Correlation Test of Residual Errors in Frequency Domain System Identification

Residual errors (deviations between measurements and system models) can be caused by several reasons: observation/process noise, nonlinear products, system transients, unmodelled dynamics, etc. The first two cannot be explained by linear models, the latter two can. Identification procedures can be stopped when the latter are not present in a reasonable size model. Therefore, we need to distinguish among these error sources. This paper analyses components of the residual errors, and suggests a simple automatic method to effectively suppress system transients and unmodelled dynamics during variance analysis. This allows robust noise analysis, and then reliable model validation. The method is based on different properties of the error sequences in the frequency domain, without the need of separate observation noise analysis or of “full-blown” nonlinear analysis. Copyright © 2006 IFAC