Online Class Imbalance Learning for Quality Estimation in Manufacturing

Abstract

Online machine learning has become increasingly important recently as more and more machines are being connected and data is being sent to the decision making node in real time. Traditional batch based machine learning is no longer suitable for such streaming data scenario. Here, an online classification algorithm to classify good and defective product, under imbalance streaming environment, is proposed. The proposed method exploits the assumption that different classes should be far away from each other. Even when the raw data might appear to be close, the algorithm learns and projects them into some specific manifold where different classes are far from each other. The algorithm classifies good and defective product in an imbalanced environment where good product outweighs defective product. The algorithm uses only single pass of the data, where the data is used once and then discarded. The approach is then being validated using industry data and the result indicates better performance in term of G-Mean and F1-score.