Industrial PLS model variable selection using moving window variable importance in projection

• An assessment of existing variable selection methods based on Partial Least Squares (PLS) models.
• Formulation of a new selection method using moving window variable importance in projection (MW-VIP) to target selection of local correlations.
• Introduction of a set of evaluation criteria to assess the prediction performance, robustness, and monitoring sensitivity of the PLS models along with results using industrial data-set.
• Application of proposed approach and evaluation methods on an industrial dataset.

Soft sensors (or inferential sensors) have been demonstrated to be an effective solution for monitoring quality performance and control applications in the chemical industry. One of the key issues during the development of soft sensor models is the selection of relevant variables from a large array of measurements. A subset of variables that are selected based on first principles and statistical correlations eases the model development process. The resulting model will perform better and will be easier to maintain during the deployment stage. In the current literature, data-driven variable selection methods have been investigated within the context of spectroscopic data and bioinformatics. In these studies, the variable selection methods assume that the inherent correlation in the entire data set remains fixed. This is not the case in common industrial processes. In this paper, existing variable selection methods based on partial least squares (PLS) will first be evaluated. Second, we will present a new approach called moving window variable importance in projection (MW-VIP) to target the selection of correlations present in segments or small clusters. Finally, a set of new evaluation criteria will be presented along with industrial data set modeling results to demonstrate the effectiveness of our proposed approach.