29 Jun 2023
High-speed imaging and acoustic signals combine with neural networks to evaluate automated joining on the fly.
by Mike Hatcher in Munich
Artificial intelligence (AI) in the form of neural networks trained to recognise welding and brazing defects offer a route to automated laser materials processing systems - although training those networks can be time-consuming and complex.
That was the take-home message following a technical forum session at LASER World of Photonics featuring three presentations dedicated to the topic.
First up, Christian Knaak from the Fraunhofer Institute for Laser Technology (ILT) outlined the challenge of assessing weld seam quality during laser processing of zinc-galvanized components used in the automotive industry.
Using three high-speed imaging cameras, including one thermal imager, Knaak and colleagues set about training a neural network capable of predicting defects like cracks and pores.
They based their algorithmic model on no fewer than 43 welding trials, generating tens of thousands of images per weld in the process, finding that a near-infrared camera capturing 1690 nm light gave the most accurate predictions.
With such huge data generation the speed and throughput of the system is critical, but Knaak said that a latency of 4 ms could be achieved at data rates high enough to yield predictions of eight different defects occurring - including spatter, seam collapse, and incomplete joining - as the weld progressed.
Imaging might be the most obvious way to track and predict weld quality, but it isn’t the only way to do it. Andreas Wetzig from the Fraunhofer Institute for Material and Beam Technology (IWS) in Dresden said that his team had also been evaluating acoustic signals captured by laser microphone as a monitoring technique.
Focusing initially on cutting applications, they trained a neural network model in collaboration with software experts at Fraunhofer USA, finding good predictions of key characteristics like surface roughness, kerf width, and dross formation.
That AI model was able to interpret signals from the acoustic data such as maximum amplitude and spatial period, yielding predictions of cutting quality. The IWS team is also working on welding applications, although Wetzig said that this was at an earlier stage of development.
In his opinion, the kind of real-time feedback offered by AI and neural networks trained on extensive data “will be implemented in real-world applications soon”, with likely deployments in the automotive, aerospace, and other sectors.
Germany’s Scansonic is one company that is already testing the market for such products, with Michael Ungers representing the Berlin-based firm at the Munich forum.
Fewer inspection stations needed
Discussing the use of image-based AI feedback in laser brazing applications, he reiterated the emphasis on extensive data training for accurate modeling. Scansonic’s “SCeye” product, which is being demonstrated at this week’s trade show, is said to be capable of spotting pores less than 0.5 mm in diameter and spatter of 0.2 mm.
The plan is to transfer that knowledge to welding applications, with Unger noting significant differences depending on the material being welded.
Adapting the approach for laser welding is also the subject of the BMBF-funded “DIREAL” project, whose participants include Fraunhofer ILT and welding systems firm CLOOS.
Ungers pointed out that the integration of defect-detecting AI into laser welding optics would have a cost benefit as it would significantly reduce the number of inspection stations currently used on a production line.
And, like the other speakers in the session, he emphasized that the main challenge was the sheer amount of training data required for each process.
Nevertheless, the overall message from forum chair Peter Abels of Fraunhofer ILT was clear enough: “AI in laser processing is not just a dream,” he said. “We can already buy such systems.”
ILT is hosting its own conference dedicated to the subject later this year - and with the likes of Trumpf and Microsoft set to take part, it looks like the era of AI-assisted laser materials processing is up and running.