Graphics visualizing normalising flows in machine learning (AI)

Project: In-line visual inspection using unsupervised learning

The purpose of this project is to introduce and improve machine learning- assessment of the quality of massproduced industrial (steel) products, and in an easier and more efficient way detect quality problems. The goal is to introduce deviation detection based on normalizing flows in real industrial processes, and strengthen the existing End-to-End solution within visual inspection.

Project information

Project manager
Diana Unander
Other project members
Welf Löwe, Morgan Ericsson, Jonas Nordqvist, Felix Viberg, Martin Kroon and Björn Lindenberg
Participating organizations
Linnaeus University, Gimic and Gunnebo Industries, which is part of The Crosby Group and SKF
Financier
Vinnova (within the call for Advanced and innovative digitization autumn 2022)
Timetable
1 January 2023 – 31 December 2025
Subject
Computer and information science (Department of Computer Science and Media Technology), Mathematics (Department of Mathematics), Mechanical Engineering (Department of Mechanical Engineering). All subjects from the Faculty of Technology.

More about the project

Historically, industrial quality control has relied on manual inspection, where personnel assess quality and identify defects by eye—an approach that is both costly and prone to errors. Recently, machine learning-based anomaly detection has been introduced to replace manual inspection. While this enables automation, these methods require collecting training data, which must be manually annotated to teach the system what is correct and incorrect.

Modern industrial quality inspection should instead be based on machine learning with minimal annotation requirements, known as unsupervised learning. These models quantify the probability of an image being normal or anomalous. A promising technique in this field is normalizing flows, which have demonstrated high accuracy in detecting anomalies in standard datasets. However, standard datasets do not necessarily reflect real-world production data, necessitating further investigation.

This project brings together a unique group of experts from manufacturing, automation, quality assurance, and research—combining the necessary knowledge and experience to achieve its goals. Gimic, a startup specializing in industrial quality assurance, provides a cutting-edge platform for implementing these methods. SKF and Gunnebo Industries contribute expertise in manufacturing processes and industrial requirements, along with extensive production data. The developed methods will be tested and deployed within these companies. Researchers from Linnaeus University bring scientific expertise to improve and generalize existing methods.

The project is part of the research initiatives within the Data Intensive Software Technologies and Applications (DISTA), the Smart Industry Group (SIG) and the Linnaeus University Centre for Data Intensive Sciences and Applications (DISA).

Current

Staff