Bayesian system identification of dynamical systems using highly informative training data

This paper is concerned with the Bayesian system identification of structural dynamical systems using experimentally obtained training data. It is motivated by situations where, from a large quantity of training data, one must select a subset to infer probabilistic models. To that end, using concepts from information theory, expressions are derived which allow one to approximate the effect that a set of training data will have on parameter uncertainty as well as the plausibility of candidate model structures. The usefulness of this concept is then demonstrated through the system identification of several dynamical systems using both physics-based and emulator models. The result is a rigorous scientific framework which can be used to select 'highly informative' subsets from large quantities of training data.