Quality Analysis Using Motif Discovery in Small Data Sets

Abstract

Manufacturers are increasingly interested in applying machine and deep learning to automate quality control monitoring to save time and costs. A typical approach is the use of Convolutional Neural Network (CNN) to assess and classify quality issues in images of the production outputs. However, many companies report project challenges and prohibitive computational costs. The challenges arise because of lack of big data required for machine and deep learning model training and lack of data scientists with domain knowledge to build and run complex models. Today manufacturers operate at six sigma levels of quality, and thus have extremely few defect samples. In many cases production changes frequently, and thus the data collection and model training have to be re-done for the new outputs. In this paper we propose two approaches - motif discovery from small sample data and visual motif discovery being simple and intuitive for any engineer to use. We compare them for ease of use and computational efficiency with a Convolutional Neural Network (CNN) and conclude that our proposed approach significantly improves the existing performance, provides fast and reliable training with small samples, and empowers engineers to do it themselves (DIY).