ABSTRACT
I. INTRODUCTION System identification usually implies the modeling of an unknown system from its input-output data. The system model may be either nonparametric or parametric [1]. Usually, in the time domain, a linear system can be described by its nonparametric model, i.e., its impulse response model. An important benefit of impulse response identification is that no a priori knowledge about system order and dead time of the process is necessary but only a rough estimate of the settling time. Recently, it has been found that realization algorithms based on the system impulse response (Markov parameters) can be effectively applied to the problem of state-space model identification. Such algorithms include recent im provements based on the singular value decomposition techniques such as the eigensystem realization algorithm (ERA), the eigensystem realiza tion algorithm using data correlations (ERADC), etc. [2]. In such methods, the Markov parameters are used to determine a balanced state-space model, and the model order is determined by the Hankel singular values. This leads to a clear trade-off between model order and identification quality in terms of a singular value plot. It is obvious that such realization algorithms require accurate estimates of the Markov parameters. Therefore how to identify the system impulse response
effectively is viewed as an important task in system identification theories and techniques.
