Identification of Hammerstein System with Unknown Order by Neural Networks

Abstract

This paper proposes new identification schemes with neural networks for a Hammerstein system which consists of a nonlinear memoryless element followed by a linear dynamical subsystem. The nonlinear element are approximated by the artificial neural networks and the linear subsystem is expressed by the transfer function model. The connection coefficients and thresholds of the neural networks and the parameters of the linear subsystem are estimated by the renewal back-propagation method. The structure of neural networks and the order of the linear subsystem are selected by the minimum description length criterion. It is demonstrated in digital simulation that the proposed identification technique is efficient for the system with various types of nonlinearities.