Eurotec: Looking back, how would you assess the impact of AI on your operations?
RB: For us, the key thing is that our clients can now carry out operations that were previously impossible. From their perspective, there is clearly a clear dividing line between the pre-AI and post-AI eras.
JP: The presentation of the Pathabene algorithm sparked many ideas among our clients, including one that was particularly interesting: using this technology to create a new form of visual recognition. A year of extremely complex work has enabled us to develop what we now consider to be the best visual recognition algorithm on the market.
Everyone thinks their product is the best… so what makes yours any different?
JP: Conventional algorithms have their limitations, particularly when it comes to surface conditions and consistency in part recognition. Feedback from our customers shows that we have succeeded in pushing these boundaries. For example, we have a watchmaking client whose parts to be welded are close to their final state, with surfaces that are mirror-like. The images captured by Mister-Laser’s optical system sometimes appear white or black, and it is very difficult for conventional visual recognition to identify the part without adjusting the lighting. Furthermore, the lighting setup has to be adjusted every time the room is changed, which is time-consuming. With one exception, we have so far been able to carry out all projects using standard, fixed lighting.
In practical terms, what improvements have been made?
RB: The comparison below shows different methods of visual recognition. The top two examples were created using Pathabene and Recobene, which utilise AI algorithms. Below are the improved versions of pattern matching and circle recognition from the latest version of Forbeam. In each case, we have simulated the laser beam to avoid any side effects that might arise from the geometry of the weld, so that we can focus solely on visual recognition. We have deliberately chosen light, dark, rough and smooth surfaces, and the pins feature different chamfers and bores with a wide variety of surface finishes.
It can be seen that Pathabene has correctly identified all the cases. The welded geometry is not perfectly circular because, as a reminder, Pathabene determines the welding path itself, which is why the visible joint plane has been used more as a centre point. As the AI adapts to different conditions, the lighting and image capture settings are basic and can remain the same from one series of parts to the next, even if the parts themselves are different. Recobene was also successful, but focused on the pin. The circular welding geometry was defined by the machine operator. The alignment must therefore be perfect for the weld. The working conditions are the same as for Pathabene and do not require any special attention during set-up. However, the recognition speed is 10 times faster than with Pathabene. Pattern matching quickly revealed its limitations when dealing with visual variations in the parts. It is clear here that we would need to adjust the lighting to make it either grazing or dome-shaped, though there is no guarantee this would resolve the issue. Any change in the production line would require adjustments, which would delay the machine’s commissioning. As the lighting is very close to the workpiece, care must also be taken during automation to prevent the two from colliding. Finally, the circle search function has identified circles, but with such a wide diameter tolerance that the centring is only approximate. The chamfers would need to be improved to enhance the search quality, but this would be much more expensive to produce.
How does it work?
JP: We tested commercially available solutions, such as YOLO V8, but its technical limitations soon became apparent. After numerous trials, we reverted to the ResNet model, which had been used successfully in Pathabene, but with a specialised final layer designed for the reliable and accurate detection of objects of different sizes.
The integration into Forbeam of annotation editing—a fundamental feature used in any learning process—was a key element of the project: it is intuitive and streamlined, making this task a real breeze. As can be seen in the following image, the number of parameters is extremely limited. The centring tool must be adjusted to the workpiece and sized so as to obtain the information required for training, whilst aligning the circle with the workpiece to be welded.
Increasing this base data is crucial: the user annotates between one and four images of their parts, taken using the Forbeam tool, and the training server transforms them into thousands of different images by varying the angle, brightness, size, noise, etc. The training focuses on the common essence shared by all these images, which makes the model particularly reliable and resilient.
Pathabene needs a training server. Does Recobene need one as well?
RB: Yes, but as customers may need both algorithms for their production work, we decided to stick with a single training server and added Recobene. Customers have two options: a local server at home or remote access to our server, which is based in Switzerland in a secure location. This development also took several months of work to create a single interface that supports two different types of learning.
Which watchmaking applications can be improved using this algorithm?
RB: All welds are inspected using machine vision. For example, the surface of certain watch bearings is so shiny that even slight angles on the part or very slight roughness can make a huge difference to the surface’s brightness, generating spurious information that can interfere with visual recognition:
Other components such as pullers, rings, oscillating weights, gear trains, dials, straps, clasps, cases, bearings, bridges, pins, barrels and tenons share the same characteristics. That is why we have introduced a new generation of optics and cameras that improve image resolution for ring fitting by a factor of four.
On the left, the image from the standard camera. In the centre, the image from the new high-definition camera. On the right, the same camera as in the centre but with the new image magnification lens. Thanks to these new features, it is possible to further improve the precision of welding on the edge of a coil and cause less damage to it.
Welding of a pinion to a clockwork schaft, carried out using visual inspection.
