Removal of Insufficiently Informative Data to Support System Identification in MISO Processes

Model accuracy can be degraded by using insufficiently informative data for parameter estimation. This effect can be observed in case of model structures where the input transfer function and the noise transfer function share the same denominator. Equation error models such as ARX, ARMAX or ARARX are examples of model structures exhibiting the latter characteristic. Removal of insufficiently informative data for parameter estimation may increase model accuracy as reported for SISO systems. In this paper, a search method to find informative intervals in MISO systems is introduced. The removal criteria are based on comparison between means of a regression matrix associated to a model class and a threshold related to the conditioning of the matrix. These matrix means are used as quality factor to discard insufficiently informative data. The performance of the search method is demonstrated in a case study.