Issue link: https://iconnect007.uberflip.com/i/1424540
NOVEMBER 2021 I SMT007 MAGAZINE 29 tensively use on our inspection systems. But AI capabilities go beyond advanced machine vi- sion, and we are already working on some ad- vanced AI applications that will further revolu- tionize the implementation of inspection. Johnson: In the beginning, inspection setup and learning used a board or a finished assem- bly as the gold standard. You created a refer- ence image of that, and then as long as every- thing matched that, it passed. Right? Vey: No! Johnson: Alexia, you're wagging your finger at me for that comment (laughs). Talk about how it should be. Vey: Mycronic has always refused to follow this image comparison system. is is the rea- son why our AOI has always been, and still is, one of the most stable AOI systems on the market. Because if you only compare images, then you are not robust to all the variations be- tween PCBs, between PCB suppliers, and be- tween components, and this will generate false calls. ere are some systems where you teach the golden board one time, then you run your batch; it's fine, but one month aer that you run another batch and it's not working any- more. e reason is simple: this is due to all the supply variations, and we must cope with it. e way we have addressed this issue is by not comparing images, but by having an algo- rithm-based system. We do not use a golden board because it does not exist in reality. No customer can place components perfectly. Instead, we teach the system what it should find, and then we check that it finds it. But we teach it in a vectoral way, which means that the system receives more in- formation on the theoretical model than neces- sary, but then it can adapt to the variations. Of course, I mean, this was more the 2D part, but since we introduced 3D, the algorithm got so much simpler, because you don't have to ques- tion whether the reflection you see is the com- ponent, some solder, or something else. With 3D, the component is here or it's not present, with no image comparison at all. I know that not all AOI manufacturers have this strategy, but we went to this strategy and history has shown us that this was the right choice. Just think of all our key accounts who have been us- ing our AOI technology for more than 10 years on multi-lines with the same library, with ex- tremely stable results. Johnson: Tell me about the results information that gets provided back. How accurate is it in spotting false positives or false errors? Does the system become more accurate and more precise over time for your users? Vey: Yes. We usually have a strategy of a cen- tral library for all products. is way, you have a learning curve, and aer a few products, your library stability increases. You will get fewer false calls. Some high runners in the automo- tive sector with very high-test coverage have managed to have more than 95% FPY with our system. Johnson: When you're talking about the com- ponents in the library, are you talking about the board assemblies as a component, or the individual chips that are going on the board? Vey: e individual chips. We create a library model for each individual package type. Dowd: ere are a couple of layers to this tech- nology. Within the AOI and the SPI equip- ment, we have what we like to call a vision en- gine, which we have been developing for years to do this analysis. It has proven to be strong in the industry; where we have deployed it, it has been beneficial to our customers. And in the traditional ways of inspection, right? Pre- venting defects or detecting defects and help- ing reliably weed them out and not get stuck with too many false positives.