The large volume of process variables collected from manufacturing applications has jeopardized process control activities. The Partial Least Squares (PLS) regression has been widely used for variable selection due to its ability to handle a large number of correlated and noisy variables. This paper presents a method for selecting the most relevant variables aimed at predicting product variables. For that matter, variable importance indices are developed based on PLS parameters and used to guide the elimination of noisy and irrelevant variables. Variables are then systematically removed from the dataset and the performance of the predictive model evaluated. When applied to five manufacturing datasets, the proposed method retained 31% of the original variables and yielded 6% more accurate predictions than using all original variables. Further, the proposed method outperformed the traditional Stepwise method regarding prediction accuracy.
Variable selection; PLS regression; Variable importance índices
