Neural networks and evolutionary computation for real-time quality control

Quality control in general and automated quality control in particular are assuming major importance in modem society as technological SNStems are becoming increasingly complex and highly interconnected. Traditional statistical process control techniques are inadequate to address control problems in automated processes because of the high degree of data correlation characterized by such processes. Classical process control methods depend on simplifying assumptions of plant linearity and time-invariance to make the problem analytically tractable. They are therefore limited in effectiveness of the control of complex, nonlinear, multivariable processes.

This dissertation attempts to overcome some of the limitations and shortcomings of traditional quality control methods through the integration of two technologies, neural networks and evolutionary computation. An autonomous control system prototype has been developed to control (maintain quality variables within desired limits) a process by providing high level adaptation to changes in the plant, environment, and control objectives. This technology utilizes memory and learning techniques to overcome limitations of traditional control methods, namely data autocorrelation, requirements of simplifying assumptions, and requirements of a priori information about the process. The robustness and applicability of this integrated technology is demonstrated though results obtained from tests involving simulated processes of varying degrees of complexity.