Data selection for system identification (DS4SID) from logged process records of continuously operated plants

Abstract

Abstract Use of historical logged data can be considered for system identification if performing dedicated experiments is not possible. Continuously operated plants are examples of processes where experiments for system identification are typically restricted due to a possibly negative impact on production. However, process variables are logged for long periods of time which results in large databases that are a valuable source of information for model estimation. Automatic selection of informative data intervals can support system identification when use of logged process data is addressed. A new method is presented that differs in several aspects from current approaches. Firstly, interval bounding is performed using the gradient of a norm associated to the resulting information matrix which decreases interval misdetection. Secondly, process data do not need to be normalized for change detection. Thirdly, an instrumental variables identification method is used which offers robustness to autocorrelated noise. Lastly, the proposed selection technique can be applied to multivariate processes. The performance of the proposed method is demonstrated in a case study implemented in a lab-scale chemical plant.