Self-enhanced optimization and transfer learning for accurate prediction and quality enhancement in friction stir welding

Friction Stir Welding (FSW) improves material properties by refining the grain structure, which leads to enhanced strength and resistance to defects. This study proposes a novel Transfer Learning-based prediction model for Friction Stir Welding (TL-FSW) performance. Two datasets are employed to assess the capability of the GoogleNet with Transfer Learning (GNTL) model for predicting friction stir welding (FSW) performance. In the case of Aluminium Alloy (AA) 7075-T651, the model estimates surface roughness using key process variables. For AA6061, the model predicts the ultimate tensile strength (UTS) by utilising parameters such as rotational speed, welding speed and axial force. All three parameters are optimised via a Self-Enhanced Motherhood Optimisation (SE-MO) strategy to improve the accuracy of the prediction. The proposed TL-FSW model achieves lower prediction errors, with MAE values of 0.1678 and 0.1574 and MAPE values of 0.014 and 0.012 at 60% and 90% training data, respectively.