Classification of Welds Using Neural Networks trained on Syntetic Data
This paper proposes a U-net-based neural network trained on synthetic weld data generated via an Ornstein–Uhlenbeck process and ISO 5817 parameters to effectively classify weld regions and imperfections, demonstrating that such models can successfully generalize to real-world inspection scenarios without the need for labor-intensive ground truth labeling.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every time a metal structure is built to withstand the relentless push and pull of wind, waves, or traffic, the joints holding it together are the most critical weak points. These joints, known as welds, are where two pieces of metal are fused, but the process is rarely perfect. Tiny variations in the shape of the weld—how rounded the edge is, how steep the slope, or if there are small dips in the metal—can concentrate stress and eventually lead to cracks. For decades, engineers have relied on human inspectors to look at these welds, often using simple rulers and gauges to measure them. While this method works, it is slow, prone to human error, and lacks the precision needed to catch the smallest, most dangerous flaws. As industries demand stronger, lighter, and safer structures, the need for a faster, more accurate way to inspect these connections has become urgent.
The solution lies in digital visual testing, a modern approach that uses lasers to scan welds and create detailed three-dimensional maps of their surfaces. Instead of a human eye guessing where a measurement should be taken, a computer can see the entire shape at once. However, this flood of data brings a new problem: teaching a computer to understand what it is looking at. A human inspector can easily spot the difference between the smooth base metal, the raised bead of the weld, and the transition zone where they meet. A computer, seeing only a cloud of points, cannot make that distinction without being taught. This is where artificial intelligence, specifically a type of computer program called a neural network, comes in. These networks are excellent at recognizing patterns in images, but they usually require thousands of examples to learn, each one carefully labeled by a human expert. For welds, finding enough real-world examples with perfect labels is difficult, expensive, and time-consuming.
To solve this bottleneck, a team of researchers from the Royal Institute of Technology in Sweden and a partner company developed a clever workaround. Instead of waiting for thousands of real welds to be scanned and labeled, they built a virtual factory to create them. They wrote a computer program that generates thousands of synthetic weld scans, complete with perfect, mathematically precise labels. In this digital world, the researchers could control every variable, from the size of the weld to the amount of noise in the measurement, ensuring that the computer knew exactly what each part of the image represented. They then used these synthetic images to train a neural network, teaching it to identify the base metal, the weld bead, and the transition zones. The central question was not whether the computer could learn from these fake images, but whether it could then apply that knowledge to real, messy welds found in the actual world.
The researchers trained their system on a dataset of one thousand synthetic scans, representing both straight butt joints and corner fillet joints. To make the training data realistic, they did not just draw perfect lines; they introduced natural variations. They simulated the way a real weld might wobble slightly along its length or how a laser scanner might introduce tiny measurement errors. They used a mathematical process to ensure that the changes in the weld's shape from one point to the next felt organic, much like how a real welder's hand might drift slightly over a long seam. Once the network was trained, the team tested it on scans of real welds produced by two different laser scanners. The results were promising. The computer, having never seen a real weld during its training, successfully identified the different regions of the welds in the new scans. It could distinguish the raised bead from the flat base metal and locate the transition zones with a high degree of accuracy.
The study suggests that this approach is a viable path forward for automating weld inspection. By using synthetic data, the researchers bypassed the laborious task of manually labeling thousands of real scans, which often requires expert judgment and can introduce subjectivity. The network learned the underlying geometric rules of what a weld looks like, rather than just memorizing specific images. While the system performed slightly better on scans with a lower resolution that matched its training data, it still demonstrated a strong ability to generalize to higher-resolution scans of real-world objects. The researchers note that the synthetic data is an approximation and that real-world factors like rust, spatter, or unusual lighting were not included in the training. However, the fact that the network could still find the weld regions in complex, real-world scans indicates it learned the essential shapes rather than just the artificial details of the simulation.
This work does not claim to have solved every problem in weld inspection, nor does it suggest that synthetic data is a perfect replacement for all real-world testing. Instead, it offers a powerful tool to jumpstart the development of automated systems. The ability to quickly generate large, perfectly labeled datasets means that engineers can now train sophisticated computer vision models much faster than before. These models can serve as the first step in an automated pipeline, narrowing down the area of interest so that other algorithms can measure specific dimensions with greater precision and repeatability. By removing the human element from the initial identification of the weld, the process becomes more consistent, reducing the variability that has long plagued manual measurements. The study concludes that while the gap between synthetic and real data remains, the bridge built by these neural networks is strong enough to carry the weight of future advancements in structural safety.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.