A stable neural network-based identification scheme for nonlinear systems

Abstract

This paper presents a stable neural identifier for multivariable nonlinear systems. A state-space representation is considered based on both parallel and series-parallel models. No a priori knowledge about the nonlinearities of the system is assumed. The proposed learning rule is a novel approach based on the modification of the backpropagation algorithm. The boundedness of the identification error is shown using Lyapunov's direct method. As a case study, identification of the dynamics of a flexible-link manipulator is considered to demonstrate the effectiveness of the proposed algorithm. Simulation results for a two-link planar manipulator and the Space Station Remote Manipulator System (SSRMS) are presented.