Maximum Likelihood identification of Wiener-Hammerstein system with process noise

Paper in proceeding, 2018

The Wiener-Hammerstein model is a block-oriented model consisting of two linear blocks and a static nonlinearity in the middle. We address the identification problem of this model, when a disturbance affects the input of the non-linearity, i.e. process noise. For this case, a Maximum Likelihood estimator is derived, which delivers a consistent estimate of the model parameters. In the presence of process noise, in fact, a standard Prediction Error Method normally leads to biased results. The Maximum Likelihood estimate is then used together with the Best Linear Approximation of the system, in order to implement a complete identification scheme when the parametrization of the linear blocks is not known a priori. The computation of the likelihood function requires numerical integration, which is solved by Monte Carlo and Metropolis-Hastings techniques. Numerical examples show the effectiveness of the identification scheme.