Applications of principal component analysis and factor analysis in the identification of multivariable systems

Abstract

The identification of a multivariable stochastic system, usually, involves the estimation of a transfer function matrix, which is a general function of frequency. This estimation involves inversion of a large Hermitian matrix, which sometimes may become unwieldly. In this paper we describe how "principal component analysis" in the frequency domain may be used to replace the input/output variables by some function of smaller dimensions without much "loss of information." The analogy between the "factor analysis" of time series in frequency domain and the minimal realization of state space models is pointed out. The principal component approach described in this paper is applied in the case of a simulated system.