Active inspection with knowledge distillation for cost-effective fault prediction in manufacturing process

Manufacturing processes involve various inspections aimed at identifying faulty products before they reach the final stages. These inspections are typically categorized into basic, conducted for all products, and advanced, conducted only for selected sampled products due to cost constraints. Recent advancements have leveraged inspection data to train machine learning models that predict potential faults in manufactured products. However, models using only basic inspection results, referred to as the basic model, often underperform compared to those that also use advanced inspection results, referred to as the advanced model, due to limited information. In this study, we propose a novel approach to train a basic model using knowledge distillation from an advanced model, achieving high prediction accuracy with reduced inspection costs. Additionally, we incorporate this distilled basic model into an active inspection framework during the inference phase to further improve the cost-effectiveness of inspections. Within this framework, the distilled basic model and the advanced model are selectively used, optimizing the usage of inspection costs. The effectiveness of our approach is demonstrated through a case study involving real-world data from a semiconductor manufacturer.