hrvatski

Identifikacija nelinearnih sustava dinamičkom neuronskom mrežom

ABSTRACT-In this Ph.D. thesis an approach to identification of nonlinear dynamic systems with dynamic neural network is presented. An attempt to establish a nonlinear dynamic discrete-time neuron model, called Dynamic Elementary Processor (DEP) has been done. This dynamic neuron disposes of local memory, in that it has dynamic states. Based on the DEP neuron, a Dynamic Multi Layer Perceptron Neural Network is proposed. The momentum method is applied in order to accelerate the convergence of proposed extended dynamic error back propagation learning algorithm. The main advantage of proposed dynamic neuron model is that it reduces the network input space. Further it offers a great potential in solving many problems that occurs in system modelling with a special emphasis on the systems with characteristics such as nonlinearity, time delays, saturation or time-varying parameters. The proposed supervised learning algorithm is tested in prediction of an nonlinear chaotic system, known as Glass-Mackey time-series. As an another application of the proposed Dynamic Neural Network (DNN), the identification of a dynamic discrete-time nonlinear system whose measurement data are spoiled with noise is performed. DNN is also used for identification of the laboratory model of an air-heater, developed with the aim to educate students and researchers in the field of automatic control. Finally, this neural network is trained to imitate an adaptive nonlinear robot control algorithm based on inverse dynamics of the full robot model of RRTR structure. The learning results are presented in terms that are insensitive to the learning data range, and allow easy comparison with other learning algorithms, independent on machine architecture or simulator implementation.

dinamičke neuronske mreže; identifikacija dinamičkih sustava

nije evidentirano