Transfer learning and augmented data-driven parameter prediction for robotic welding

Robotic welding envisioned for the future of factories will promote high-demanding and customised tasks with overall higher productivity and quality. Within the context, robotic welding parameter prediction is essential for maintaining high standards of quality, efficiency, safety, and cost-effectiveness in smart manufacturing. However, data acquisition of welding process parameters is limited by process libraries and small sample sizes, given complex welding working environments, and it also requires extensive and costly experimentation. To address these issues, this study proposes a transfer learning and augmented data-driven approach for high-accuracy prediction of robotic welding parameters. Firstly, a data space transfer method is developed to construct a domain adaptation mapping matrix, focusing on small sample welding process parameters, and a data augmentation method is adopted to transfer welding process parameters with augmented sample data. Then, a DST-Multi-XGBoost model is developed to establish a mapping relationship between welding task features and welding process parameters. The constructed model can consider the relationship between the output, which reduces the complexity of the model and the number of parameters. Even with a small initial sample size, the model can use augmented data to understand complex coupling relationships and accurately predict welding process parameters. Finally, the effectiveness of the developed approach has been experimentally validated by a case study of robotic welding.