Several correlated and noisy variable are collected from industrial processes. This paper proposes a method for selecting the most relevant process variables aimed at classifying production batches into classes based on multiple criteria (e.g., sensibility and specificity). Production batches are inserted into two classes. The method first applies the PLS regression (Partial Least Squares) on process data and derives a variable importance index. A classification/elimination procedure is then carried out, and a weighted Euclidian distance is generated to identify the recommended variable subset. When applied to the testing set of real industrial data, the proposed method retained average 12% of original variables. The recommended subsets yielded 9% higher sensibility, from 0.78 to 0.85, and 20% higher specificity, from 0.64 to 0.77. Simulation experiments are also performed.
Variable selection; Multiple criteria; PLS regression
