PSI - Issue 75

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

124

5

extraction and abstraction on increasingly larger local regions. Afterward, global point cloud properties are captured by MLPs. The network structure is illustrated in Figure 2 (d). The implementation from [20] was used in this work.

(a)

(b)

Quadratic RANSAC

Linear RANSAC

d

d

Data points

Quadratic RANSAC fit

Intersection → Weld toe

Y

X

Linear RANSAC fit

Radius determination at weld cross sections

(c)

by fit

(d)

spline fit Points from scan Points in the area of interest

Y

X

Area of interest Max( (x))

Curvature

(x)

X

Fig. 2: (a) Illustration of application of RANSAC Algorithm on point clouds, (b) Determination of weld toe from weld cross section with RANDSAC (c) Determination of the weld toe by the curvature method, (d) Deep neural network structure of PointNet ++ [13]

3. Evaluation of 2D-weld profiles

As mentioned, determining the AoI (weld toe) is crucial for evaluating the geometrical parameters of the welded joints. For this reason, the di ff erent methods described here are applied to the segmentation of weld cross-sections, illustrated in Figure 3. Both fillet weld types contain irregularities resulting from start-stop positions or by weld spatters. Around 200 cross-section were manually annotated using the open source tool LabelCloud [17] from the Github repository [1]. Each cross section was divided into weld, weld toe and, base material, as shown in Figure 2. Segmentation accuracy is calculated by the ratio of correctly determined points to the number of total points, given as a percentage. The segmentation for all cross sections was carried out according to Section 2. For the curvature-based segmen tation, a fixed area of interest of 20 points was used (around 700 µ m distance). In contrast, for the RANSAC- and curvature-based method, a floating value for the width of the area of interest is used, based on the predicted weld toe angle α pred and the weld toe radius r pred = 1 κ , where κ is taken from the maximum kappa value (Max( κ )):

r pred · π · α pred resolution · 180

points on edge =

(3)

The results are shown in Figure 4. As depicted, PointNet ++ achieved the highest accuracy ( > 90 %). For relative simple weld geometries, however, (i.e., automated GMAW welds), the RANSAC-based approach reached similar accuracies. In contrast, for complex weld geometries such as manual welded FSAW welds with a high degree of variation, the RANSAC algorithm yielded significantly lower accuracies.

4. Evaluation of 3D-weld geometry

In a 3D model, curvature and radii can be e ff ectively evaluated by estimating covariance matrices for each point using a hybrid KD-tree search, which captures local geometric properties based on a specified radius and neighbour