Collaborative Project »DIPOOL«

Learning production technology – using artificial intelligence (AI) in production

Manufacturers are increasingly facing several challenges: They have to shorten development, production and delivery times, respond to growing cost pressure and fulfil higher quality requirements. To meet them, they are using modern production equipment as a key component in their manufacturing processes. At the same time, the processes used to develop such equipment are also changing rapidly due to digital transformation. In summary, production technology must meet the increasingly volatile demand for varying products (material, design, function) with the highest possible efficiency and flexibility. Laser machines are predestined to meet these challenges. However, they also require increased automation to make their operation economical. The joint project DIPOOL, scientifically coordinated by the Fraunhofer Institute for Laser Technology ILT, aims to improve the autonomy and agility of such systems by combining the unique temporal and spatial programmability and controllability of lasers with suitable machine learning (ML) methods. The goal was to achieve automatic and robust monitoring, quality assurance and to optimize laser machines for changing production tasks.

Smart laser systems

For the cutting process, AI system monitoring is used on the demonstrator system developed by Automatic-Systeme Dreher GmbH in the project for laser blanking, the continuous cutting of blanks from a conveyor belt. The system has two laser cutting heads that can be used to process a joint area of approx. 5.5 x 1.6 square meters, thereby delivering high productivity and flexibility thanks to this arrangement and the high-speed cutting process. A laser robot cell from LBBZ GmbH serves as a demonstrator of digital optimization technology for 3D welding.

In its innovative, scientific approach, Fraunhofer ILT is imposing "minimally invasive" laser modulation patterns on the machining process. The process responds to this continuously with particularly characteristic, condition-dependent signals. These response signals were recorded using a smart, newly developed cutting head from Precitec GmbH & Co. KG with a sampling rate of 50 kilohertz and then evaluated using AI methods. Various ML methods were tested in DIPOOL by the Institute for Industrial Information Technology IIIT at the Karlsruhe Institute of Technology KIT, with convolutional neural networks (CNNs) achieving a classification accuracy of 97.9 percent. For the monitoring, decision-making and control modules, Marx Automation GmbH integrated the AI model into a process computer with a field programmable gate array (FPGA) for data acquisition, pre-processing and controlling the laser. The cycle time for data acquisition, pre-processing and inference is 1 to 2 milliseconds, so that real-time capability can be achieved even with the fast processes involved.

During laser beam welding, the multispectral sensor developed by 4D Photonics GmbH is used to record and evaluate the emitted process radiation in 16 channels each in the visible and near-infrared spectral range in a manner selective to the wavelength. Various methods of stochastic and multispectral analysis were qualified and quantified in the course of the project to detect or localize an unintentional under- or overshoot of the weld penetration limit at an early stage. Using a selection of varying alloy compositions and weld seam configurations, the institute evaluated the multispectral sensor for use as an AI-supported process monitoring method. In extensive tests under controlled conditions, it was able to reproducibly detect when a machine either exceeded or fell short of the weld penetration limit.

The laser cutting and laser welding solutions developed in DIPOOL show that AI-supported methods and innovative techniques for improving signal quality and the information content of process data can significantly contribute to process safety and reliability.