Nonparametric Preprocessing in System Identification: a Powerful Tool

In this paper, we study the properties of existing nonparametric methods for estimating the plant and noise transfer functions of a linear dynamic system. The analysis is based on the recent insight that leakage errors in the frequency domain have a smooth nature that is completely similar to the initial transients in the time domain. This not only allows us to understand better the existing classic methods, but also opens the road to new better performing algorithms. The paper includes the output error setup, the errors-in-variables setup, and measurements under feedback conditions. Eventually, some of the methods are illustrated in the analysis of a vibrating metal beam.