Lavrynenko K. Neural network identification of nonlinear dynamic plants on the multilayer perceptron basis

The thesis is devoted to the solution of neural network identification problem for nonlinear dynamic plants and is based on the multilayer perceptron in apriori and flowing indeterminacy conditions concerning researched plants and influencing noises properties. Neural network models are realized on the multilayer perceptron (MLP) basis. The existing procedures of MLP parameters adjustment are analyzed and the algorithm of its learning is synthesized on the basis of the spread Kalman filter. Is offered to realize MLP learning with the help of multistage projective procedures, which use a circumscribed amount of the information. With the purpose of improving computing properties of the given procedures and resistance of learning process are developed them factorized form. The exposition of nonlinear dynamic plant with the predictors help is considered on the basis of a Kalman filter and the appropriate neural network models in spacious condition are obtained. Is carried out a simulation of different learning MLP algorithms with the help MATLAB 6.1, Trajan Neural Network Simulator 3.0 and IntelligentPad.