Model Structure Determination and Identifiability Problems in System Identification

摘要

The canonical structure of linear systems is examined and specific canonical forms are constructed. It is shown that although a general stochastic model is not identifiable, its associated steady-state kalman filter is identifiable if a canonical form is used. A non-iterative method is developed for estimating the parameters (including model order and noise covariance) of a steady-state Kalman filter. Finally, the concept of local identifiability is discussed and sufficient conditions are derived for local identifiability of parameters in terms of the Fisher information matrix. (Author)