PSI - Issue 57

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

115

4

In total 2540 different combinations of radius, angle and weld width were realized as point cloud with a point distance of 0.25 mm in x-direction. Additionally, equivalent FE models were set up for the FE-solver ABAQUS 2019, which were used to calculate the SCF at the weld toe. The evaluation point was set to the location of the maximum evaluated principal stress in the weld radius. A symmetry in the weld centerline was assumed to reduce the size of the model, loading was applied as bending load on the far end of the model. The material was modelled purely elastic with a Young’s modulus of = 210 GPa and \ = 0.3 . Quadratic 8-node plane stress elements with an element size of 0.025 mm in the vicinity of the weld toe were used.The histogram in Figure FIXME gives an overview over the calculated SCF. The SCF range from 1 to 3.5 with most of the results being in the range from 1.3 to 2.5. The overview of the generated SCF is shown in Figure 2 (c). The histogram in Figure 2 (b) illustrates the distribution of the SCF. As shown the majority of the SCF range a value of 1.3 and 2.5 with a maximum around a value of 3.5. 3. Artificial neural networks for 2D-profile regression 3.1. Architecture of the ANNs Comparable less ANN architectures are available in literature that are designed to analyze 3D point cloud data. The PointNet architecture (Charles R. Qi et al. , 2017), see Figure 3 (a), was the first deep learning architecture that directly processes unordered point clouds. It was designed for classification tasks (Kaleci, Turgut and Dutagaci, 2022). In this ANN, points are passed through network layers independent from each other. High dimensional features are extracted only through weight-shared Multi-Layer Perceptrons (MLPs) from lower-dimensional feature channels of each point separately. To reduce the effect of variability due to rotation, PointNet architecture also includes the T-network trained to predict transformations that align the input coordinates and intermediate features into a canonical pose in spatial and feature space.

(a)

(b)

(c)

Fig. 3. (a) PointNet architecture, (b) PointNet++ architecture, (c) 2DLaserNet architecture (Kaleci, Turgut and Dutagaci, 2022)

A further development of the PointNet architecture is the PointNet++ architecture (Charles Ruizhongtai Qi et al. , 2017) utilizes neighboring points while extracting local features. The PointNet++ architecture for classification employs three successive set abstraction (SA) layers, as shown in Figure 3 (b). This approach of applying feature extraction and abstraction to increasingly wider local regions mimics conventional convolutional neuronal networks (CNN), which operate on receptive fields of varying sizes (Kaleci, Turgut and Dutagaci, 2022). After the SA layers, global properties of the point cloud are modeled through MLPs.