Selecting the best variables for classifying production batches into two quality levels

Datasets containing a large number of noisy and correlated process variables are commonly found in chemical and industrial processes, making it hard for engineers to identify the key variables. Typically, the goal has been to identify the most important process variables and create a model using these to predict product variables. Here the focus is not on predicting product variables but on correctly classifying the outcome of each production batch into two quality classes, conforming and non-conforming, for example. The objective is to reduce the number of process variables needed for classification by eliminating noisy and irrelevant ones. We compare several approaches that combine data mining classification techniques with Partial Least Squares (PLS) regression. We propose new indices that measure variable importance based on PLS parameters. Alternative approaches are applied to five manufacturing datasets. With the recommended approach, only 9% of the variables are used for classification, while the classification accuracy slightly increased from average 0.79 to 0.84, compared to using all variables.