Research On Learning Algorithm For Fuzzy Neural Networks Based On Extended Kalman Filter

The fuzzy system has the function of simulating human brain's reasoning; and the neural network has adaptive learning ability, generalization ability and parallel processing ability. The fuzzy neural network not only combines the advantages of fuzzy system and neural network, but also can overcome each of their shortcomings, so it becomes a hot research issue in intelligent control field. Because fuzzy neural network is simple and practical, it has been widely used in industrial process control,signal processing and other fields. Inherently, a fuzzy neural network is a neural network, and its structure is mainly determined by the number of fuzzy rules. The structure of a network has an important impact on its generalization ability and it is still difficult to determine the structure of a fuzzy neural network is rational or not. A fuzzy neural network is usually trained with the back propagation algorithm. The back propagation algorithm has a lot of limitations, such as slow learning speed, easily falling into local minimum points. Kalman filter,a kind of real-time recursion algorithm, being used to deal with random signals, it primarily takes advantage of the statistical properties of system noise and observation noise, and it takes the measured values as the input and the estimated values as the output. The input and the output are connected by the time update algorithm and the observation update algorithm, the signals that needed to be processed are estimated in the light of the state equation and the observation equation, Kalman filter is essentially an optimal estimation method. It can not only reduce the network learning cycle, but also optimize the network structure by using Kalman filter to adjust of the parameters of the neural network. In order to construct an effective fuzzy neural network, a self-organizing learning algorithm is presented in this dissertation, and the structure identification and parameters adjustment of the fuzzy neural network can be conducted simultaneously in the proposed algorithm. The main contents of this dissertation are as follows:â‘ The basic theories of fuzzy systems and neural networks are introduced briefly, and their properties are analyzed.â‘¡The basic principles of the discrete and extended Kalman filter are elaborated; The calculation processes of the two kinds of filters are deduced; the applications of the global extended and decoupled extended Kalman filter in neural networks training are studied. â‘¢A self-organizing learning algorithm for fuzzy neural networks is constructed, system errors, accommodation boundaries and error reduction ratios are considered in establishing the growing criterion of the fuzzy rules, the network restrictively adds fuzzy rules according to the growing criteria. Because the former part of the network is nonlinear and the rear part of the network is linear, the extended Kalman filter is adopted to adjust the premise parameters of the network and the Kalman filter is adopted to adjust the consequence parameters of the network in the process of adjusting the parameters. The simulation results show that the resulting fuzzy neural networks with the algorithm have compact structures and possess the advantages in the aspects of approximation accuracy and generalization ability.