Feature Selection to address High-Dimensionality in Industry 4.0 Multi-emitter Laser Modules Assembly Lines

Abstract

Industry 4.0 has increased data depth and breadth in high-tech manufacturing, but high-dimensionality and sparsity persist. High-dimensional space's sparsity makes classical learning and knowledge extraction algorithms ineffective and error-prone. Dimension reduction methods like feature selection seem to address this problem. This study addresses these challenges by conducting a comparative analysis on a real laser assembly industrial case of high dimensions. We explore five standalone methods—NCFS, RReliefF, MRMR, RFE, and Lasso—applied to datasets from two laser modules (d-serie and s-serie). Additionally, two hybrid methods—RReliefF-RFE and MRMR-RFE—are evaluated, broadening the scope of feature selection strategies. Time efficiency prioritizes RReliefF, NCFS and Lasso, while RReliefF-RFE, NCFS and Lasso excel in interpretability, achieving significant predictor reduction without compromising accuracy. The study thus provides insights into the selection of FS methods in a challenging industrial laser assembly setting.