A Time-Domain Fractional Approach for Wiener-Hammerstein Systems Identification

Abstract

This paper describes a new approach to initialize Wiener-Hammerstein models for iterative prediction error minimization. The key idea is to parameterize the division of poles and zeros of the best linear approximation (BLA), between the two linear subsystems. Taylor expansion is used to handle the parameterization in the time-domain. Results regarding uniqueness of the initial estimate are proved on a low order example. The initial estimate generated with this approach allows to avoid problems with local minima and it is used for the final optimization in all model parameters.