Machine Learning-based Automatic Defect Classification in Real Manufacturing Settings

A substantial amount of classification tasks in real manufacturing settings face significant challenges such as skewed class distributions and multiple performance metrics to meet. This research presents a method to tackle these difficulties and applies it to automatic classification of surface defects on display glass. The proposed method consists of two key stages: (1) using threshold-moving to address class skewness and meet specific quality targets, and (2) using a simple ensemble to consider multiple quality objectives in a classification problem while keeping computational costs low. The results of an industrial case study demonstrate that the proposed approach is a practical replacement for manual inspection of display glass. Future works could optimize the ensemble approach for improving classification performance by investigating different techniques, exploring the use of more sophisticated models, and finding the optimal ensemble size and composition.