AI in the factory digitalization strategy
gives electronics manufacturers the power to create the optimal production environment

The IPC APEX show in San Diego earlier this year was a huge success – we all enjoyed pre-pandemic levels of visitors to our booth and the atmosphere felt energetic and optimistic – it was truly refreshing. Beyond the positive buzz on the show floor, I was particularly impressed by the overall interest level attendees expressed in implementing Artificial Intelligence (AI) tools to improve their manufacturing operations.

For years our industry danced around the idea of AI, speculating about what it could do for PCBA manufacturing, but until very recently, there wasn’t a large pull from electronics manufacturers to put in the effort to properly give AI a try. I used to even joke that AI in electronics manufacturing was more like a science project, not something that was ready to make a tangible difference to my customers. How times have changed – everywhere I turned at the show, I either saw or heard the term AI. Personally, I am unsure if these claims can be substantiated with proven use cases, but it’s encouraging to see everyone so interested.

One thing I do know for certain is that together with our parent company, iTAC Software AG, at Cogiscan we have AI-powered tools with field proven use cases to provide advanced analytics and feedback loops for electronics manufacturers to better optimize their performance by reducing defects, maximizing output, and increasing overall throughput. These new tools genuinely have the power to revolutionize factory intelligence and prediction – I’ve experienced it firsthand.

Sadly, the vast majority of AI projects fail, with a Forbes article citing an 85% failure rate, as little forethought is put into what the AI actually needs to solve. Instead of looking at an AI or machine-learning (ML) platform as a tool to solve a specific manufacturing problem, such as attrition rate, many customers approach it as a data only project. Collect all the available data from the operation so the AI tool has data, ­and from there figure out what problems to solve.

These customers are collecting data from the shop floor and putting it into a data lake in the cloud without a clear structure – so they end up with a bunch of data that is not structured in a way that the algorithm can efficiently learn from. It's the wrong approach. There are specific requirements and steps that must be followed to properly train an algorithm and create a use-case model in an efficient manner and run it in both a cost effective and high-speed way.

To kick-off an AI project successfully there are three clear steps to follow: one, assign the AI project a purpose in advance; two, collect factory data in a flat structure to send it to the cloud in a way that's conducive to machine learning; and third, to ensure fast processing while using the AI during production, ensure the trained algorithm is transferred from the cloud and embedded into a use case that runs on the edge. When running directly from the cloud, you will experience too many latency issues which leads to an unwanted bottleneck problem.

So, how did we do it differently? Well, beyond the years of R&D and millions invested, we partnered closely with major strategic electronics manufacturers to merge their domain expertise with our platform to develop and test use cases where the AI is trained and programmed to solve a specific manufacturing headache.

Our most popular use cases today involve reducing false calls at automated optical inspection (AOI) and correlating solder paste inspection (SPI) data directly to specific screen printer variables. These enable manufacturers to prevent issues from happening instead of waiting for something to go wrong before reacting – put simply, going from reactive to prescriptive analytics.

Basically, everybody who's running an SMT line has nightmares with false calls – even with a $200k+ machine, an operator still must sit in front of the machine or at an offline station to filter out the false calls. This adds to labor cost while diminishing line productivity. With our platform, you don’t need to invest so much effort; we’ve helped manufacturers reduce false calls by about 60%.

With the AOI use case, we train an algorithm to review every defect flagged by the AOI and detect the false calls with a high degree of confidence. Displayed on a simple dashboard (pictured above), operators can see precisely why the AI flagged a false call, providing the necessary proof to SMT operators that the AI catches the majority of false calls and doesn’t make mistakes. Once trust is established and the AI system is sending automated messages directly to the AOI software, the operators can instead focus their limited time on more important improvements and leave the tedious, repetitive nature of flagging false calls to these intelligent systems.

Not just applicable to the AOI process, we are currently working with manufacturers to develop similar solutions for both automated x-ray inspection (AXI) and SPI machines. I am particularly excited about the development with AXI because first pass yield (FPY) at AXI tends to be quite low due to the inferior quality of x-ray images relative to the cameras used in AOI machines, so our AI platform will make significant improvements here for manufacturers using AXI machines.

As time goes by with the AI in-use on the factory floor, it gets better and better at identifying false calls and catching the subtlest defects that are often overlooked by the human eye. In fact, to improve the algorithm’s accuracy it can be retrained over time based on your factory environment and specific products. With this retraining, your AI use case becomes more effective at anticipating failures and even predicting analyses for future performance.

The critical nature of the solder printing process on the SMT line is well known by us all; in fact, over 60% of the PCB assembly soldering defects are attributed to this process. Any inadequate variable transferred at the printing operation to the PCBA leads to defects as well as substantial reworking and repair costs. This is why so many manufacturers have looked to closed-loop feedback between SMT machines – and most notably between the printer and SPI machines.

While not a new concept, electronics manufacturers have been striving to achieve this for years – made incredibly complicated by a mixed vendor ecosystem, most manufacturers have yet to experience the benefits of correlating and automatically exchanging data between machines on the SMT line.

Comparing results data from the SPI machine to what's happening inside the screen printer is one arena in SMT manufacturing ripe with opportunity. Unfortunately, it’s basically impossible for a human being to efficiently correlate all that data, and to correctly predict when test yields will drop due to multiple variables changing simultaneously in the screen printer.

That’s where our AI platform has helped a few key electronics manufacturers – by training an algorithm to process, analyze, and understand the relevant variables within the screen printer and how those impact the inspection results at the downstream SPI machine. Once trained, our platform can predict and prevent problems before they even happen. For example, if a combination of variables at the screen printer start to deviate from normal (even the subtlest of changes), our model could predict if and when these changes will lead to an actual problem and trigger an alarm.

The ubiquitous nature of AI in our personal and professional lives tells me it’s here to stay. Electronics manufacturers who are early adopters of this technology can truly distinguish themselves from the competition by exploring ways to make process optimizations, leverage cost savings, and determine novel insights. Intentionally implementing AI with clear use cases has countless benefits (with some solutions we humans have yet to imagine, though I have no doubt AI will discover new suggestions for us before long).

Like most worthwhile endeavors, implementing the right AI platform will not be easy and should be looked at as a long-term journey that clearly complements your overall factory digitalization strategy. With our team of data scientists and AI experts, we partner closely with our customers to ensure we’re rolling out an AI solution that gets integrated seamlessly into their specific manufacturing environment and evolves right alongside them.