基于粗集与聚类的神经模糊建模

【作者】 蔡志峰；

【导师】 张阿卜；

【作者基本信息】 厦门大学 ， 控制理论与控制工程， 2002， 硕士

【摘要】 在现代复杂工业控制中，由于被控对象通常具有多变量、严重的非线性、强耦合、大滞后、分布参数、时变以及种类繁多的干扰，使得基于精确数学模型的常规控制方法已无法获得满意的动静态控制效果。近年来，基于神经网络和模糊逻辑的神经模糊控制得到了广泛的应用。在众多的神经模糊建模方法中，数Jang于1994年提出的ANFIS-自适应神经模糊建模方法最为突出，它的自适应性质使得它几乎可直接应用于自适应控制和学习控制；事实上，它可以替代控制系统的任意神经元网络并执行同样的功能。然而它和绝大多数神经模糊建模一样总是不可避免地存在着这样一个难题，即结构辨识问题，也就是如何合适地划分输入输出空间，如何从观测数据中提取出较为简化的模糊规则库，避免规则库爆炸问题。 由Z.Pawlak提出的粗糙集理论是一种刻划不完整性和不确定性的又一个强有力的数学工具，能有效地分析不精确(imprecise)、不一致(inconsistent)、不完整(incomplete)等各种不完备的信息，还可以对数据进行分析和推理，从中发现隐含的知识,揭示潜在的规律。 本文在学习已有文献的基础上，结合作者三年来的学习与实践，利用聚类算法较好的数据组织与分类能力以及粗集理论强大的数据分析与分类能力，分别提出了两种从系统的观测数据中提取模糊规则的方法，构造类似于ANFIS模型的SCANFIS模型和RSANFIS模型，使得规则库的规模不再随系统的输入维数呈指数增长，一定程度上解决了维数灾难问题，为复杂的多输入系统的建模提出了一种切实有效的建模方法。具体的仿真试验表明这两种建模方法对于复杂的多输入系统的建模，在效率与精度上比ANFIS模型及其他模型都要好。更多还原

【Abstract】 In the control of complicated modern industrial plants,because most of the plants have many input variables,high non-linearity,tight coupling,serious transmission lag,distributed parameters,time -varying and many kinds of disturbances,the conventional control methods based on the exact mathematical model can not achieve both approving dynamic and static results. Recently,Neuro-fuzzy Control base on the Neural Network Theory and Fuzzy Logic System was used widely and successfully. In most of these Neuro-fuzzy modeling methods,the ANFIS (Adaptive -Network-based Fuzzy Inference Systems) method which was proposed by Jang in 1993 is the most prominent one,its adaptive property made it possible to be used in adaptive control and learning control directly. In fact,it can replace any Neural Network of the control systems and carry out the same function. However,it has the same problem as most of the Neuro-fuzzy systems - rule explosion problem. It is a difficult problem relates to the structure identification of the system and concerns with the partition of the input and output space and the rule generation of the raw sample data.Rough set theory,introduced by Zdzislaw Pawlak in the early 1980s,is another new powerful mathematical tool to deal with vagueness and uncertainty. It can analyze the imprecise,inconsistent and incomplete information effectively,find out the connotative knowledge and detect the potential rule of the system under consideration.In this paper,combine with what the author learned and studied during the three years,using the Cluster Algorithm and the Rough Set Theory,the author propos two new rule generation methods and put forward two new Neuro-fuzzy modeling methods with the T-S type of rule consequent,which named SCANFIS and RSANFIS respectively. Both methods can get over the rule explosion problem and confine the size of the rule base to a reasonable one. They give effective solutions to the modeling of the complicate multi-input systems. The simulation results show that both methods are more powerful than ANFIS method in modeling complicate multi-input systems concerning with efficiency and precision of the model.更多还原