System identification is a multidisciplinary field which focuses on obtaining dynamic models from plant data. Identification is often the most demanding and time-consuming step in the implementation of advanced control technology in the process industries. The focus of CSEL's research on this topic is in the area of control-relevant identification, which takes advantage of the fact that the intended purpose of system identification is control design. As a result, improvements in all facets of the identification problem (experiment design, model structure definition, parameter estimation, and model validation) can be obtained. Control-relevance issues have been examined with regards to both linear prediction-error methods and certain classes of restricted-complexity nonlinear systems.

Our most recent efforts in the area of control-relevant identification involve the use of constrained, minimum crest factor signals for accomplishing “plant-friendly” identification testing. Research on this topic is conducted in collaboration with Professor Hans Mittelmann from the Department of Mathematics and Statistics, with funding provided by a grant from the American Chemical Society- Petroleum Research Fund.

Model-on-Demand Identification and Control of Process Systems

In recent years we have been pursuing the concept of nonlinear identification and control through a formalism named Model-on-Demand (MoD). MoD is a “data-mining” technology inspired by ideas from local modeling and database systems technology. In MoD estimation all observations are stored on a database, and the models are built "on demand'' as the actual need arises. Local models constructed by the Model on Demand predictor use only small portions of data, relevant to the region of interest, to determine a model as needed. The variance/bias tradeoff inherent to all modeling is optimized locally by adapting the number of data and their relative weighting. The MoD approach enhances local modeling and provides the potential for performance rivaling that of global methods (such as nonlinear ARX models, wavelets, fuzzy models, and neural networks) while involving substantially less detailed knowledge of model structure from the user and much more reliable numerical computations.

Since 1996 we have been investigating the MoD estimation framework as an effective, practical means for modeling nonlinear process systems; much of this work has been done in collaboration with researchers from Linköping University in Sweden in the Division of Automatic Control. The research has included such diverse topics as the systematic design of databases for MoD estimation using pseudo-random and minimum crest factor multisine input signals, an experimental study on a brine-water mixing tank, and the development of a comprehensive MoD-based Predictive Control methodology.

Funding for this work has been provided by a National Science Foundation Graduate Research Traineeship and the Honeywell International Foundation.