Neural network approach to dynamic process control

Abstract

There is a growing interest in the application of Neural Networks (NNs) to the intelligent control of a nonlinear dynamic process which cannot be described by mathematical models. One of the most relevant features of NNs is self-organizing which provide a good representation of dynamic process through learning by examples. Additional relevant features of NNs to process control are adaptability to a change in environment, robustness with respect to noise, and generalization over the set of previously learned instances. In this thesis, we present a neural control architecture, named Neural Network with Partial Inversion (NN-PI) for the modeling and control of the process. In this architecture, a forward model of the process is obtained using a backpropagation learning algorithm and then the partial inversion method of the trained network produces the control signals, given the process output and current states of the process. This architecture would resolve the problems of the conventional neural control approaches, such as (1) the one-to-many mapping problems which may prove unable to find an inverse in the case of multiple-inputs/single-output of the process and (2) the training and analysis complexities resulted from placing the nonlinear NN in the feedforward or feedback path of the indirect inverse control approach. NN-PI architecture is applied to two complex industrial processes, melting process in T.V. glass-bulbs production and traffic signal control. NN-PI architecture would be well-suited to the traffic signal control in a single intersection. However, a single NN used for traffic signal control in urban traffic networks must be large and thus it requires a long time to train such a large network. In order to overcome this problem, we propose the two-level hierarchical control architecture through the decomposition of the process control problem into manageable subproblems. Through a series of experiments, we found that this architecture can reduce the...