Artificial intelligence at Nestlé: Innovative process control with deep learning

May 10, 2022

GO BEYOND: DISCOVERY #20.

Artificial intelligence with deep learning software from SICK: At the Nestlé production site in Osthofen, a 2D snapshot camera in conjunction with the integrated SICK software package “dStudio” detects transparent objects in any position. And the system is continuously learning as well.

The health science division of the world's largest food company Nestlé offers specialized products for people with specific nutritional requirements. For example sip feeding and supplemental food, which are manufactured at the Osthofen plant in the Rhinehessen region.

The powder produced here under state-of-the-art conditions automatically ends up in special cans at the end of the manufacturing process. Before filling each container, a measuring scoop for easy and precise dosing is added. As part of the quality control process, each can is checked to ensure that a scoop has actually been applied.

“While a human eye can easily recognize whether a scoop has been enclosed,” says Marcus Kauf, an automation technician at Nestlé, “at a filling speed of over 80 cans per minute it is no longer possible to do so without errors.”

 

 “Spoon check”: a real challenge

That’s why the check was performed instead by a specially installed camera that counted the colored pixels of the plastic spoon. Nestlé has recently started using a colorless scoop, however, to improve the recycling rate. There are therefore now gray-tinged transparent scoops sitting on top of aluminum foil closing flaps with a similar hue ... or perhaps not always? Being difficult to identify on the corrugated, embossed, and reflective metal – the traditional solution with sensor-based image processing had hit its limits. 

The camera quickly and reliably checks every pack to see whether the dosing scoop is actually enclosed or not – with an error rate approaching zero.
The camera quickly and reliably checks every pack to see whether the dosing scoop is actually enclosed or not – with an error rate approaching zero.

dStudio: an innovative type of image processing

This is where the technology and know-how of SICK came into play. The key to the solution was “Artificial intelligence (AI)”. AI very quickly, reliably and consistently detects new patterns by collecting a large quantity of data that is interpreted directly by algorithms. From a theoretical view, algorithms specify a clearly defined procedure for solving specific tasks.

In practice, a 2D snapshot camera (picoCam) from SICK learns to “think” with the help of the integrated deep learning software “dStudio” from the SICK AppSpace. Using this web service, neural networks can be “trained”, in this case with images of the enclosed scoop in a wide variety of positions. Just like humans can solve problems and make decisions in an ideal situation, so too does deep leaning – only many times faster.

Camera technology keeps learning about image differences every day

The responsible expert at SICK is Klaus Keitel. The national account manager for strategic customers explains: “The taught-in decision-making algorithm is transferred to the camera system. This makes it able to independently detect significant image differences.” The camera can also be adapted to each new product without difficulty by re-training the neural network for the new circumstances. Furthermore, the intuitive operating entity interface does not require the user to have any specialist AI know how or image-processing knowledge.

Inspection on the conveyor belt

The camera quickly and reliably checks every pack to see whether the dosing scoop is actually enclosed or not – with an error rate approaching zero. But what happens if the camera identifies the “No scoop” image? Then the system stops automatically. Once the missing scoop has been added, the camera software detects this and allows the system to continue running without the inconvenience of a manual restart.

 

Engineering Tools
Artificial intelligence for SICK sensors
Deep Learning
Machine vision
Ultra compact industrial streaming cameras in accordance with GigE vision standard
picoCam2
Ultra-compact, industrial streaming cameras with GigE interface
picoCam

 

Problem-free cooperation of the parties involved

According to Klaus Keitel, the collaboration between SICK and Nestlé ran smoothly: “We conducted tests both in our laboratory and on site. This enabled us to demonstrate to the customer right up front just how confident we are with the solution.” Thanks to this intelligent type of automation, Nestlé achieved an extremely high reliability in placing the measuring scoops while at the same time simplifying the implementation and offering flexible and expandable usage possibilities. That’s why SICK will be focusing even more in future on applications in the area of AI-assisted image processing.

 

Read more

 Tailored made sensor system for logistics with a clean effect

Read more

4Sight Automatic Print Inspection System with SICK AppSpace “gets the message”

Read more

Every movement in sight: The camera-monitored assembly facility at RIBE Anlagentechnik

Read more

Do you have any questions?

Contact our experts

We make Machine Vision Accessible
Customizable and easily configured 2D and 3D machine vision solutions – driven by SICK AppSpace
We make Machine Vision Accessible
Customizable and easily configured 2D and 3D machine vision solutions – driven by SICK AppSpace
We make Machine Vision Accessible

Customizable and easily configured 2D and 3D machine vision solutions – driven by SICK AppSpace

Learn more