Korbinian Weiss holds a very interesting position at TRUMPF. Not only is he a team research and development manager, and involved in machine vision, he is also the team R&D manager for artificial intelligence (AI) at TRUMPF, one of the leading technical companies in manufacturing. Recently we caught up with him to talk about TRUMPF’s activity in AI and how that is coming to the market.
Fifth Wave Manufacturing: TRUMPF has such a history with sensors. AI relies on it too, and the model might look like this:
Receive input > consider/analyze input > decide on the best direction > take action
You’re collecting sensor input and combining it with what you know, or what the system knows, and that’s how AI decides on a path to take and then how it initiates that action.
Korbinian Weiss: Processing sensor data is an essential use case for AI, as it allows a machine to gain a deeper understanding of its environment. Sensors come in many forms, cameras being one example. Cameras are relatively easy to explain because their operation is easy to visualize. For example, if you have an image, you need to determine the location of a part or its position within a machine. AI is crucial here because it allows us to deal with various environmental factors more effectively, reducing our reliance on strict environmental controls.
The other question was about the process: “receive input, consider and analyze it, then decide on the best direction and take action.” That’s essentially how these systems work. Additionally, we would emphasize the importance of preprocessing the data, which is an essential step in this process.
We have multiple stages of data analysis involving various algorithms and AI systems. Ultimately, a decision algorithm generates an output, such as a number or an offset. This value is then passed to the control system, which determines the appropriate action. For example, the control system might move the cutting head based on the input received. This exemplifies AI’s strength in process optimization.
FWM: I’m a parent of three grown children and you want them to be happy and do a good job navigating the world. At times, I’ve thought this though: What haven’t I taught my children? In an AI system, you can’t say: What haven’t I told my system? With your children, you might know about a particular circumstance you forgot to tell them about, but with AI, yes, you need to seed it with data, but on the other hand there is a lot of self-learning that happens after the pump has been primed.
KW: It is similar to human learning. Training an AI requires a system of positive and negative feedback to reinforce correct responses. You feed data into the system, for example, you show the AI an image of a dog and label it as such. If the AI misidentifies a dog as a cat, you correct it by providing feedback. This process helps the AI learn to accurately distinguish between different inputs. It takes a lot of training, but over time the system gets better.
An AI system is somewhat like a child. You can anticipate its reactions to some extent, but it takes extensive testing to achieve high confidence in its precision and response to various inputs. You need a large test set and must probe the AI to observe its outputs, determining whether they meet expectations or if the system behaves unpredictably or irrationally—sometimes even hallucinating.
AI systems are often considered black boxes, so we must thoroughly probe them with a variety of use cases, including edge cases and common scenarios, to ensure they align with the intended application.
Yes, some of our systems learn continuously and we use incoming data to fine-tune them. Sometimes we finish training a system and consider it done. However, as we encounter new edge cases, different processes or new materials, the robustness and accuracy of the system may degrade over time. In such situations, retraining becomes necessary.
FWM: Here is a topic that will be very interesting to my readers. TRUMPF has a subset of customers who have said they are interested in being a beta tester for your equipment, along with the AI component. I have a few questions about that: 1) How do you pick the people; 2) at what point along the path of being a finished product do you roll a beta program out to these users; and 3) what has been the response to the beta program?
KW: The way we select beta customers is always very focused on the feature we want to test. For example, if we were building a system for cutting thick sheet metal, we would look for customers who have a portfolio of thick sheet metal that they run in their shop, rather than a customer who is more focused on thin sheet metal. We’re trying to find a customer that’s going to get a lot of value and a lot of use out of this new feature, because that’s really what we’re going for. We want a customer to use and test it extensively and we want them to give us a lot of feedback.
In setting up a beta test program, we have to take into account things like machine downtimes for updates, distances, and hardware that we might have to ship and install. That’s always something we have to focus on, and those additional parameters go into the selection process.
Different products have different beta stages, but with AI systems we’ve always worked with beta customers where we get the system out early and get real-world feedback because that’s something we need. It’s quite striking what we can test internally versus the diversity of what our customers are actually running on these machines. We need that diversity of testing.
Sometimes we come across an edge case that we did not expect our customers to use. So, it’s extremely valuable and extremely important for us developers to have this phase and to test the systems.
The general reactions during the beta phase are very valuable. Of course, the feedback we get can vary from customer to customer. Some customers may suggest additions, while others may say, “This is exactly what we need.” In addition, some may request specific adjustments.
FWM: Let’s talk about choosing customers for beta programs. It certainly wouldn’t hurt if you were already a blue shop, right? The reasoning behind it is that you have more of our equipment and you could take advantage of this pool of data that you’re collecting in your shop. And in fact, TRUMPF corporate can also collect that data and does so. There is some high percentage of shops that are having TRUMPF help them optimize their equipment this way. Because you do need data for AI, people who have the data, and have the equipment to take advantage of what TRUMPF has already done, will get the most out of it.
KW: That’s a very important issue: how to get the most out of the machines and the systems that we sell. It is extremely important to us as a company, and we try to do that on many levels. For me, it’s the development side of things where I have my influence, and that’s where we try to build a very good and solid setup for our customers to work with. That’s what I’m responsible for running. And if we have some issues, I like to have this feedback and improve the systems. After that, of course, it’s very important to ship updates out to our customers. So, we do over-the-air software updates and with that we can fix bugs, add features, add performance and productivity to our machines.
Adding performance and optimizing features, is only possible with a data-driven workflow. This approach relies on the systematic analysis of data to inform decisions and improvements. By using data, we can identify areas for improvement, predict future problems, and develop new products that are better aligned with user needs and operational efficiency. Data-driven workflows enable us to continuously refine processes and develop solutions that are both innovative and highly effective.
FWM: Yes. That brings up another thing for your client base: people are very used to going to a trade show and seeing what’s new once a year, right? From a machine tool point of view, that was the cadence of the changes. And it seems like all bets are off now because we’re becoming a very software-oriented business and now we’re starting to become an AI oriented business.
KW: It’s definitely worth going to trade fairs because they feature new machine types and concepts and additions. All of this will be at the big shows.
Sometimes the solutions to customer needs are more complex, and to keep the solution simple for the user, we try to address that complexity with AI systems. We try to build systems that help the operator to run the machine as effectively as possible, so we try to add AI not to make it more complicated, but rather to make it easier to use. If it’s too complicated, it won’t be used in the end. Ease of use, access and awareness of the features and functions on the machines are important issues for us.
FWM: Communicating the simple in a complex way is easy; communicating the complex in a simple way is much more difficult. And you’re doing the latter, thank you. Any final thoughts on AI?
KW: We are going to see a lot more AI in the manufacturing business in the future. I think what we’ll find out is this: we will have overestimated what will happen in two years, and underestimated how AI will transform manufacturing in the next ten years.