In-line visual inspection using unsupervised learning

Purpose and goal:

The purpose is to introduce and improve machine learning- assessment of the quality of mass-produced industrial (steel) products. 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. We want to lift the method from TRL 3 to TRL 7.

Expected results and effects:

The projektet is expected to lead to increased flexibility and scalability in the system as manual handling would be reduced. By both further developing products that industry can use for the purpose and processes where the method is used in existing equipment, the project is expected to generate knowledge and experience that can be useful for many industries.

The work is also expected to lead to the method meeting the industry eligibility requirements on all relevant KPIs, and thereby also improve the state-of-the-art deviation detection in a wider context. Approach and implementation:

The project will be carried out with a test-driven working method that begins with a requirements analysis of the training and detection processes in various industrial contexts at SKF, Gunnebo and Gimic´s customers. The analysis leads to the definition and implementation of the benchmarking KPIs and their assessment of the monitored base solutions.

Then the models are iteratively developed and trained, implemented in production environments and assessed, before the next iteration improves the unsupervised models. The work also expands when new data becomes available.