(698b) Design of Constrained Experiments for Identification of Multivariable Fir Models

A true experimental design ? where test outputs are generated from an optimization performed based on model structure and constraints ? is not typically done in practice. Independent random binary signals such as PRBS or GBN are often used in practice, but are sub-optimal when output constraints are present; further, they are known to be inefficient for the identification of ill-conditioned systems. Control relevant experiment design with a criterion of integral controllability has received significant academic attention over the last 15 years, but it has been based on a static model, with ad-hoc extensions to the dynamic case. Even with traditional design measures, there has been the challenge of formulating design problems for the dynamic, multivariable case that is amenable to online optimization. In this work, we propose an open-loop design in the time domain for multivariable FIR models that explicitly considers input and output variance constraints. A key aspect of the problem formulation is an input parameterization that results in a tractable optimization problem. Optimal results are expressed in terms of input covariance, which can be readily implemented. The method is illustrated on examples of different design criteria.