Probabilistic prediction of optimal T-joint welding parameters considering data trends and variability using generative adversarial network

Abstract

This study proposes a method for deriving optimal welding conditions to achieve the target bead geometry in fillet welding. Existing regression-based approaches, fail to sufficiently capture the nonlinearity and variability inherent in the welding process. Moreover, the lack of available raw data hinders the training of reliable models. Prediction models trained solely on deterministic data also face limitations in learning distributional characteristics, making it difficult to identify optimal parameter combinations within a confidence interval. To overcome these challenges, we introduce a model correction and augmentation technique based on a generative adversarial network (GAN). This approach calibrates the linearity of finite element analysis simulation data while incorporating the experimental variability observed in welding data. This approach preserves the linearity of simulation data and the physical variability of experimental welding data, generating reliable corrected and augmented datasets for training. The generated dataset, expressed probabilistically, was applied to prediction models to estimate optimal welding conditions corresponding to the target bead geometry. Two deep learning algorithms—a Multi-layer Perceptron and a Transformer model—were employed for performance comparison. The results indicated that the Transformer achieved lower mean absolute error and higher R² values, demonstrating superior predictive performance. Additionally, heuristic filtering was used to derive multiple welding conditions that satisfied the target bead geometry. In conclusion, this study demonstrates that GAN-based data correction and augmentation provide the foundation for constructing a reliable probabilistic dataset, while the optimal welding conditions are subsequently derived by applying prediction algorithms in combination with heuristic filtering, enabling the selection of stable welding parameters that minimize defects and achieve the target bead geometry in fillet welding.