Enhanced real-time quality prediction model based on feature selected nonlinear calibration techniques

This paper evaluates multivariate statistical calibration models for predicting final quality values from process variables. The use of a feature selection technique for multivariate calibration is also provided. Instead of using a full set of process variables, some process variables selected can be used. The objective of such feature selection schemes is to eliminate non-informative variables producing better prediction performance. In this work, genetic algorithm is used as an optimization tool. The performance of the proposed calibration model is demonstrated using real process data. The quality of the final products from the plant is not measured in a real-time basis. Due to the time delay related to measuring final quality values, reliable and timely prediction of the quality characteristics is quite important for safe and efficient operation. By adopting a feature selection scheme along with a filtering step, the prediction performance improved because of the exclusion of non-informative features. The nonlinear calibration models with feature selection and a preprocessing step were shown to produce better performance than those without these two steps. In addition, most of the calibration models considered here benefits from the use of a feature selection step in this case study.