Control relevant model identification with prior knowledge

Abstract

<p>The use of prior or accumulated knowledge for the identification of multivariable models that will generally assure the stability of multivariable model based controller designs is investigated. The systems that are considered in this study are by and large linear, time invariant, ill-conditioned, and multi-input multi-output (MIMO) in nature. The effect of different types of prior information on controller stability is studied. It is shown that use of some types of prior knowledge may improve the model quality in terms of the stability of the resulting closed-loop system, while use of other types of prior knowledge may degrade the model quality. Prior knowledge that provides information about the low-gain direction of the process has the most significant effect on controller stability. Several issues associated with incorrect prior knowledge and the sensitivity of controller stability to such an error are also considered. This leads to checkable metrics that can be used by practitioners to evaluate the sensitivity of the controller to given prior knowledge before controller implementation. The issue of model maintenance (that is re-identification of existing models) that will result in improved controller stability in MIMO controllers is then addressed. Posterior knowledge about existing controller performance can be used to re-estimate models. Two novel controller designs result from this study: a multi-model style controller, and an adaptive style controller. Finally, issues regarding closed-loop identification of single-input single-output systems are considered. In particular, it is shown that the direct method of closed-loop identification results in an improved model quality compared to 2-step methods of closed-loop identification.</p>