Welding with Artificial Intelligence

If you can see an anomaly in the welding camera image, an AI system can be trained to detect it. The AI system can detect the anomaly consistently, and create immediate alarms allowing the team to repair the faulty weld before additional welds bury the problem or the component moves downstream for additional processing. If the problem is detected again, its time for the team to check and modify the process.

The AI system can also automatically collect records of all known anomalies, offering the welding engineer the opportunity to study the welding anomaly library and refine the detection algorithms for defects that are out of tolerance. In essence, the system will improve as you use it.

Our AI approach harnesses your engineering expertise to maintain the highest quality standards. Because AI far outperforms traditional machine vision for complex and varied images (e.g., a welding image) you can expect higher accuracy in defect classification over a diverse range of images. You can also expect a tool that can be used for related applications via retraining without starting from scratch each time. 

If the defect is visible in the camera image, our AI system can detect it. All we need in order to build the system is video and your project specification. 

When considering AI approaches to identifying geometrical features, it's important to consider the welding engineering aims. Which features need to be tracked? What range of weld geometries will the system need to work on? Our preferred approach is keypoint tracking. Compared to machine vision techniques, this approach is considerably less sensitive to image variation, and is smart enough to work on multiple weld joints.

The system can be set up to measure distances between any pair of points, and can be designed to be adaptive to tack and fit up variation in the welding process. While we haven't built a control system to use the detected geometry yet, it's only a matter of time.

• AI outperforms traditional machine vision for complex and varied images, e.g. welding images. 
• Because AI learns, it leads to higher accuracy of predictions than traditional machine vision approaches and lends itself to continuous improvement.
• A single AI model can be developed to work on multiple different images; for example in a multi-pass welding application, the same model can be used to detect features from the root to the cap.
• In the same vein, AI models can be adapted to related applications with moderate training; it is not necessary to start from scratch each time. 
• AI offers real potential to control key aspects of the welding process, e.g. wire position.
• Cameras offer potential for beyond human inspection e.g. fast process dynamics and UV, IR or Xray spectra.