A neural network approach to single and multidimensional model based adaptive control

Stricter environmental regulations and a greater need for waste minimization have increased the importance of process control to chemical plant operations. However, the standard PID controller common to most plants often does not offer sufficient performance as the controller is not designed to account for the nonlinear or nonstationary behavior of most processes. The desire to improve controller performance has led to a growing interest in adaptive control. Adaptive controllers typically operate in a model based fashion, where process data is used to update the parameters of a simple linear process model so that the model remains descriptive of the process dynamic behavior. The controller parameters are then related to the model parameters in a manner which allows the controller to maintain robust performance.^ This research investigates a pattern recognition approach to adaptive control. Pattern recognition techniques analyze the input and output process response following dynamic events such as set point changes and significant disturbances. They then translate these responses into descriptive pattern features. Past pattern recognition techniques have used a rule based or expert system approach, where a series of rules translate features such as overshoot and damping into controller parameter updates. This work uses Artificial Neural Networks (ANN) for the pattern analysis task due to their inherent ability to analyze the entire pattern as opposed to particular features, making them less susceptible to measurement noise and other irregularities. The ANN\u27s are used to translate dynamic input/output responses into a measure of the multiplicative mismatch between a present model parameter and the value of the model parameter which is descriptive of the process. By focusing on parameter mismatches as opposed to actual parameter values, the adaptive strategies maintain process independence.^ Focusing on model parameters as opposed to controller parameters allows the methods to be applied to a number of model based control algorithms, including an IMC based PID structure and Dynamic Matrix Control. This work implements the adaptive techniques on a variety of challenging single and multivariable processes in an effort to determine the strengths, weaknesses and limits of the pattern recognition approach.