Date of Award

Spring 2002

Document Type


Degree Name

Doctor of Philosophy in Chemical Engineering - (Ph.D.)


Chemical Engineering, Chemistry and Environmental Science

First Advisor

Robert Benedict Barat

Second Advisor

Basil Baltzis

Third Advisor

Denis L. Blackmore

Fourth Advisor

Dana E. Knox

Fifth Advisor

Norman W. Loney

Sixth Advisor

John G. Stevens


Active feedback and feedforward-feedback control systems based on static-trained feedforward multi-layer-perceptron (FMLP) neural networks were designed and demonstrated, by experiment and simulation, for selected species from a laboratory two stage combustor. These virtual controllers functioned through a Visual Basic platform. A proportional neural network controller (PNNC) was developed for a monotonic control problem - the variation of outlet oxygen level with overall equivalence ratio (Φ0). The FMLP neural network maps the control variable to the manipulated variable. This information is in turn transferred to a proportional controller, through the variable control bias value. The proposed feedback control methodology is robust and effective to improve control performance of the conventional control system without drastic changes in the control structure. A detailed case study in which two clusters of FMLP neural networks were applied to a non-monotonic control problem - the variation of outlet nitric oxide level with first-stage equivalence ratio (Φ0) - was demonstrated. The two clusters were used in the feedforward-feedback control scheme. The key novelty is the functionalities of these two network clusters. The first cluster is a neural network-based model-predictive controller (NMPC). It identifies the process disturbance and adjusts the manipulated variables. The second cluster is a neural network-based Smith time-delay compensator (NSTC) and is used to reduce the impact of the long sampling/analysis lags in the process. Unlike other neural network controllers reported in the control field, NMPC and NSTC are efficiently simple in terms of the network structure and training algorithm. With the pre-filtered steady-state training data, the neural networks converged rapidly. The network transient response was originally designed and enabled here using additional tools 'and mathematical functions in the Visual Basic program. The controller based on NMPC/NSTC showed a superior performance over the conventional proportional-integral derivative (PID) controller. The control systems developed in this study are not limited to the combustion process. With sufficient steady-state training data, the proposed control systems can be applied to control applications in other engineering fields.