sinterqAlity: Machine Learning for Optical Quality Control

7 min readMay 30, 2022


Franz Tschimben and Thomas Villgrattner at DURST Campus, Brixen, South Tyrol (Italy)

GKN Powder Metallurgy and COVISION Lab have been collaborating on a project now for more than 1.5 years. The project and findings have the potential of changing the application of optical quality control in manufacturing: With sinterqAlity, machines do not only learn what makes a good or bad part, but contribute to improved, more sustainable output. Some weeks ago, around 20 people from different AES member companies had the opportunity to deep dive into the topic on site at DURST (one of seven founding companies of Covision Lab) and discuss the project and vision for the future with Franz Tschimben and Thomas Villgrattner…

What exactly does machine learning mean?

TV: Machine learning aims at automatically detect and generalize correlations in datasets. This approach works well if t a sufficient high amount of data with enough variance is provided.. After the training phase, the model can be used for predictions.

FT: That is correct. Machine Learning is generally also publicly referred to as “artificial intelligence”. Machine learning does nothing other than find correlations in data. Our joint SinterqAlity project is about image data — here, our software solution automatically finds defects. Unlike classical visual inspection systems, in the SinterqAlity project no defect detection parameters are programmed and hard-coded by experts, but the Covision Lab software programs this defect detection by itself.

What is special about the application in your case — and how does “sinterqAlity” work exactly?

TV: In “ordinary” optical quality inspection systems, each relevant defect pattern must be generally programmed by hand. If a new defect occurs, it must be added manually, otherwise these “new” bad parts will not be detected.

The goal of sinterqAlity is to automatically learn the “good part” behaviour without any additional intervention of a programmer. For this machine learning is used to extract the relevant features of images of approximately 50,000 parts. Once the good part is deduced it is simple: Everything that is not good is bad. So, if a new defect image appears, it is recognized immediately and the bad part is sorted out without any further intervention of a programmer.

FT: Right, what Thomas describes is commonly called “unsupervised machine learning”. This is also the main feature of our visual inspection software and our differentiator in the market.

In contrast to this is “supervised machine learning”, where the software is trained on the basis of many annotated or labelled images.

How did the idea of launching a joint project come about?

TV: GKN Powder Metallurgy is always open to new technologies and is continuously looking for new project partners. Our CDO Paul Mairl established the contact with Covision Lab. The project was and is actively driven by our top management. In that case, through Peter Oberparleiter (CEO at that time) and SVP Elmar Auer.

FT: The top level management support by Elmar Auer, Peter Oberparleiter and the intuition of Paul Mairl was and is essential for the speedy developments and results.

It is a unique opportunity for Covision Lab to work in such a close way with GKN and its people. We are learning a lot about the challenges in quality control and how artificial intelligence can ideally be used to solve them in an efficient and cost-effective way. For us, the project with GKN — where we now also work with the teams at GKN in Bad Brückenau, Germany and in St. Mary’s, United States — was the starting point for the launch of a subsection of Covision Lab, namely Covision Quality. Here we have the goal to establish a scalable software product on the market.

What requirements are necessary so that such a use case can not only be identified, but then implemented in the form of such a project?

TV: After initial clarifying discussions with the new partners and responsible personnel involved, a compact test dataset is made available. If the analyses are promising, a pilot project can be launched at a GKN plant. For this, the technical and financial potential must outweigh the costs and challenges. Once the pilot is completed, the costs and benefits of the new approach will be analyzed again and further project phases initiated if suitable. In the best case, generalization and global rollout take place and the new technology is used as a new standard in all GKN plants and for all products.

FT: I completely agree with that. The starting point is always the use case and the size of the pain point and challenge for the current visual inspection process. That’s where we can start and show our added value quickly — under the conditions mentioned by Thomas above.

What hurdles do you face in developing such a project — and what would you do differently in retrospect?

TV: GKN always starts pilot projects directly in production and often at several sites at the same time. This allows testing the new technology under real conditions, but has to cope with production priorities. For better coordination, colleagues who work in production departments are directly involved in development. In future, we will strengthen the collaboration of development and production departments as well as expand our platforms for digital collaboration.

FT: As a joint team, I think we quickly learned from the experience of the 1.5 years of cooperation and created a “blueprint” with a checklist for new installations. This is to be worked through step by step in order to create not only the ideal conditions on the technical side but also, above all, acceptance with the teams on the shopfloor, in the plant and in production. Furthermore, in recent months, based on our experience with GKN, we have invested in making the user experience (UX) better. We launched an interface for the team close to the production lines and a management panel for quality control managers to oversee key performance indicators. The software is intuitive and easy for everyone to use.

If another company now wanted to implement the same project in its own production, would that already be possible today?

TV: If other companies can provide the necessary datasets, they are welcome to use Covison Lab’s software. We as GKN are “early mover” and actively support Covision Lab in developing their software and distributing it to other companies. Moreover, we are also actively promoting Covision Lab’s software to our peers.

And what would that cost?

FT: There are typically three phases of a project to install our software. In the first phase, based on a few data points of the company (about 200 images), a first free test is performed to show the potential of our software.

In a second phase, if necessary, our software is installed on a production line. Here the collaboration between the quality control specialists from the customer and our A.I. team will be started and awareness will be created in the company with good results. This phase is billed at a monthly rate.

In the third phase, there is a broad roll-out on different production lines in different locations. The most suitable business model in this phase is “Software-as-a-Service” (SaaS) where the customer licenses the software per production line or per part. This gives the customer the right to continuous improvements of the software.

Furthermore, for the purpose of making our software run in real-time (e.g. in Bruneck we work at speeds of 200ms/ part) GPU’s are installed on-site next to the production lines.

Do you have a forecast on how this kind of approach to quality control might develop in the future?

TV: The approach has the potential to revolutionize quality control. The goal is to significantly reduce training time, especially with regard to the use of highly qualified personnel. At the same time, defect images, even those never seen by the system, should be reliably detected, while good parts are always detected as such.

FT: We are certainly still at the beginning of the A.I. revolution in the field of visual quality control. The potential is confirmed and the technology is predestined to create quality advantages, competitive advantages and savings for manufacturing companies. In the long term, quality control could become the starting point for a broader application of A.I. — not just on images, but to other production data as well. Predictive maintenance is an example.

Franz Tschimben works as CEO at COVISION LAB
Thomas Villgrattner is Head of Adaptive Technology at GKN Powder Metallurgy. and

About Automotive Excellence South Tyrol (AES)

Automotive Excellence is an association of the most relevant companies in the automotive supply industry located in South Tyrol. The aim of the innovation hub called AES is to use synergy effects for joint projects, to develop greater agility and innovative capability in the companies, to accompany them on the path to climate neutrality — and to focus centrally on the further development and qualification of existing and future employees as the core of entrepreneurial activity. Automotive Excellence Südtirol aims to contribute to positive effects for the industry as an employer and for South Tyrol as a place to work. The current member companies of AES are: Abuscom, Alupress, Autotest, GKN Powder Metallurgy, GKN Driveline, Intercable and Tratter Engineering.




The network Automotive Excellence Südtirol is an Open Innovation Hub for Sustainable Transformation in the sector mobility, automotive and beyond.