Ensuring product quality through machine learning for Altia

Ensuring quality of products by machine learning in Altia’s Rajamäki factory

Altia's historic bottling factory in Rajamäki has been in production since 1888. A total of 800 different wines and spirits, up to 70 million liters a year, is produced in the factory. There are eight production lines at use at the factory, which produce different products at different times.

Altia has a very strict process for product quality. Automatic inspection can detect missing labels, faulty closures or deficient fill level. However, manual inspection is needed in 8 production lines to detect other issues. Overseeing the production lines in order to weed out faulty bottles is a manual and boring process. Manual inspection requires the full attention of one person per production line, where the production speed can be up to 18 000 bottles per hour.

Altia wanted to improve quality assurance by automating the fault detection further, thus giving the quality assurance team the possibility to focus on more challenging tasks. With raised accuracy, waste would also be reduced, as the production line can be stopped and fixed if errors occur.

In the proof of concept, Bilot and Altia realized that this problem can be solved with image recognition and machine learning.

What we did

We helped the customer:

• Scope the problem for proof of concept
• Create the data set to test drive the algorithms
• Train their people in machine learning problem solving
• Build the prototype into the cloud and on-premise for comparison purposes
• Create the business decision materials and justifications for the actual project

Technologies

Azure Cognitive Services – Cloud solution

Keras & TensorFlow – On-premise solution

Custom firmware for Canon 60D camera

Image recognition to the rescue

In order to automate the inspection process, we set up a camera to the bottling machine, connected it to AI logic, and then connected the AI logic to the conveyor belt logic to reject faulty bottles. A cloud-based and on-premise solution were both used in order to test accuracy, and over 1200 images were used as a training set for the AI. The end results: 100 % accuracy in recognition in both the cloud and on-premise solutions. Altia now has the capability to divert quality inspection resources to more complex matters.