A Study Of T-S Fuzzy And Neural Network Based On A Clustering Algorithm

The rapid development of modern technology makes control theory develop in the direction with more complication and more acuration than before. Fuzzy control and neural networks control attract more and more people just because they have the charactors of special noneline and need not establish the mathematical model. Fuzzy system is good at expressing knowledge and its logical reasoning is similar to man's thought.But this system depends on the man's subjective factor too much and lacks the capacity of adaptation and learning.Neural network has a Variable construction and can be stored paralelly.To the most importance it has the characters of self-organization and self-learning.But its network parameters lack the physical meaning and easily trap in the local convergence at the same time.To combine these two system for absorbing their advantages seemly is a inevitable tendency. T-S model is a novel fuzzy reasoning model which replaces the parameters of traditional reasoning system with linear partial equation .Therefore, It can generate complicated nonlinear equation with fewer fuzzy rules.But the conclusional parameter is not a fuzzy rule but a linear equation,It can not be got from expert's experience and operational data directly.We must refine the parameter with some algorithm.By constructing T-S fuzzy neural networks we can solve the problem easily. But this model can't change the weakness in essence that it is difficult to withdraw the rules ,optimize the input space and theconstruction of system must be designated in advance. The model's application and development are restricted for this reason. In this paper we try to solve the problem. We put clustering algorithm into T-S fuzzy neural networks and withdraw system's character,optimize the input space in this method in order to set up a self-adaptive model.Firstly this paper analyzes and compares some types of typical clustering algorithms,and an improved algotithm is developed refering to this weakness in the meanwhile.The closest clustering algorithm is very simple and has a low wrong rate.In order to simplify the model,reduce calculation and improve the system's real-time responding capacity ,we construct the system in this method.Firstly this paper consider the influence of sample's outputs upon the model.Secondly the paper combines the nonsupervised algorithm with the algorithm based on gradient.Thirdly we adjust the distributing condition of the clustering points according to the density of data's distribution.Finally we construct a model based on T-S Fuzzy and Neural networks. This model can get rule's number ,confirm system construction and initialize original parameters according to clustering result automatically.The result parameters can be adjusted by BP algorithm.Focusing on function approximation problems we evalute its performance.In the process we can draw the conclusion that the new algorithm is better than the old one in adaptation ,computational complexity and the modeling accuracy. Finally this paper analyzes the stabilization of T-S fuzzy model.