David Fehrenbach

David is Managing Director of preML and writes about technology and business-related topics in computer vision and machine learning.

The Digital Hub Applied Artificial Intelligence Karlsruhe collects use cases of AI from practice. With preML, we are a partner of the de:hub Karlsruhe and are very happy that this initiative exists. I put myself to the Hub team’s questions about our AI-based quality inspection of concrete elements and don’t want to withhold the answers from you. Enjoy reading!

What tipped the scales in favor of developing an AI-based use case?

Suppliers of automatic visual quality control systems have been around for some time. These focus in particular on the automotive, food and electronics sectors. Some industries have been passed over, for example the construction industry. The main reason for this is that there are unfavorable conditions for computer vision systems here: Dirty environments, large components, high product variance and customer-dependent quality requirements. Take, for example, one of our systems used to inspect concrete tunnel segments. Component production is different for each new tunnel, depending on diameter, tunnel boring machine, geology and country-specific quality requirements. By using AI algorithms, we were able to create a system that meets these requirements, is robust against anomalies, and can quickly adapt to changing conditions.

What were your expectations or requirements for the AI application?

Our first customer for the system had a precise idea for defects that needed to be detected. He could draw on experience from hundreds of projects, which was reflected in a very detailed specification. In addition to anomalies of any kind, various tasks such as the presence of built-in parts and safety-related issues were to be checked. We check these requirements very specifically and compare them with similar projects we have done in this area and current (and published) results from science and research. In the case of automated visual inspection tasks, we set ourselves the expectation that our systems will solve the task better than a human could. If the system wins this competition, then a package of advantages results directly: One is faster and cheaper, more consistent and one has the results documented directly and comprehensibly.

Which partner did you rely on with which technology for the development?

In our first project, our customer was the world market leader for technical systems in its field. This partnership was (and is) extremely beneficial for us. Through the contacts and network, we were able to get initial training data, we were able to access expert knowledge in the field, and it brought us the needed trust from the end user in the solution.

The use of AI solutions must also be accompanied by a change in competencies. How have you dealt with this challenge in relation to your employees?

As a startup, we are AI experts and naturally prepared for such projects. However, we notice that the topic of AI is often hyped in the industry, which we see as a danger. This can quickly fuel false expectations. For example, many expect self-learning components in their applications because this is familiar from the media. That’s why we always try to explain the basics of AI at the beginning of a project and are also glad that there are institutions like de:hub Karlsruhe that do a lot of educational work.

In what timeframe did they implement the AI solution in your company?

We have already developed Proof of Concepts (PoC) in one week, which solve problems with real image material using AI algorithms. Overall, however, we work for industry, which means that processes can take much longer. Sometimes we are already involved in factory planning, but it can take two years until the factory is built and production starts. If the factory is already built, about 3-4 months is a good planning horizon.

A conclusion to your AI solution:

In our view, AI is an important driver of automation processes in industry. Our solution shows that medium-sized businesses can also benefit from AI – especially from individual automation solutions. The advantages of automated visual quality inspection with AI are obvious for our application:

  • The inspection is standardized according to clearly defined quality guidelines.
  • The inspection takes about 20-30 seconds including full documentation
  • Current quality metrics are captured in real time and reports are generated automatically
  • People are kept away from potentially dangerous situations
  • Position defense as an innovation leader in the industry.

One final piece of advice/tip to other companies looking to apply AI:

We recommend that companies always follow the path from the problem to the technology and not from the technology to the problem. The goal should not be an application of AI, but the best possible solution to a problem. Then AI can be a powerful tool, but it is far from being the solution to all problems. To use AI efficiently, we recommend ensuring data availability and considering the long-term use of an AI application from the beginning. In the area of automated visual quality checking, preML is happy to help you with this.


David Fehrenbach

David is Managing Director of preML and writes about technology and business-related topics in computer vision and machine learning.