A Data Selection Method for Large Databases Based on Recursive Instrumental Variables for System Identification of MISO Models

Experiments in industrial plants to collect data for system identification are constrained due to production and safety requirements. In such situations, logged historical data can be instead used for system identification instead. However, these recorded data are predominantly stationary in continuously operated plants since processes are seldom excited during normal operation. Performing system identification with such data will yield numerical problems. Alternately, the “most” informative sequences can be extracted and used for system identification. Current data selection methods have several drawbacks. They are constrained to Single-Input Single-Output (SISO) modeling problems. The methods are not robust against correlated noise which is a disadvantage when using real data sets. Moreover, setting design parameters requires some information about the process that is not always available. In this contribution, an alternative data selection method for system identification is presented and evaluated in a case study. In contrast to current approaches, the proposed method does not require data normalization to detect transient changes. It can be used in Multi-Input Single-Output (MISO) systems operating in open or closed loop. An instrumental variables (IV) method is used in the algorithm which provides robustness against non-white noise. Results from a simulation case study of a multivariable system show that models with similar accuracy are obtained when using the intervals retrieved by the data selection method as when using the entire data set.