Searching for informative intervals in predominantly stationary data records to support system identification

Abstract

Performing experiments for system identification is prohibitive in some processes due to restrictions in production or due to requirements for a safe operation. In such situations, system identification with historical data records can become a suitable alternative. Nonetheless, these data records are mainly stationary because few transient changes occur. Thus, the available data records can be considered “little” informative and the model quality may be affected using these data sets. In this contribution, a novel search method to find informative data in large mostly stationary data sets is presented. In contrast to available search methods, independent moving windows are used to screen the data and detect transient changes. The performance of the proposed method is demonstrated on case studies. Results demonstrate that parameter biases of a selected model decrease and small changes in variances of the parameters are observed using the selected subset for parameter estimation.