Issue link: https://iconnect007.uberflip.com/i/1502623
20 SMT007 MAGAZINE I JULY 2023 pixel matching, it can be more effective with machine learning and deep learning. All that past data gets appropriated into the model, which gets better over time. Sometimes we'll bring up a system for a client that's 93% accu- rate, but within a couple months we're at 97% because of that feedback loop. Shaughnessy: How do you start developing this? Tell us about that process. We start with a traditional computer vision model that we trained to recognize general PCBs from a proprietary PCB dataset we've been collating over the past few years. en we repurpose that "foundational model" for our customer's product. at allows us to train against a preferred particular client very quickly, given their range of PCBA products. Now, the AI model is only about 30–40% of the entire system. ere's the soware which must do all the configuration and the nice user interface you give it. ere are the optical imag- ing algorithms, which, in our case, allowed us to bring a hardware solution to market that's much lower cost than we typically see. Johnson: What other areas in the assembly line do you see your kind of AI application making sense? Where can this grow? One of the use cases is conformal coating. ere is a lot of interest in solder formation and detecting aer the reflow oven. Badly applied solder, I'm told, is a hard problem to crack. How do you determine bad from good in a solder joint? Beyond that, think of anywhere where you could benefit from increased inspection. e earlier I can detect something and prevent it from going through other parts of the work- flow, I'm saving on materials. Johnson: That sounds to me like you're starting to plug into the CFX data flow. I think we will see the emergence of real interoperability standards so that you can feed data across all elements of your line and ben- efit from those feedback loops. Some of our larger clients are asking for this feedback loop between manufacturing and design. Johnson: What other data from the line is useful in doing your job? RGB imaging data is just our starting point. ere's 3D imaging, which has benefits for large and tall through-hole components that you have to look around. ere are X-rays and other modalities. Recently, we wrote a paper on acoustic analysis. e underlying question is whether there is anything about the acous- tics of the environment that we can lever- age, or is there just too much human noise? What about vibrations of the machines—does that affect alignment? ere's a lot of data to collect. Sometimes we become too enam- ored with collecting data, rather than asking whether it's effective data. Right now, it's all the different ways of collecting that imaging data. DarwinAI VQI (visual quality inspection) system for PCBA.