A comparison of different discriminant analysis techniques in a steel industry welding process

Abstract

The present work compares several statistical discriminant analysis techniques applied to a steel industry welding process. Data from 85 variables collected from 18,605 links, classified as Good (18,381), Defective (195) or Bad (29) from laboratory analysis, are available. Process engineers want to find out which variables explain the main differences between the three defined types, so they can implement effective action to reduce the percentage of Defective and Bad links. The approaches used are SIMCA, Global PCA, PLS-DA and Fisher's Linear Discriminant Analysis (LDA). The dataset comprises two kinds of variables, one for the chemical properties of the links, and the other related to the welding process. All the above approaches basically lead to the same results and match the ones derived from the more traditional Fisher s Linear Discriminant Analysis (LDA) technique. The pros and cons of the approaches used are discussed.