Investigation of the influence of different welding methods on the prediction accuracy of machine learning methods for fatigue strength prediction of butt joints based on geometry scans

Abstract

Butt welds produced using different welding methods differ in their weld surface geometry. Due to that, using a combined feature space with varying weld methods does not seem appropriate for good prediction results, as indicated in the previous studies [5]. So, in this thesis, the transferability of pre-trained ML models on a new LH target subset with slightly varying weld qualities is investigated to check whether there is a sufficient transfer between the trained model and the subset. Therefore, the XGBoost ML model was trained on datasets consisting of specimens of only one welding method at a time and a pairwise combination afterwards. So, a total of six ML models, each with 30 different model configurations, were available after the training. These models were used to predict the three LH target subsets constructed based on steel grade and stress ratio R. Error measures were used to evaluate the prediction accuracies of different models for the three LH target subsets. Model interpretation using SHAP values was performed to assess the feature contributions and the influence of the data used for training the model on the predictions. This provides insights into whether geometrical and load-related features are adequately taken into account for the prediction. Using this, the thesis aimed to answer the question of whether the transferability is sufficient between the FCAW and LH target subset, the SAW and LH target subset and their combinations in comparison with the LH and LH target subset because a good transferability between the LH trained model and the LH target subset is expected.