Evaluation of Experiment Designs for MIMO Identification by Cross-Validation

Guidelines for design of input perturbations for identification of multivariable systems are given. The focus is on control-oriented identification, but for comparison some general-purpose designs are also considered. The information needed for design of input perturbations such as double rectangular pulses, pseudo-random binary sequences, and multi-sinusoidal signals can be obtained from a simple step test. Such a test also gives information about the directionality properties of the system. This information can be used to design control-oriented experiments as shown in the paper. The design techniques are illustrated by realistic simulations of a moderately ill-conditioned 3×3 system. The identified models, and thus the experiment designs, are evaluated by cross-validations. According to this evaluation, the directional input designs are superior to standard input designs for identification of MIMO systems.