|
There are many distinct commercial sensor systems for monitoring or controlling laser processes. Most of them are based on camera or photodiode technology. Furthermore, the introduction of real in-process measuring systems in laser material processing like OCT has significantly increased the safety of both defect detection and process control [5][6][7][8][9]. The focus of this contribution, however, relates to the use of artificial intelligence algorithms to "See New Things". We will discuss how classified, physical properties can be derived from already reliable process information - "seeing the unseen", so to speak. Instead of defining complex rules for algorithms, the use of Data Science and Machine Learning methods reveals hidden structures in noisy unstructured data and make it possible to find the relationships of the data to the physical measurement. There are several systems available in the market which capture or measure and analyze physical effects of the process zone and their properties during the laser processing, but none of these effects and properties stand for “seam quality” for themselves, which is typically defined by mechanical, geometrical and metallurgical properties of the solidified seam. The process properties like emitted visible and thermal radiation or geometrical values of the melt pool or the penetration depth of the laser forming the keyhole (penetration depth) only provide indications of how the desired quality might be achieved. When welding thousands of seams per shift during serial production for Li-Ion batteries in e-mobiles, very often users would like to stop their fully automatic machines once the laser process is suddenly running out of previously set boundaries. Such an approach allows users to find systematic faults where and when they occur, rather than at the end of the line where quality inspection takes place and the line is still running, producing more and more parts with the same faults. When it comes to the use of artificial intelligence algorithms the following questions must be answered: ▪ Is it possible to map an input time series to a real value? ▪ Can we get out the strength of the weld from the process emissions? ▪ Is the information buried in the signal? The vision, which can be represented by Figure 1, is that by using the data from process emissions and specific data models, a relationship to physical quantities can be established. Thus, AI goes a significant step beyond the simple "good-bad" statement.
|
