Research Of System Identification Based On Type-2Fuzzy Neural Network

Neural Networks(NNs) have the advantages of self-adaptive learning and nonlinear mapping. And the use of Fuzzy Logic System(FLS) is due to the ability of processing fuzzy information. The combination of Neural Networks and Fuzzy Logic System has become a topic research direction. They make the Fuzzy Neural Networks(FNNs) become a powerful tool of nonlinear system identification. In this paper, research of type-2fuzzy neural network for nonlinear system identification is done as follows:(1) For nonlinear dynamic system identification problem, we designed a novel structure named a Recurrent Type-2Compensatory Fuzzy Neural Network(RT2CFNN). The usage of type-2fuzzy set makes the system deal with uncertainties effectively. And the compensatory operation makes the inference system more flexible to optimize the fuzzy reasoning. The design of recurrent factor makes the system have the ability of memory. A method is given to determine the initial parameters of the antecedent part of fuzzy neural network based on Fuzzy C means clustering algorithm and its interval of the fuzzy degree. We deduced the learning algorithm based on the gradient descent algorithm. The stability and convergence of the proposed algorithm is analyzed and we give out an optimal learning rate theorem. The results of experiments show that the novel learning algorithm can efficiently improve the accuracy of dynamic system identification while using less rules and iterations.(2) For the problem of instability of identification result caused by the random production of the consequent part initial parameters of the fuzzy neural network, we proposed a type-2fuzzy neural network based on a Hybrid learning algorithm. It uses the recursive least squares with forgetting factor(RLSFF) to initialize the consequent part parameters. This method makes the identification result stable while accelerates the learning speed, and to some extend it can avoid falling into local optimum. The result of experiment shows that the further modified method improved the modeling accuracy and the modeling result is stable.(3) We applied our proposed algorithm to the identification of the TE process. The result of experiment shows that the proposed algorithm has better predictive performance than BP network and type-1TSK FNN. It further verified the effectiveness of the proposed algorithm.