A Predictor Form State-Space Identification Algorithm Using Multivariate Linear Regression

This paper describes a novel MIMO state-space identification algorithm that is based on multivariate linear regression rather than the usual subspace techniques such as orthogonal and oblique projection. We first estimate the Markov parameters of the predictor using multivariate regression, then the state sequence is estimated using singular value decomposition via an equation central to our approach, and finally the A, B, C, K matrices are computed again by multivariate regression. Our algorithm is in predictor form, so it is suitable for both open-and closed-loop cases. Numerical experiments show the accuracy of our algorithm.