ON THE RELIABILITY OF SYSTEM IDENTIFICATION TECHNIQUES

Abstract

The reliability of the system parameters identified by analysing input-output data pairs is addressed. The least square time domain identification procedure is used to identify the system matrices. The issue of incomplete representation through a limited availability of response measurements and the resulting uncertainty in the estimated system parameters is studied. To access the robustness of parametric identification different durations of excitation signals and different quality of response signals are used. Statistical analysis of identified parameters is performed to estimate their variability. An important finding is that the system parameters aregrossly underestimated even if the variance is not large. It is shown that even with these obviously wrong estimates of system parameters the computed natural frequencies remain reasonably stable irrespective of signal duration, signal quality, and availability of responses. Therefore, natural frequencies are more suitable for use in health monitoring and model updating studies