Note: This content is accessible to all versions of every browser. However, this browser does not seem to support current Web standards, preventing the display of our site's design details.

  

Presentation Title

Back
Abstract:

System identification refers to the art / science of building a mathematical model of a dynamical system based on observed data from the system. As the interest in model-based control continues to mount, many experts are pointing to system identification as the key bottleneck. Currently, models for practically all model-based control applications are obtained through plant tests.

System identification basically consists of two steps: data generation and model fitting. The data generation step is crucial in many ways. Not only does it fundamentally determine the nature and accuracy of the system characteristics that are identified, it is the only step requiring interaction between the user and the process and is therefore the most costly and time-consuming step. There lies a great incentive to optimize the experimental conditions, e.g., the choice of the input signals, so that most informative data can be generated in the shortest possible time span with minimal intrusion to the on-going operation. On the other hand, most research efforts in system identification have been directed to the other problem, i.e., selecting the best model from a model set based on given data, and little science exists on data generation currently. This is especially true in the multivariable system context, where the current practice amounts to varying input channels independently without any special regard given to existing plant interactions. What seems to make the optimal experiment design problem difficult to handle from a theoretical standpoint is the fact that the optimal design depends on many unknown factors such as the plant. This means that there has to be some form of feedback from the plant to the test protocol. This may be done, for instance, by physically closing a loop between the plant output and plant input resulting in the so-called ``closed-loop identification," or by continually redesigning the input signals on the basis of the current plant model / uncertainty information resulting in ``iterative / adaptive identification.

In this seminar, we begin by examining why and how the current practice of system identification, even when carried out in the most ideal manner, can fail, leading to multivariable models that are unsuitable for control applications. We point to poor experiment design as the main cause of this. Then, we introduce an optimization-based procedure to generate informative data for highly interactive plants. In this approach, we propose to formulate the input design problem as a mathematical program: maximize the model quality subject to constraints on outputs, number of data, and inputs. This requires a mathematical expression (objective function or "input design criterion") that measures model quality. Traditionally, model quality is measured by some simple function of the error covariance matrix of the parameter estimates, which does not truly reflect the end goal. We develop an objective function, which is directly related to the closed-loop performance of the model. Because the closed-loop performance of a particular model depends, among other things, on the unknown system, we feedback the collected information to the input design scheme, resulting in iterative / adaptive identification. We use an example of a non-ideal high-purity distillation column as a vehicle to demonstrate the potential pitfalls of the current practice and to assess the benefits of the proposed procedure.

http://atom.ecn.purdue.edu/~jhl/>http://atom.ecn.purdue.edu/~jhl/
Type of Seminar:
Public Seminar
Speaker:
Dr. Jay H. Lee
School of Chemical Engineering, Purdue University, U.S.A
Date/Time:
Jul 27, 1999   10:30
Location:

ETH-Zentrum, ETL K 25, Physikstrasse 3, 8006 Zuerich
Contact Person:

Couson
No downloadable files available.
Biographical Sketch:
Jay H. Lee was born in Seoul, Korea, in 1965. He obtained his B.S. degree in Chemical Engineering from the University of Washington, Seattle, in 1986, and his Ph.D. degree in Chemical Engineering from California Institute of Technology, Pasadena, in 1991. >From 1991 to 1998, he was with the Department of Chemical Engineering at Auburn University, AL, as an Assistant Professor and an Associate Professor. Since 1998, he has been with School of Chemical Engineering at Purdue University, West Lafayette, where he currently holds the rank of Associate Professor. He has held visiting appointments at E. I. Du Pont de Nemours, Wilmington, in 1994 and at Seoul National University, Seoul, Korea, in 1997. He was a recipient of the National Science Foundation's Young Investigator Award in 1993. His research interests are in the areas of system identification, model predictive control and nonlinear estimation.