PSI - Issue 75

Jörg Baumgartner et al. / Procedia Structural Integrity 75 (2025) 120–128 Jo¨rg Baumgartner / Structural Integrity Procedia 00 (2025) 000–000

127

8

The described approach to identify the weld toe line can also be used to simplify the annotation process of scanned welds for machine learning purposes, illustrated in Figure 7. At some areas, manual adjustments may be needed if irregularities are present close to the weld toe. Therefore, a semi-automatic process need to be applied. As shown in Figure 7, with a limited training data set of around 200 mm 3D-scan of a fillet welded joint with one layer, an agreement of 91 % was reached even if not all irregularities are removed in the labeled data. The authors assume that with a higher quantity of high-quality data, the prediction accuracy could be significantly increased.

5. Discussion

In this paper, various evaluation procedures employing deterministic algorithms and machine learning are de scribed. The evaluation algorithms for 2D profiles are well advanced and lead to reliable (but still improvable) results. However, the data quality and consistency are crucial for a successful application. In this work, a limited number of 2D profiles were annotated or labeled by hand using manual tools. The e ffi ciency of the process may be increased by automated algorithms, but in the end there is still the question regarding a universal definition of the AoI or weld toe. The evaluation algorithms for detecting welds in 3D point clouds are less advanced. Although the algorithm for identifying the weld toe lines by 3D curvature analysis, subsequent clustering and consideration of the base plate surface by RANSAC is fast (approx. 10 seconds for a 30 mm long weld) and leads to reliable results, the following steps still need improvement, especially if: • the weld and especially the weld toe geometry is subject to high variation. In this case, no continuous weld line can be derived. • the surfaces of the welded plates are bent, and the RANSAC algorithm does not detect a plane surface. In this case, a RANSAC-algorithm that detects bent resp. cylindrical surfaces need to be implemented.

6. Conclusion

This work focuses on segmentation and determination of 2D weld profiles and 3D scans of welded joints by rule based and machine learning algorithms. The following conclusions can be drawn:

• The deep neural network PointNet ++ shows high accuracy regarding the semantic segmentation of 2D-weld profiles, while the rule-based RANSAC and curvature-based algorithm show similar accuracy for simple cases (automated welded joints with a low geometrical variation). While the ground truth, the ’real’ weld toe, was manually annotated from 2D profiles, the training data may vary from user to user. An universal definition may be needed regarding the definition of the weld toe. • While curvature-based weld toe detection works well for smooth single-layer weld transitions without irregu larities, the accuracy of this method drops significantly if there are several features, such as spatter, which lead to high variation of curvature in 2D weld profiles. • A method is proposed that allows an evaluation of curvature for 3D point clouds and, with it, feature lines, such as the weld toe line. For high-quality welds, this leads to reliable results; however, additional work is still needed to capture the overall weld direction of complex geometries and irregular welds. • Accurate annotation of 3D point clouds of welds is essential to apply machine learning techniques to detect welds and weld toes. A semi-automatic approach could be an e ff ective solution, where the automatic application of a 3D curvature method is complemented by manual refinement to address any irregularities. This work shows di ff erent approaches for the determination of relevant features in point clouds for weld quality assessment. In terms of fatigue, this includes the weld toe or weld seam, where fatigue failure is typically expected (notch e ff ect). Thus, this work underscores the need for a standardized definition of the weld toe to form the basis of an automated quality assessment based on 3D scanning.