PSI - Issue 57

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

113

2

notch effect and depends on the local surface geometry and the load case (Zerbst et al. , 2018). Beside other influences, like coarse grain microstructure in the heat affected zone that benefits the crack propagation (Kucharczyk et al. , 2018) or harmful tensile residual stresses (Hensel et al. , 2018), the stress concentration at the weld toe is the main mechanism for crack initiation and responsible for the comparable low fatigue strength of welded joints (Schork et al. , 2017). For this reason, there is the need for a fast and reliable quantification of the SCF for weld quality control. Usually, the weld toe radius and flank angle and plate thickness are used to calculate the SCF of a weld toe, as those parameters have shown to be most influential (Kiyak, Madia and Zerbst, 2016; Schork et al. , 2017, 2020). Multiple SCF solutions for different types of welded joints are available, for example the approximation formulas by Kiyak et al. (Kiyak, Madia and Zerbst, 2016), Anthes et al. (Anthes, Köttgen and Seeger, 1993), Rainer (Rainer, 1978) or Lawrence (Lawrence, Ho and Mazumdar, 1981). Newer methods for the assessment of SCFs according to Oswald et al. (Oswald, Mayr and Rother, 2019; Oswald, Neuhäusler and Rother, 2020) and Dabiri et al. (Dabiri et al. , 2017) are based on artificial neuronal networks (ANN) for the assessment of SCF. However, it should be mentioned that comparisons between SCF evaluated using analytical solutions and SCF evaluated directly using the finite element (FE) method by Ottersböck et al. (Ottersböck, Leitner and Stoschka, 2021) and Schubnell et al. (Schubnell et al. , 2020) show much higher variation in the results of the FE calculations than the analytical solutions predicted. These FE based evaluations are only possible with a high effort, which is not practical for most applications. As mentioned, nearly all current SCF solutions are based on geometrical parameters ( , , ). The usage of 3D scanning technique allows the evaluation of the geometrical parameters with a comparable low effort (Schubnell et al. , 2020; Renken et al. , 2021). However, previous investigations (Schubnell et al. , 2020) show high uncertainties and scatter regarding the evaluated parameter with different measurement methods or different measurement systems (scale and resolution dependency). Furthermore, the geometrical parameters cannot always be defined properly according to the real geometry of welded joints, especially for the weld toe radius a measuring instruction is missing (Schork et al. , 2017). Also, it should be mentioned that the geometrical parameters (weld toe radius and flank angle) are in general approximations of the real weld geometry. The real weld shape is more complex. For these reasons the determination of the SCF based on geometrical parameters will always have a certain deviation from the real SCF at the weld toe. The aim of this work is to develop an approach for the direct determination of the SCF of welded joints based on the surface scans (real weld toe geometry), illustrated in Figure 1 (a). To build up this complex relation artificial neural networks are used. In this first study input data for the ANN is created by using artificial 2D-profiles (or 2D-scans) that were generated according to a given range of geometrical parameters of welded joints, illustrated in Figure 1 (b). In this approach the 2D-profiles (X- and Y-coordinates) were used as input for an ANN to determine the SCF. The determination of input (or training) data for the ANN is described in section 2; the network architecture of the used ANN, adjustments for regression, postprocessing, training are summarized in section 3; the results of the SCF assessment by ANN compared to previous methods is given in section 4; followed a short conclusion and outlook in section 6. Nomenclature weld toe radius ANN Artificial neural network flank angle MLP Multi-Layer Perceptrons plate thickness SA Set abstraction weld width CNN Convolutional neuronal network sum of square residual (R2-loss function) total sum of squares