A NEW NEURO-FUZZY IDENTIFICATION MODEL OF NONLINEAR DYNAMIC-SYSTEMS

Abstract

Multilayer neural networks with error back-propagation learning algorithms have the capability of learning an arbitrary continuous nonlinear function with examples of input and output sample pairs and a great potential for identifying nonlinear dynamic systems with unknown characteristics. A fuzzy system is composed of fuzzification of input, reasoning (or inference by fuzzy rules, and defuzzification of fuzzy output. In general, there are some difficulties in finding suitable fuzzification and defuzzification methods and fuzzy rules. Formation of fuzzy rules with complex input-output relationships can be replaced by building neural networks with input and output sample pairs. A neuro-fuzzy identifier is proposed to have a cascaded structure of fuzzification, neural network, and defuzzification and additionally is shown to be able to compensate for fuzzification and defuzzification en or of fuzzy logic. Computer simulation shows that neuro-fuzzy identification is very effective in modeling the fuzzy system whose fuzzy rules can not be obtained easily.