Utilizing artificial neural networks for stress concentration factor calculation in butt welds

In the current paper, the stress concentration factor in butt welded joints experiencing axial tension and bending loads is analyzed by means of neural network-based models. The configurations are considered in single-V and double-V forms, which is a differentiation that has received insufficient consideration in calculation of stress concentration factors of transverse butt welds using parametric equations. Differentiation of the weld form is of considerable significance as the symmetrical and non-symmetrical shapes of this weld type influence the maximum principal stress value at the critical spot, which makes utilization of one similar equation for both forms inappropriate. Design of experiments by Taguchi method is used to generate the required profiles with variable local weld parameters. The analysis is also extended to consider joints with axial misalignment, and numerical models are implemented to train the corresponding neural network. This network explicitly taking the misalignment into consideration was able to estimate the stress concentration factors with a higher degree of precision than common solutions using reference stress concentration factors (for different load cases) and magnification factors. All the neural network-based models in this study yielded more accurate results than currently available parametric equations for stress concentration calculation of transverse butt welds and, furthermore, the neural network-based models were able to provide accurate results for a broader range of local weld parameters.