A cautious approach to minimizing industrial process variability

Reducing product and/or process variability in manufacturing is a fundamental issue in quality improvement programs. Off-line robust design and on-line engineering process control are among the most widely used methods for accomplishing this. In robust design, the objective is to select off-line the fixed controllable input settings in order to make the process output response robust to variations in the uncontrollable noise. In engineering process control, the objective is to actively adjust on-line the controllable inputs in order to make the output response robust to certain types of external disturbances and/or uncontrollable noise variables. The effectiveness of these methods depends strongly on the accuracy of the process models that are used to relate the output response to the controllable inputs, noise variables, and disturbance. Since industrial process models must always be estimated empirically, some level of model uncertainty is unavoidable. If model uncertainty is not properly taken into account, both off-line and on-line process control methods may be completely ineffective and may even increase variability.; This dissertation develops methodologies for incorporating model uncertainty information into off-line and on-line process control strategies for reducing industrial process variability. The intended purpose is to create robust design and engineering process control tools that are themselves robust to model uncertainty. A Bayesian approach is adopted, in which the model uncertainty is characterized via the posterior distribution of the model and model parameters. This allows model uncertainty to be treated simply as an additional source of variation within the overall objective of minimizing process variability. In addition to an estimated process model, the primary additional information that is required is the posterior covariance matrix of the model parameters, which is readily available with most commercial statistical software packages.; The control strategies possess a property that has been referred to as caution in stochastic dynamic programming and adaptive control, where the selected input settings are generally less aggressive than if model uncertainty is neglected. The examples throughout this dissertation demonstrate that the proposed approach can substantially increase the effectiveness of variation reduction methods and improve their robustness to modeling errors.