Fast model order selection of the nonlinearity in hammerstein systems

Abstract

In the paper we show a new three-stage algorithm identifying the Hammerstein system nonlinearity. The algorithm is designed to work when a poor a priori knowledge is available and when the measurement data set is small. In the first stage, a deconvolution routine is applied to output signal in order to diminish a (usually) harmful influence of the linear dynamic component. Such filtered output is then used in a standard nonparametric estimate (be it kernel or orthogonal series one) to recover the unknown characteristics. Finally, the results of nonparametric estimation are used in the algorithm selecting the best parametric model. The entire proposed approach is illustrated by the simulation example.