Stable and efficient neural network modeling of discrete-time multichannel signals

Abstract

This paper presents a neural-network-based recursive modeling scheme that constructs a nonlinear dynamical model for a discrete-time multichannel signal. Using the so-called radial-basis-function (RBF) neural network as a generic nonlinear model structure and the ideas developed in the classical adaptive control theory, we have been able to derive a stable and efficient weight updating algorithm that guarantees the convergence for both the prediction error and the weight error. A griding method developed in [11] based on the spatial Fourier analysis has been modified and applied for setting up the RBF neural net structure. Simulation analysis is also carried out to highlight the practical considerations in using the scheme.