基于神经网络的自适应逆控制研究

【作者】 李仁俊；

【导师】 杨承志；

【作者基本信息】 昆明理工大学 ， 控制理论与控制工程， 2003， 硕士

【摘要】 控制理论的目标就是使得给定动态系统(对象)的特性尽可能准确地按照指定使用者的要求运行。这种对象控制问题可以分成三个独立的任务：对象的动态稳定性、对象动态控制和对象扰动控制。常规的控制方法是用反馈同时来处理和控制这三个任务，这时只能在三个方面进行折中处理以寻求控制的最优解。 自适应逆控制是一种分别处理这三种任务的控制方法。首先，对象是稳定的，对象不稳定的可以进行镇定。其次，对象是使用前馈控制器来控制的；第三，扰动控制器用来消除对象的干扰。自适应滤波器既可以作为控制器又可以作为扰动消除器，然后自适应算法分别优化使之达到各自的最优控制。 具体地说，自适应逆控制是用被控对象传递函数的逆作为串联控制器来对系统的动态特性作开环控制，从而避免了因反馈而可能引起的不稳定问题；同时又能做到对系统动态特性的控制与对象扰动的控制分开处理而互不影响。这是一个十分可贵的特点。且该控制器是自适应的，并将其调节到使对象及其控制器的总体动态响应达到最优。在其中使要用到反馈的，但反馈仅限在自适应过程本身采用。和传统的控制不一样，自适应逆控制采用反馈不是为了控制系统中的信号流动，而是用于控制系统中的可变参数。 以前自适应逆控制的研究主要集中于运用信号处理的方法尤其Wiener滤波器理论来建模与控制。这种方法对于研究线性系统来说是相当成功的。但对于非线性系统来说，还用这种方法来研究其效果不是太理想。尤其对于某类不能用精确数学模型描述的对象，其逆模型用传统的方法难以建立。而神经网络本质的并行结构在处理实时性要求高的自动控制领域中所显示出的极大的优越性使得逆控制在控制系统和调节器的设计中得到进一步发展。 作者把神经网络在非线性系统建模与控制中的优点引入自适应逆控制中来。从而提出一种基于对象—正模型—逆模型建模的神经网络自适应逆控制系统结构。我们采用基于BP算法的前馈神经网络构造对象辨识器和逆控制器。经过针对不同的对象，在环境及其参数变化时进行大量的仿真实验，证明了所设计的控制结构是合理和有效的，并且具有很强的鲁棒性。同时也对神经网络自适应逆控制系统的稳定性进行了初步的探讨与分析，最后我们对于神经网络自适应逆控制系统中值得关注的几个问题进行了分析与总结。 神经网络自适应逆控制是一种用神经网络构成无模型直接自适应控制系统的方法。具有对模型要求低、鲁棒性好、自适应能力强、适用于数字计算机控制昆明理工大学硕士学位论文中英文摘要的优点，将之用于工业过程控制中可以取得较好的控制品质。更多还原

【Abstract】 The goal of control theory is to make a given dynamic system (the "plant") behave in a user-specified as accurately as possible. This objective may be broken down into three separate tasks: stabilization of the plant dynamics; control of plant dynamics; and control of plant disturbance. Conventionally, one uses feedback to treatall three problems simultaneously. Compromises are necessary to achieve good solutions.Adaptive inverse control is a method to treat the three control tasks separately. First, the plant is stable or the plant is stabilized; secondly, the plant is controlled using a feedforward controller, thirdly, a disturbance canceller is used to reject plant disturbances. Adaptive filters are used as controller and disturbance canceller, and algorithms adapt the transfer functions of the filters to achieve excellent control respectively.Concretely, adaptive inverse control is an open loop control toward dynamic characteristic of system by acting inverse of plant transfer function as series controller, so it avoids instability as a result of feedback. At the same time, because the control of dynamic characteristic of system and plant disturbance is to be treated separately and that has not influence mutually. It is an important trait. Moreover this controller is adaptive and regulates collective dynamic response of plant and controller to achieve excellent control. Feedback is used in control, but it is just limited to adaptive process. Adaptive inverse control dominates alterable parameters rather than signal flow in system not so as conventional control.The research of adaptive inverse control mainly concentrated on modeling and control by utilizing the way of signal process especially Wiener filter theory ago. This method is fairly successful to research linear system. But it is not an ideal method for nonlinear system. Especially toward the plant without exactly mathematic model, it is difficult to exact inverse model for conventional method. However neural network with natural parallel structure shows huge advantage in the field of automatic control with high real-time. Consequently it promotes the more development of inverse control in the design of control system and regulator.Author introduces the advantage of neural network in modeling and control ofnonlinear system into adaptive inverse control. Consequently the paper puts forward a structure of neural network adaptive inverse control based on plant-positive model-inverse model modeling. Feedforward neural network based on BP algorithm constructs identification of plant and inverse controller. For different plants with the variety of circumstance and parameter, a mass of simulations testify the project structure of control to be reasonable and effective, besides strong robustness. At the same time the stabilization of adaptive inverse control system based on neural network is discussed and analyzed simply. Ultimately several problems deserved attention are analyzed and summarized.Adaptive inverse control based on neural network is a method of non-model direct adaptive control system constituted by neural network. It has an advantages of low modek good robustness , strong adaptation and is applicable for the control of digital computer. It may acquire the better control quality if adaptive inverse control based on neural network applies to industrial process control.更多还原