A search method for selecting informative data in predominantly stationary historical records for multivariable system identification

Multivariable processes are frequently found in continuously operated plants as in chemical, pulp or metallurgical industry. These processes are frequently operated in closed-loop to achieve a secure operation and/or to meet production objectives. Hence, for reasons of plant safety and in order not to violate product quality performing experiments for system identification can be forbidden in the plant. Despite these limitations for collecting relevant data, historical records are logged for years of operation and it would be natural to use them for system identification. These data are available at no cost for analysis. Nonetheless, historical process records contain stationary data which usually represents most of the entire data set. In this contribution, a search method is presented which extracts informative data in multivariable processes to support system identification. The assessment of the search method is evaluated in a case study and benefits of removing non-informative data are demonstrated comparing parameters computed with the entire data set and with a subset formed by informative elements of the original set.