Abstract
This paper presents a procedure to identify linear multivariable systems in the state space representations from the input-output data contaminated by noise. The input-output relation is identified by using the generalized least squares identification which is a modified Clarke's algorithm, and the (I/O) relation is minimally realized in the practical sence. This minimal realization, named the ε-minimal realization, is found effective for the structural determination in the identification, where the dimension of an ε-minimal realization is depending on the choice of ε. A method to determine ε by the statistical test is also given. The proposed method is effective as also shown by some numerical examples.
