Development and implementation of a novel adaptive filter for non-linear dynamic system identification

Abstract

This paper presents a novel approach for the identification of nonlinear dynamic systems using dynamical filter weight neuron architecture. A sliding mode strategy is proposed for the synthesis of an adaptive learning algorithm for the neuron, whose weights comprise the first-order dynamic filters with adjustable parameters. This approach is known to exhibit robust characteristics and fast convergence properties. Experimental results on nonlinear dynamic systems, governed by difference equations, demonstrate the effectiveness of the proposed approach. Further, a meaningful comparison has been presented with a recent system identification technique that uses Karhunen-Loeve transform approach. The identification error comparison exhibits the robustness and reduced computational cost of the improved sliding mode filter weight algorithm over the later.