A Framework For Observational Data-Based Response Surface Methodology

Response Surface Methodology (RSM) is an integrated tool for optimization purposes based on an experiment. The designed experiment ensures the researcher fully controls all factors that potentially influence the response and simultaneously fulfills the orthogonal assumption among factors. On the other side, conducting DoE for a continuous production process raises difficulties since it should be interrupted during experiment runs. Meanwhile, the rapid development of production data acquisition systems provides stored records or observational data with potentially useful information for supporting process optimization. This paper proposes an alternative framework for adopting observational data for RSM analysis to reduce the need for real experimentation and ongoing production disruption. Referring to three stages of classic RSM and adopting the instance selection concept in the data mining context, the proposed framework aimed to achieve an observational data condition similar to an orthogonal D-optimal DoE based on criteria of Variance Inflation Factor (VIF) and determinant of matrix containing factor levels. It starts by applying a genetic algorithm for iteratively selecting an orthogonal subset of observational data and generating new actual experiment points to satisfy an orthogonality criterion. Then, a linear RSM model is fitted and continued by adding new experiment points. Then a standard numerical optimization method is applied to search among factor levels that optimize the response. A simulated data-based case study was taken in this paper, aiming to maximize the response of a production process with some pre-determined factors. The proposed framework has been implemented successfully, orthogonality of the data subset is achieved, and an optimal solution is found. Both criteria show acceptable results and raise some improvement opportunities.