Process monitoring for quality–a feature selection method for highly unbalanced binary data.

Defect detection is among the most critical modern challenges in manufacturing science. The state-of-the-art in this field involves generating only a few defects per million opportunities. Process Monitoring for Quality is a philosophy driven by big data and aimed at rare event detection applied to quality control. This is accomplished using binary classification and empirical knowledge discovery through feature interpretation, which is facilitated by feature selection methods. These analytical tools help identify the driving features of a system, which are then used in a manufacturing context to plan and design randomized experiments to determine optimal process parameters. This work presents a new filter-type feature selection method based on the separation between classes. As shown in previous studies, predictive ability is strongly correlated to the distribution of margins. Because manufacturing-derived data sets for binary quality classification tend to be highly or ultra-unbalanced, the proposed method is designed to analyze these data structures effectively. The method properties and ability to select high-quality features are illustrated through three case studies. First, virtual features are used to demonstrate the method procedures. Then the method is used to analyze a manufacturing-derived data set, from which the most relevant feature is identified and used for process redesign. Finally, five of the most widely used methods are compared with the proposed method using an analysis of publicly available data sets. The empirical results demonstrate a significantly improved prediction ability of the features selected by the proposed method. [ABSTRACT FROM AUTHOR]