No abstract

We present results concerning the parameter estimates obtained by prediction error methods in the case of input signals that are insu ciently rich. Such input signals are typical of industrial measurements where occasional stepwise reference changes occur. As is intuitively obvious, the data located around the input signal discontinuities carry most of the useful information. Using singular value decomposition techniques, we show that in noise undermodeling situations, the remaining data may introduce large bias on the model parameters with a possible increase of their total mean square error. A data selection criterion is then proposed to discard such poorly informative data so as to increase the accuracy of the transfer function estimate.

No abstract

No abstract

No abstract