Comparison of Some Time Domain System Identification Techniques using Approximate Data Correlations

Most time-domain methods used for modal analysis perform a curve fit to impulse response data and use the least squares method as an integral part of their formulations. It is well known that when the data are corrupted, least squares leads to biased parameter estimates, with the damping values being especially sensitive. A number of non-iterative techniques that attempt to eliminate the bias – instrumental variables with delayed observations, double least squares, correlation fit and total least squares - are compared with least squares in terms of the approximate data autocorrelations used in the curve fit. The theoretical comparison is complemented by statistical comparisons upon simple simulated systems. As well as the ability of the methods to reduce the bias on damping estimates, ease of implementation and computational requirements are also investigated.