How do you proceed from there?
Petring: “The process is called feature engineering: We don't feed the AI with all the process data, but with intelligently prepared data. The trick is to find the best compromise between data quantity and data quality because you can also overtrain a system in a machine learning environment. If the trainers overdo it with too much data, then the AI will only learn by rote. And, as soon as data deviates from what has been learned by rote, it is overwhelmed and can no longer make a sensible decision.”
How do you ensure data quality?
Petring: “We don't constantly unleash the laser power on the process, but modulate it minimally. We periodically change the power very slightly and see how the process responds. To do this, one would ideally assume linear response behavior and build control loops on that. Unfortunately, this does not work for laser material processing because we are dealing here with strongly nonlinear processes. The processes react comparatively “sensitively,” but also quickly to parameter changes. Accordingly, the process signal response also changes quickly depending on the process status with a characteristic signal pattern. We and, in the future, AI will exploit this, for example, to recognize in real time whether the control system can increase the cutting speed even further or should perhaps even reduce it somewhat in order to achieve optimum cutting quality. Without this modulation technology, only the signals of a so-called unquestioned process are generated, which can quickly be misinterpreted or not interpreted at all. This is why the integration of process monitoring in industrial laser machines has failed very often in the last 30 years.”
So the goal is to increase the reliability of interpretation?
Petring: “Exactly. At the moment, we are therefore primarily focusing on supervised learning in the DIPOOL project: We are testing the response behavior with minimally invasive laser modulation (MILM). With it, we obtain characteristic signals that are easier to interpret. In supervised learning, we thus train the ML system, usually an artificial neural network, to draw the right conclusions from the signal patterns and to initiate appropriate measures if necessary: They can range from instructions to the plant operator all the way to the automatic adjustment of process parameters. The trained “model” is then integrated into the plant control system as a so-called inference engine.”
Which sensors do you use?
Petring: “For example, multispectral sensor technology with different wavelengths, which is being developed by 4D-Photonics. Additional information can be derived from the different signal heights at the various wavelengths. The broad color spectrum is comparable to a dynamically changing rainbow, which AI is predestined to evaluate. For laser beam cutting, Precitec specifically addresses the boundary conditions in a 2D flatbed cutting system during sensor development. Compactness and, above all, robustness of the solution are top priorities, in addition to the highly dynamic acquisition of the process responses.”
Do you see any chances of convincing small- and medium-sized companies of these solutions?
Petring: “Today, a system of this type can also be implemented for SMEs because the prices for laser and computing power have fallen significantly. Virtually all manufacturing companies are now under constant pressure to innovate – not least from competitors in China. This increases the willingness to invest in innovative technology. In addition, the process understanding for AI systems is also growing among medium-sized companies.”
Please give us an assessment. How much more efficiently can companies cut or weld with lasers thanks to AI control?
Petring: “That's a simple question, but of course it can’t be answered universally. Nonetheless, I roughly estimate that the overall equipment effectiveness (OEE) can be increased by 25 percent.”
The DIPOOL project will run until summer 2024. Is there already an end user in sight?
Petring: “Yes, the BILSTEIN GROUP has even set up its own subsidiary BILCUT in Hagen specifically for this purpose, which will be cutting shaped blanks for the automotive industry at high speed starting in 2025. DIPOOL project partner Dreher will build a laser cutting system based on the technology we have developed. This will enable BILCUT GmbH to supply individually designed blanks with excellent economic and ecological performance.”
BMBF research project DIPOOL
- Project funding in the program "Future of Value Creation - Research on Production, Services and Work”
- Project duration: July 7, 2021 to June 30, 2024
- Project Management Agency: Project Management Agency Karlsruhe (PTKA), Karlsruhe Institute of Technology (KIT)
- Automatic-Systeme Dreher GmbH, Sulz a. N.: Project coordination, creation of the requirements profile for a smart laser blanking line, construction of a prototype
- LBBZ GmbH, Geilenkirchen: Creation of the requirements profile and evaluation of learning 3-D laser welding technology
- Precitec GmbH & Co. KG, Gaggenau: Integration of smart system technology and sensor technology for smart laser cutting
- 4D Photonics GmbH, Isernhagen: Combination of multispectral welding process sensor technology with AI methods
- Marx Automation GmbH, Düren: Implementation of machine learning algorithms in laser machine control systems
- Fraunhofer Institute for Laser Technology ILT, Aachen: Development of robust process control for learning laser machines
- Karlsruhe Institute of Technology (KIT) - Institute for Industrial Information Technology IIIT, Karlsruhe: Development of an efficient signal analysis with machine learning