Digital twin for weld pool evolution by data-physics integrated driving

Abstract

Perception of the welding process is a widely recognized challenge in intelligent manufacturing. The critical issue involved is accurately detecting the evolution and state of the weld pool, with the weld backside width being a crucial parameter. A prevalent approach involves training deep learning algorithms with welding process data to uncover latent patterns for predicting the backside width of the weld pool. However, this data-driven method heavily relies on welding process data and label data, which are prone to systematic and cumulative errors during acquisition and fabrication processes. Consequently, traditional regression methods restrict both accuracy and generalization capabilities of these models. To address this limitation, this study proposed a loss function for a regression model estimating weld backside width based on maximum likelihood estimation principles, and prior functions and transfer learning strategies were employed to enhance the prediction accuracy of the regression model. In addition, the weld surface width, arc voltage, and welding current were combined with a simplified heat source model to effectively visualize the cross-section of the weld pool. A digital twin system was developed to record, analyze, and visually characterize weld pool evolution.