Doctoral Dissertations

Guiran Chang

Date of Award

12-1991

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Electrical Engineering

Major Professor

J. D. Birdwell

Committee Members

J. S. Lawler, M. M. Trivedi, C. F. Moore, T. W. Wang

Abstract

In our modern world with heavy competition, it is important that the productivity, product quality, and other performance measures of a production process be improved continuously without interruption or degradation of its normal operation. New tools are desirable which can monitor a process' dynamics and economic performance, learn new knowledge about how to do the job better based on past experience, and adjust the behavior of the process to improve its performance, all in real-time, so that its long-term economic performance can be optimized. Artificial intelligence and expert systems techniques, coupled with system cultivation and optimization methods, provide a new approach to the solution of this problem. The research described in this report is an attempt to implement such a tool, a knowledge-based supervisor, targeted at an example problem of tank level scheduling. For this example, the goal of the supervisor is to select a tank level set point schedule so that the average cost of the purchased supply is minimized. System cultivation methods, proposed by B. Moore, are used to extract information or knowledge about the process, and the long-term optimization goal is achieved by using short-term optimization repeatedly. After a prototype software system is implemented, it is applied to other problems, including a more practical inventory scheduling problem, set point optimization of a continuous stirred tank reactor (CSTR) and a continuous bioreactor, which is a special case of CSTR, under parameter uncertainty. The research is based on simulation models. The simulation programs are implement ed on a Sun microcomputer system, together with other numerical routines. A simple rule-based expert system and the system cultivation procedures are implemented on a TI Explorer II Lisp machine. For the tank level scheduling problems, linear programming is used in the optimization. The results demonstrate that the supervisor, guided by heuristic rules, can find a better schedule for supply purchase so that the overall cost is minimized under given constraints. For CSTR and bioreactor control, dynamic optimization is achieved by continuously searching out the optimal set point with small excursions introduced to the nominal set point. The result is promising. The current work provides a background for further research in knowledge extraction and performance improvement of the process. A complete knowledge base could be constructed from historical measurement data, and can be used to determine the control or scheduling actions. Other optimization methods should be investigated, and the effect of each short-term optimization on the long-term goals can be assessed to guarantee the optimization result. For construction of a more realistic software system, an implementation utilizing the C++ language and X Window System on a Sun workstation was investigated.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS