Realization-based system identification with applications

The identification of dynamic system behavior from experimentally measured or computationally simulated data is fundamental to the fields of control system design, modal analysis, and defect detection. In this dissertation, methods for system identification are developed based on classical linear system realization theory. The common methods of state-space realization from a measured, discrete-time impulse response are generalized to the following additional types of experiments: measured step responses, arbitrary sets of input-output data, and estimated cross-covariance functions of input-output data. The methods are particularly well suited to systems with large input and/or output dimension, for which classical system identification methods based on maximum likelihood estimation may fail due to their reliance on non-convex optimizations. The realization-based methods by themselves require a finite number of linear algebraic operations. Because these methods implicitly optimize cost functions that are linear in state-space parameters, they may be augmented with convex constraints to form convex optimization problems. Several common behavioral constraints are translated into eigenvalue constraints stated as linear matrix inequalities, and the realization- based methods are converted into semidefinite programming problems. Some additional constraints on transient and steady-state behavior are derived and incorporated into a quadratic program, which is solved following the semidefinite program. The newly developed realization- based methods are applied to two experiments: the aeroelastic response of a fighter aircraft and the transient thermal behavior of a light-emitting diode. The algorithms for each experiment are implemented in two freely available software package