PSI - Issue 75

6

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

125

(a)

(b)

Start-Stop-Position

Manual weldedfillet welds (GMAW) 200 cross sections

Manual repair welded fillet welds (FSAW) 200 cross sections

Cross section

spatters

PointNet++ Accurency 94.7 %

PointNet++ Accurency 90,9%

Manual labelling

Manual labelling

Fig. 3: Application of PointNet ++ for semantic segmentation of weld cross sections with irregularities (a) GMAW weld with start-stop-points, (b) Manuel FSAW weld with spatters

Total accurency BM accurency Weld toe accurency Weld accurency

0 10 20 30 40 50 60 70 80 90 100

Accurency [%]

Curvature

RANSAC

PoitNet++

Curvature

RANSAC

PoitNet++

Automated GMAW weld

Manual FSAW repair weld

Fig. 4: Determined accuracy of semantic segmentation of 2D weld profiles

constraints. The eigenvalue decomposition of these covariance matrices produces curvature metrics, defined as the ratio of the smallest eigenvalue to the sum of all eigenvalues, providing a quantitative measure of the local shape characteristics of the point cloud. By filtering areas with high curvature, such as those where r < 1 mm, point clusters can be identified, as illustrated in Figure 5. These clusters can then be separated using a clustering algorithm, such as the one described by [8]. The resultant connected regions can be further refined by excluding areas outside the weld line, for example, regions with small radii around spatter. This allows the extraction of feature lines, such as those corresponding to weld toe radii. By this approach, various clusters of regions with high curvature can be identified, for example, ripples or the small weld end crater visible in Figure 5. These clusters can be removed if information from the 3D RANSAC algorithm is included, Figure 6. In the majority of welds, two planar or bend surfaces (plates of pipes) are connected that can be identified by RANSAC. Subsequently, all clusters can be removed that have a certain distance of, e.g., d > 0 . 5mm from the surface. By this approach, imperfections in the weld, such as weld toes between cover beads, can be removed from the weld toe identification. Irregularities away from the weld toe line can be removed by looking at the number of points in each cluster. It is expected that the cluster containing the weld toe is the largest one. Only irregularities close to the weld cannot be distinguished. In the final step, a 3D interpolation function can be fitted to the identified nodes along the feature line. This spline enables the definition of normal surfaces from which 2D weld profiles can be extracted, providing a detailed representation of the weld geometry.