Resistance Spot Welding (RSW) is a popular technique for joining sheet metals. Due to the involvement of multiple process parameters, ensuring continuous quality assessment is crucial. While machine learning (ML) methods have demonstrated effectiveness in monitoring welding quality, their application is often limited by the high cost and time required to collect sufficient training data. A promising solution is the integration of prior knowledge into the ML process through transfer learning (TL), enabling the development of more generalized models. This study proposes a TL-based methodology for RSW quality monitoring in the case of limited datasets. An experimental campaign was conducted to generate a target domain dataset comprising welding points produced under varying process conditions. A neural network was trained to predict the nugget diameter, which is typically more difficult and time-consuming to measure. Subsequently, TL techniques were employed to transfer knowledge from a model trained to predict the peak load of tensile shear tests to the nugget diameter prediction model. The source domain dataset used for TL included samples obtained under diverse experimental conditions, encompassing different materials, welding parameters, and electrode types. The results demonstrate that TL enhances model generalization and predictive performance across the full range of nugget diameters, including challenging cases with extremely small or large values that typically hinder accurate prediction. Accordingly, the model performance improved by 25%, achieving a mean absolute percentage error of 5.76%. These findings confirm the potential of TL to improve model robustness, particularly when applied to novel experimental setups. The study provides both theoretical and practical contributions, illustrating how laboratory-generated source domain datasets can be effectively leveraged to support quality monitoring in different production setups.
Antal et al. (Mon,) studied this question.