Casting plate defect detection using motif discovery with minimal model training and small data sets

Manufacturers are increasingly applying machine and deep learning to automate production quality monitoring to save time and costs. The most widely used approach is Convolutional Neural Network (CNN) trained to detect quality issues in production output images. While the approach achieves high accuracy, many companies face challenges implementing it. Many manufacturers lack both the big data sets required for machine and deep learning model training and the data scientists having the domain knowledge to build and run complex models. Today manufacturers have implemented lean manufacturing and six sigma quality controls which result in small defect samples that are not sufficient for modeling. Some manufacturers also change the production outputs frequently which does not permit enough time for data collection for model building. In this paper, we propose two motif discovery based approaches that work within the constraints of modern manufacturing. The first approach is programmatic motif discovery learning patterns from small data samples. The second approach is a self-service visual motif discovery that is simple and intuitive for engineers not versed in data science. We compare the proposed approaches with a CNN and conclude that our proposed methods achieve higher accuracy, have significantly lower computational costs, and empower engineers to do it themselves.