Professor Emeritus Thomas McAvoy

Ph.D., Princeton University, 1964

Room 2231 A.V. Williams Building
Phone: (301) 405-1939
Fax: (301) 314-0523
Email: mcavoy@umd.edu

Dr. McAvoy's research has focused on process control. Over the past four decades he has studied control system operability, distillation control, and the integration of process design and control. In 1983, he drew together results on a key operability index, the relative gain array and published a monograph dealing with this important subject. Afterwards he and his students developed a new operability index, the relative disturbance gain. This index extends relative gain analysis and it accounts for the effect of process disturbances on control system performance.

Research Interests

Currently, Dr. McAvoy is part of a team of University of Maryland engineers, scientists, and oncologists awarded an NIH grant for a project entitled "Feedback Control and Inferential Modeling for Radiotherapy." The project's goal is to develop a new approach to planning and conducting radiation treatment in cancer patients. While receiving treatment, a patient's respiration may cause a tumor to move, making the delivery of radiation to the right place more difficult, and irradiating healthy tissue in the process. Dr. McAvoy's team is developing a motion-synchronized "treatment couch" that uses feedback control and inferential skin markers to follow tumor movement and direct radiation therapy. Dr. McAvoy has four decades of experience developing feedback control systems and also specializes in inferential sensing, making him a key member of the development process.

Recently, Dr. McAvoy shifted his research focus to artificial intelligence. Specifically, he is looking into neural networks, including combining them with expert systems. Neural computing is a fast emerging technology that holds great promise for the chemical/petroleum industries. In contrast to standard computation, one uses a data set to "train" a neural computer. In a sense the neural computer learns about a problem area. Neural nets have been used to develop non-linear dynamic models for use in model based process control. The control approach has been applied to distillation and reactor examples, and it will be used for waste water control. An active program to improve both the speed of training and accuracy of neural networks is currently underway. This research includes the combination of process models, when they are available, with neural networks. An especially promising area of research involves the combination of neural networks with widely used statistical techniques including partial least squares and principal component analysis. The resulting networks are used to predict both steady state and dynamic relationships among process variables. Initial results have been excellent. Current neural net applications include modeling and control of activated sludge waste water treatment and fire/emission detection.

Selected Publications

"Nonlinear PLS Modeling Using Neural Networks," J. Qin and T. J. McAvoy, Computers and Chemical Engineering,16, 379-391 (1992).
"Long-Term Predictions of Chemical Processes Using Recurrent Neural Networks: A Parallel Training Approach," H. Su, T. J. McAvoy, and P. Werbos, I&EC Research,31, 1338-1352 (1992).

"Intelligent Control," K. Astrom and T. J. McAvoy, J. of Process Control,2, 115-126 (1992).

"Comparison of Four Neural Net Learning Methods for Dynamic System Identification," S. Qin, H. Su, and T. J. McAvoy, IEEE Trans. on Neural Networks,3, 122-130 (1992).

Determining Model Structure for Neural Models By Network Stripping," N. Bhat and T. J. McAvoy, Computers and Chemical Engineering,16, 271-281 (1992).