PSI - Issue 57

Jan Schubnell et al. / Procedia Structural Integrity 57 (2024) 112–120 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

118

7

scatter and have the best R2 scores with 0.965 and 0.964 compared to the SCF solution of Rainer (Rainer, 1978) (R2 score 0.772). The results from PointNet++ also shows a high performance (R2 score 0.947).

Fig. 5. Scatter plots for comparison of the stress concentration factor taken from FEM and the stress concentration factor determined by different SCF solution of Anthes et al. (Anthes, Köttgen and Seeger, 1993) (a), Rainer (Rainer, 1978) (b), Kiyak (Kiyak, Madia and Zerbst, 2016) (c) the proposed method (d) (PointNet++ trained by virtual 2D-profiles 4. Conclusion An approach is proposed in this work to determine SCFs of welded joints directly from 2D-profiles by ANNs. Compared to former work (Oswald, Mayr and Rother, 2019; Oswald, Neuhäusler and Rother, 2020) the X- and Y coordinates from this profiles were used as input for the ANN. For this, two CNN architectures for the point cloud classification PointNet++ and 2DLaserNet were adjusted for the point cloud regression. For the training the R2-loss function were used. FE-simulations were used to create the SCF input data. For this, artificial 2D-profiles were created in a determined range of the geometrical parameters (weld toe radius , flank angle and weld width for a fixed plate thickness of =10 mm) of a representative double-V butt joint. The performance of the ANN were determined according to the R2 score. Also, the results of the ANN regarding the determined SCF were compared with three analytical SCF-solutions. Following conclusion can be made: • Both modified ANNs PointNet++ and 2DLaserNet show similar performance with a high R2 score (>0.94) and can be used for 2D-point cloud regression in general. • Both ANN shows a better agreement between the determined compared to the from the FE-analysis than the analytical solutions of (Rainer, 1978; Anthes, Köttgen and Seeger, 1993) and a similar agreement according to the SCF solution of (Kiyak, Madia and Zerbst, 2016). Thus, the proposed approach seems to be reasonable for the determination of the SCFs of welded joints. However, it is expected that usage of ANN for point cloud regression for the determination of SCF at real 2D or 3D-scans have an even higher potential because no further approximation of the weld geometry by geometrical parameters are needed, see Figure 1 (a). For this reason, further work should be performed based on real surface scans.