Separation methods for dynamic errors-in-variables system identification

Unlike standard output error models, both input and output observations of errors-in-variables systems are corrupted with noise. As practical applications often fall into the errors-in-variables category, estimation methods that can simultaneously derive model and noise parameters are of particular interest. In this paper, we explore separation mechanisms applied on the combined input and output observation matrix to partition observations into two sets, after which it is possible to perform parameter estimation over the individual sets, and the estimates may, in turn, be compared. Minimizing the distance between parameter estimates, it is shown that we may infer a noise structure. Once a noise structure estimate is at hand, a maximum likelihood estimation may yield model parameter estimates.