Determination of stress concentration factors of welded joints from 3D-surface scans by artificial neural networks

The surface geometry and the stress concentration of welded joints show a large variation and are individual for each joint at each position. This is one reason for the conservative fatigue assessment of welded joints. In the past the determination of stress concentration factors (SCF) by Finite Element (FE) simulations was based on the approximated surface geometry of the weld, defined by weld toe radius and flank angle or other geometrical parameters. In this work a new approach is presented to directly determine SCFs of welded joints based on the 2D-profile (coordinates) of the weld surfaces. For this, two convolutional neural networks (CNN) PointNet++ and 2DLaserNet for point cloud classification are modified to perform regression on 2D-profiles. As input parameter artificial 2D-profiles were generated. The artificial neural networks (ANN) were trained by using the SCFs determined by Finite Element (FE)-simulations based on the virtual 2D-profiles. Both ANN show a high performance (R2 score) for the determination of SCFs. Comparison of the proposed method with three analytical solutions shows in two cases a higher agreement and in one case a similar agreement.