Application of back-propagation neural networks to system identification and process control
Broussard, Mark Randall
Doctor of Philosophy
Certain properties of the back-propagation neural network have been found to be potentially useful in structuring models for process control applications. The network's relative simplicity and its ability to learn by example are potentially important in the effort to develop automated continuous on-line system identification. The capacity of the network to form nonlinear mappings enhances research designed to advance nonlinear system identification techniques. Since most real processes are nonlinear, this prospect can have wide impact. The unstructured nature of the network model was found to be controllable by techniques developed in the study. Care must be taken to identify and train the network with consistent data that contains sufficient dynamical information. Model-based fine tuning of a controller using a network model that was identified with closed-loop data was successful for the linear and nonlinear systems examined. The utility of the model is a function of the dynamical history of the process. When the information content of the data is sufficient, the network can capture the most important features of system behavior so that fine tuning can be based on optimal parameters such as integral absolute error. This method offers a more complete picture of tuning options than that of other fine tuning techniques such as trial and error, which are not based on a system model. The techniques developed in the tuning effort may be extended to closed-loop model identification for the purpose of controller redesign. In this case, successful identification probably depends on the continuous on-line identification to correct for modeling error.