How artificial intelligence optimizes laser processing

Artificial intelligence is currently a hot topic in all sectors – including laser technology. How to choose an AI solution for better processes.

laser processing
laser processing

How can data and algorithms be used to increase the quality and efficiency of processes in laser processing? Researchers, laser technology manufacturers and users are asking themselves this question at the same time, because with the emerging Industry 4.0, with solutions artificial intelligence (AI) are also becoming increasingly important. The first solutions in the field of laser technology that uses artificial intelligence already exist.

How is artificial intelligence used in laser material processing?

For example, machine builder Trumpf uses artificial intelligence to control laser systems using voice commands. With the system equipped with a marking laser, the system operator can issue all relevant control commands such as ‘Open/close the door’, ‘Start the marking process’ or ‘How many products have you marked today?’ speak directly into a microphone. The laser system responds accordingly and executes the voice command. You can find out more about the machine in our article ‘Trumpf: AI moves into laser material processing ‘.
However, significantly more researchers and manufacturers are dealing with AI systems in the field of quality assurance – some AI applications are already being developed, especially for laser welding. Most of the time it is about the quality control of the welds. The Fraunhofer Institute for Laser Technology (ILT) is also researching in this area.
“We want to use our AI solution to recognize different quality categories in laser welding,” reports Christian Knaak. The scientific employee of the Fraunhofer ILT deals with Zero Defect Manufacturing and in this context with Deep Learning and spoke about it at the first ‘AI for Laser Technology Conference ‘ in Aachen.
“We want to detect seam collapse, binding errors and incorrect seam widths,” Knaak continues. “And we also want to recognize when the seam is in order.” For this purpose, the measurement data from welding processes – in this example the camera images of the weld seams – are evaluated and the system is then supposed to make statements about the quality of the seams. There are two possible approaches: the classic machine learning approach and deep learning.

What do I need for a successful AI project?

“AI needs a sustainable database. And a team that implements the project together and in which the individual members have the necessary specialist knowledge from the departments involved.”
Christian Kohlschein, Hotsprings GmbH
 
“Machine learning sometimes needs bad process results. If too many results are good, then the algorithm says that everything is good. But then the AI ​​​​solution is useless.”
Stephan Schwarz, Mercedes Benz AG
 
“The biggest challenge when introducing AI solutions is the cultural transformation of the company. Faster technology and better algorithms alone cannot make a difference.”
Benjamin Kreck, Microsoft
 
“When selecting suitable algorithms for quality analysis, the domain knowledge of the employees is required. This is the only way to find suitable classifiers and algorithms.”
Christian Knaak, Fraunhofer Institute for Laser Technology

What is the difference between classic machine learning and deep learning?

In machine learning, knowledge about the area of ​​​​application (so-called domain knowledge) plays an important role. Image processing algorithms are used to extract and classify features from the camera images. “We have five classes,” explains Knaak. “These are typical images of welding tests with examples of lack of fusion, seam collapse, increased night width, OK seams and no seam.”
The features of these images are then used as a ‘fingerprint’ for the various seam conditions to be recognized by an algorithm. The appropriate algorithm is selected by determining the recognition rates of various algorithms using a training data set. The algorithm with the best detection rate can then be used for the process data.
Deep learning requires less domain knowledge but requires significantly more data. “The special thing about this method is that the task of extracting features, which was taken over by image processing algorithms in classic machine learning, is now handed over to a neural network,” explains Knaak.
This neural network recognizes the features and processes the data in several layers. This makes it possible to analyze very complex image features. “The end result is the same as with the classic approach, where you need a relatively large amount of domain knowledge,” reports Knaak. Both methods are therefore suitable for analyzing the seam quality, only the prerequisites differ.