The Method Of Fuzzy Rules Extraction Based On Fuzzy Neural Network

Fuzzy control and fuzzy system modeling have been used widely with the development of almost 40 years. However, owing to the fuzzy control theory is not perfect, there are many problems are in the research and exploring phase. Among which, fuzzy rule automatically generating is one of research hotspot in fuzzy control fields. Moreover, both objective definition fuzzy language and modifying the membership functions real time are the two main matters in the process of fuzzy rules automatic extraction.This thesis aimed at above problems puts forward an approach of fuzzy rules automatic extraction according as measure data and propose a RBF net structure, which can be good for description the fuzzy rule interpretable characteristic.The thesis is mainly composed of four parts.In the first part, the thesis expatiate the status in quo and development of fuzzy control in brief.In the second part of the thesis, we mainly discuss the method and process of input and output automatic division. First, fill in vacant values of sample data via spline insertion. Then, we use a method -based on density and grid which often be employed in data mining automatically generating the number of effective clustering according the sample data after insertion; After that, applying K-average value cluster the sample data, realize the input and output space automatic partition. This method solves the lack of determining partitions by person's experience beforetime. At last, dispersed the data based on the projections of sample space clustering on every dimension, offered the objective foundation for the fuzzy subset definitionof each variable in the fuzzy system.Based on above work, in the third part, actualize fuzzy rules automatic generation applying fuzzy RBF neural net. First, according to analyze and compare the net structure in common use, bring forward a new RBF net structure that not only clearly express the partition of input -output space but also directly offer the description form of fuzzy rules structure. This net structure enhances the descriptive ability to of fuzzy system. After that, discuss in detail the process of deciding the net construction and parameter, learning algorithm and language C code.At last, take as an example, verity the effectiveness of model through 500 and 1000 group sample respectively learn. The average relative error is smaller than 6%.The innovations of this thesis are as follows:l.In this paper, we realize the number of valid clustering have been automatically ascertained according to apply the data mining method, which solve the question of K-average value clustering man-made determine k in advance.2. A new RBF neural network structure is proposed that is able to express effectively the fuzzy system explainable character.