Study of Conditional ML Estimators in Time and Frequency domain System Identification..pdfVIP

Study of Conditional ML Estimators in Time and Frequency domain System Identification. J. Schoukens, R. Pintelon and Y. Rolain Vrije Universiteit Brussel, Department ELEC, Pleinlaan 2, 1050 Brussels, Belgium email: jschouk@vnet3.vub.ac.be; tel: (02) 629 29 44 fax: (02) 629 28 50 Keywords: linear identification, stochastics, estimation tages [7]. The covariance matrix after a DFT is dominated by its diagonal so that the full MLE can be approximated by a Abstract weighted least squares estimator using the diagonal terms as a nonparametric weighting. These weighting coefficients (being This paper studies the impact of replacing the full covariance the variances of the measured spectra) can be easily obtained matrix by its main diagonal in maximum likelihood from a small number of repeated measurements if the excita- estimation (MLE) of linear dynamic systems in the time or tion is periodic [8]. This need for periodic excitations is the frequency domain. First the source of the nondiagonal entries major disadvantage of this approach to frequency domain is studied, next the impact of neglecting these terms on the identification. A detailed analysis shows that the neglected efficiency of the estimators is analyzed. Finally the nondiagonal entries in the covariance matrix can be explained equivalence between the time and