Methodology for real-time feedback variable selection for manufacturing process control: theoretical and simulation results

This paper explores the effectiveness of a proposed methodology for selecting feedback variables for real-time feedback control (RFC). Both analytical and simulation results are presented. In many manufacturing precesses, the important product characteristics cannot be measured in real-time and therefore cannot be directly controlled using RFC. Although the product characteristics may not be fed back, the benefits of RFC, which include reduction of variation through disturbance rejection, may be gained by feedback of process variables closely related to the product characteristics. A general condition under which RFC mill reduce process variation, expressed in terms of sensor noise, process disturbance characteristics and process noise, is derived. Generally process knowledge is used to selected the process variables appropriate for feedback, however in many cases this knowledge is not sufficiently quantitative. A methodology which utilizes statistical analysis of experimental data has been developed for the purpose of identifying the best process variables to regulate in order to minimize variation in the product characteristics. The methodology includes the following steps: design of experiments, candidate model selection, final model selection, check for controllability, and verification, sin efficient, exhaustive search of all possible regression models which satisfy the constraints imposed by the RFC control problem is used to implement the candidate model selection step. The effectiveness of the methodology is evaluated using simulation. Through simulation a wide range of conditions were explored in a relatively short period of time. Situations considered include various degrees of sensitivity to process disturbance, limitations in sensor availability and variation in the importance of unmeasured process variables.