Frequency-domain subspace system identification using non-parametric noise models (Q1614389)

The paper is devoted to a study of the stochastic properties of a frequency-domain subspace algorithm using the noise covariance matrix. The true noise covariance matrix is replaced by a sample covariance matrix, obtained from a small number of independent experiments. A basic system model in the frequency-domain subspace is developed. An original system identification algorithm is obtained. It is analyzed with the assumption of an exactly known system input and a proper system. The case of not meeting these assumptions is also discussed. The strong consistency, the strong convergence, the convergence rate, and the asymptotic normality are preserved. In the absence of model errors the uncertainty is slightly increased. Some illustrative real and simulation measurement examples are given.