Description

The application of machine learning is of high significance for quality control tasks in the manufacturing industry due to large volumes of machine-generated data. However, labeling data is costly and labor-intensive. In this study, we evaluate the role of manual labeling and the moderating effect of autoencoder-based pre-training in optical quality control using real-world industrial data. We observe that pre-training substantially elevates the classification accuracy for small amounts of labeled data. With increasing amounts of labeled data available during fine-tuning, however, we find diminishing returns, analogous to recent concerns raised in non-industrial applications.

Share

COinS
 
Jan 17th, 12:00 AM

Rethinking Pre-Training in Industrial Quality Control

The application of machine learning is of high significance for quality control tasks in the manufacturing industry due to large volumes of machine-generated data. However, labeling data is costly and labor-intensive. In this study, we evaluate the role of manual labeling and the moderating effect of autoencoder-based pre-training in optical quality control using real-world industrial data. We observe that pre-training substantially elevates the classification accuracy for small amounts of labeled data. With increasing amounts of labeled data available during fine-tuning, however, we find diminishing returns, analogous to recent concerns raised in non-industrial applications.