AI Monitoring of Laser Welding

Laser welding has become parti­cularly well established in highly automated production, for example in the automotive industry, because a laser operates with virtually no wear, is very fast and offers high precision. But until now, the quality of a weld seam could only be documented retro­spectively, either by means of X-rays, magnetic analysis methods or by dissecting indi­vidual samples from production. Real-time monitoring of the weld quality would be a major advantage.

X-ray images of a laser beam hitting an aluminium plate, taken at the European Synchrotron ESRF in Grenoble. (Source: Empa)

While in conduction welding only the surface of the material is molten, in deep pene­tration welding the laser beam pene­trates quickly and deeply into the material and produces a thin hole filled with metal and gas vapors. If this keyhole becomes too deep, the vapour pressure of the metal vapour decreases while the surface tension of the molten metal increases. The keyhole becomes unstable and can even­tually collapse, leaving a pore in the weld seam – an unwanted fault in the material. It is therefore important for the quality of laser welding seams to detect the moment when the keyhole becomes unstable. This has not been possible to a suffi­cient degree until now. It was only possible to look into the keyhole from the top using optical methods.

A group of Empa researchers led by Kilian Wasmer has now succeeded in precisely detecting and docu­menting the moment of instability in laser deep pene­tration welding. To do this, they are using an inexpensive acoustic sensor on the one hand, and measuring the reflection of the laser beam on the metal surface on the other. The combined data are analyzed within only 70 milli­seconds with the help of arti­ficial intelli­gence – convolutional neural network. This allows the quality of the laser welding process to be monitored in real-time.

The researchers recently demons­trated the accuracy of their monitoring method at the European Synchrotron ESRF in Grenoble. Using their laser, they melted a keyhole into a small aluminium plate, which was simul­taneously scanned by hard X-ray radiation. The entire process, which takes less than a hundredth of a second, was recorded with a high-speed X-ray camera. The result: the indi­vidual phases of the welding process could be correctly identified with more than 90 percent certainty.

Once the laser beam hits the metal, the first phase of the heat conduction welding process begins – only the surface is being molten. Subse­quently, a stable keyhole is formed, which wobbles with longer exposure times. Sometimes the keyhole spits out liquid metal, similar to a volcanic eruption. If it collapses in an uncon­trolled manner, a pore is formed. All these phases can now be detected in real-time. In some experi­ments, the researchers succeeded in creating pores in the weld seam and then closed them again with a second laser pulse. The formation of the pore could be detected with 87 percent certainty, successful removal with as much as 73 percent. This method of error correction is extremely promising for laser welding. Until now, the pores in a weld seam could only be detected after completion of the work. Now, the location of a pore is already known during the process. Post-proces­sing with the laser can be started immediately.

The moni­toring process can be used not only for laser welding, but also for quality control of 3D-printed metal parts. In the powder bed process – one of the most common methods used in 3D metal printing – a laser beam passes through a layer of metal beads and welds them together. If pores appear, the laser could be directed towards the defective area a second time to remove each pore subsequently. However, this can only be done with the help of real-time moni­toring, because any pores that have been created must be eliminated immediately before they are covered by further layers of metal.

“One advantage of our moni­toring method is that the acoustic and optical sensors we use are in-expensive and robust and can be easily retro­fitted in existing industrial plants,” says Kilian Wasmer, who coordinated the work. His colleague Sergey Shevchik, who developed the artificial intelli­gence for signal proces­sing, is pleased with the high computing speed at moderate hardware costs. “We use graphics processors that can calcu­late several tasks in parallel. Such processors are also used in modern game consoles and are available at a reasonable price. Thus, the rapid technical progress in Play­station and Co. has helped our work a lot.” (Source: Empa)

Reference: S. Shevchik et al.: Supervised deep learning for real-time quality monitoring of laser welding with X-ray radiographic guidance, Sci. Rep. 10, 3389 (2020); DOI: 10.1038/s41598-020-60294-x

Link: Laboratory for Advanced Materials Processing (LAMP), Swiss Federal Laboratories for Materials Science and Technology (Empa), Thun, Switzerland

Speak Your Mind