Artificial intelligence & SMEs: Our path to intelligent assembly testing
Is artificial intelligence only suitable for highly standardised processes in the industrial sector? Not for SMEs? Far from it!
A black cube in front of the selective soldering wave - and now?
Bluish light illuminates the conveyor belt on which a THT assembly is slowly travelling. The assembler at the next table is already working on the next assembly. On its way to the selective soldering wave, the assembly passes through the black box before the parts are soldered. Just the daily routine of an EMS service provider - but what is the cube doing there?
„It's quite an innovation,“ says Andreas op den Winkel, one of the two project managers at A+B Electronic.
There are 2 cameras in the box that take 2 images of each assembly passing through, 4 images per assembly. With around 600 assemblies a day on this system, 5 days a week, that makes around 12,000 images a week. LED strips provide optimum lighting inside, while the black walls shield the light from outside.
Just a nice technical gimmick? Hardly.
„We're talking about an automatic and, above all, self-learning inspection of THT assemblies.“
Artificial intelligence, in other words. Self-learning neural networks. Backpropagation. Industrial image processing. All buzzwords that are not directly associated with EMS service providers. Industry 4.0 and Smart Factory with AI are no longer dreams of the future. Instead, the topic of AI is gaining momentum in the industry.
SNAP GmbH – Moving humans with technology
SNAP GmbH is a company that has set itself the goal of developing brain-computer interface systems (BCI systems for short). They specialise in working with AI technology and have joined forces with us to tackle the intelligent assembly verification project.
Artificial intelligence, neural networks, backpropagation .... and EMS?
In spring 2021, Andreas op den Winkel and another colleague from the test equipment construction department at A+B Electronic are asked whether they would like to take on a project in the field of intelligent assembly testing. Development of the hardware, construction of the photo box, project management and communication with the project partner, SNAP GmbH. So a project plan is drawn up and a concept written.
Many questions arise at the beginning. Where will the system be positioned in the process? Is the quality of the images sufficient? How can the lighting be optimised? Dr Corinna Weber is CEO of SNAP and remembers: “The most important thing for us was to start taking images and digitising the data. We took care of programming the artificial intelligence and mapped the interface between hardware and software. Ultimately, the aim is to analyse image data as intelligently and automatically as possible.”
The test phase: familiarisation with the algorithm
“We are currently in the test phase. The system is learning what a circuit board is and will then identify and classify components in the next step,” says op den Winkel. As with a new - and admittedly perhaps somewhat unfamiliar - employee, the first step is familiarisation: What is a capacitor? What is an IC? Where does which component belong?
Now it gets technical. The automated learning process is lengthy and not easy. Using large amounts of data, such as images of components, an artificial neural network learns to recognise patterns in the data. And also to recognise deviations. The training runs via what is known as backpropagation, or error feedback. In simple terms, various scenarios are run through in the system in order to detect possible errors and effects and automatically make adjustments.
Challenges of AI image recognition
“For the project, we are working with classic methods of artificial intelligence and building a system that can optimise itself,” explains Dr Weber. Of course, new technologies always come with challenges. “In the field of AI, the biggest challenge is probably the heterogeneity of the data. The data comes from different sources and is processed differently. We have to find a solution here so that employees don't have any additional work, for example when changing projects on the assembly line. The system must automatically pull the information it needs to check the PCB.” The motto is: the more effort you put into the set-up and test phase, the better the automation works in the ongoing process.
The problems with the hardware
In terms of hardware, one problem lies in the component heights. Op den Winkel summarises: “Interferences such as light must be excluded so that the images are as consistent in quality as possible. And the higher a component is, the higher the shaft through which the part enters the box has to be. This also allows more light to enter from outside.”
Once the system recognises all the components, the next step is to optimise the AI processes. “Then we have to integrate the circuit board data. This specifies the position of the components on the board. We can then define so-called ROIs, regions of interest, for image processing. This means that the entire image does not have to be viewed pixel by pixel, but only a predefined area. This makes the evaluation and learning process more efficient.”
How EMS and artificial intelligence come together at A+B Electronic
Of course, there are already AOI processes in the SMD and THT areas. “But we have to programme them ourselves and adjust them again and again,” says op den Winkel. “We have to adapt the programme every time too many pseudo errors occur, for example. By then, however, our colleagues in the revision department already have many assemblies to check, which in fact contain no errors. This allows the system to adapt itself in real time. We have much less to do with it.”
“Or especially in the THT area. It's enough to twist a capacitor on the module. It then has to be desoldered manually, turned round and soldered back in. That takes time. But if the AI recognises the error beforehand, the assembly line stops, the colleagues remove the assembly, correct the error and that's it. During revision, the time-consuming quality checks can be reduced to random samples while maintaining the same quality.”
Artificial intelligence in industrial SMEs
Production processes are becoming more reliable and efficient through the use of AI. Even in medium-sized companies. “We are actually surprised that AI is not yet used so frequently in production. It can take some of the burden off employees, because AI thinks for itself,” says Dr Weber.
Quality control in particular can benefit from this, as the large amounts of data constantly provide new insights and information about the processes. The topic of AI in production can also be taken further. “It could be used in the PCB assembly process itself - or in device assembly,” says Andreas op den Winkel. “Or networking the machines with each other to streamline the entire production process,” considers Dr Corinna Weber.
While the project team ponders further possibilities of artificial intelligence in our manufacturing processes, the pick-and-place operator continues to work with concentration. Another circuit board passes through the photo box. 4 more pictures are taken. Perhaps not spectacular from the outside, but a huge step towards the smart factory on the inside.
Would you like to gain an insight into our production site in Huntlosen and see the intelligent box for yourself? Then find out more about us here - or contact us directly with no obligation. Then we can talk about future projects and how we can grow together.
