恢复过程较长。针对上述缺陷给出两种改进方法：第一利用遗传算法来优化单神经元的权值从而实现对系统的控制。结果表明该方法能够很快地搜索到一组较优的权系数，消除了初值对控制性能的影响，加快系统收敛速度，具有较好的动态性能和鲁棒性；第二基于CMAC的单神经元复合自适应PID控制器，使CMAC网络的学习过程包括整个系统控制过程，仿真结果表明该方法具有自适应能力强、实时性好、抗干扰能力强等优点。(3)分析了基于BP网络的自适应PID控制的模型、算法及特点。将神经网络用于控制器的设计或直接学习计算控制器的输出，一般都要用到系统的预测输出值或其变化量来计算加权系数的修正量。但实际上，系统的预测输出值是不易直接测得的，通常做法是建立被控对象的预测数学模型。所以为了提高控制效果，需要建立合理的模型来计算预测输出。本文利用最小二乘法和神经网络建立被控对象的预测数学模型，用该模型所计算的预测输出取代预测输出的实测值，对基于BP网络的自适应PID控制器的权值调整算法进行改进。仿真结果表明算法的有效性。(4)针对一类未知、不确定、时变的SISO离散非线性系统，利用神经网络对被控对象的正向模型辨识，将神经网络的输出作为被控对象输出的预报，在此基础上设计出控制律，构成神经自校正控制方案。仿真结果表明控制算法的有效性。(5)利用神经网络逆模型辨识的思想，提出一种神经网络模型参考自适应控制器设计方案，并给出设计步骤与算法，它适用于任意非线性系统，更接近于工程实际。理论分析和仿真结果证明了该方案的合理性和有效性。更多还原

【Abstract】 The process of industrial production is sometimes nonlinear anduncertain and is difficult to establish accurate mathematic module, sousing routine PID controller is hard to achieve perfect controlling effect.Neural network (NN) has favorable nonlinear mapping performance andhigher parallel information processing capability, and it has been the mostengaging theory and method for nonlinear system modeling, recognitionand control.On the base of consulting literatures both here and abroad, themodel of NN identification, PID controller based single-neuron, PIDadaptive controller and the NN adaptive controller are discussed. Aimed atthe shortages of PID, the connection between NN and adaptive controllerare researched in order to accelerating convergence-speed and theaccuracy. This paper are based on the researching of the model,realizations and the performance of the algorithms. The main task can beconcluded as follows:(1) Aimed at the a sort of Non-linear system, the paper proposed an state observer based on the BP NN inverse-model, this model can observethe state of system in real time. And the analysis of theory and simulationshowed that the state observer is very good.(2) There have some problems in the single-neuron adaptivecontroller:①the choosing of weight coefficient. The initial value of theparameter has strong effect on the control performance②The slicingproblem of weight coefficient. when the parameter is impregnated, thestudy ability will lost.③The rise time is long, anti-jamming capability isinfirmness.In order to solve the problem, two improved method are proposed:the first is using GA to optimize the parameter, and the simulation showedthe way can find the best parameter and eliminating the effect of initialvalue. The second is a compounded adaptive PID controller based onCMAC. the study process of CMAC involved the total control process ofsystem, the simulation showed the method had the strong adaptive abilityand anti-jamming capability etc.(3) Analyzed the model, algorithm and characters of the adaptivecontroller based on BP NN. In order improve the effect on control, theappropriated model must be seted. Using the idea of NN contrary modelrecognition, the text put forward a designing scheme that NN model makes self-adaptive controller as reference and lists the designingapproach and arithmetic. It applies to any nonlinear system, so it is closerto engineering practicality. The theory analysis and emulate result provethe rationality and validity of this project.(4) Aiming at a kind of unknown, uncertain and time-variant SISOdiscrete nonlinear system, the text uses the forward model recognition forNN to the controlled object and makes the output of NN as prediction ofthe output of controlled object, and based on this the text works out thecontrolling rule and forms the project of neural self-revise control. Theresult of emulation indicates the controlling arithmetic is effective.(5) Using the idea of NN contrary model recognition, the text putforward a designing scheme that NN model makes self-adaptive controlleras reference and lists the designing approach and arithmetic. It applies toany nonlinear system, so it is closer to engineering practicality. The theoryanalysis and emulate result prove the rationality and validity of thisproject.更多还原