Robust machine learning modeling for multi-parameter prediction in friction stir welding of naval brass: a case study towards industry 4.0

This study establishes a comprehensive machine learning (ML) framework for accurately predicting essential friction stir welding (FSW) attributes in naval brass, focusing on weld temperature, strength, and hardness. To address insufficient experimental data, a customized variational autoencoder (VAE) was used to generate high-fidelity augmented datasets. The fidelity of the augmented datasets was verified through distributional similarity analysis and feature-correlation divergence measures, ensuring the synthetic samples strictly adhered to the underlying physical-statistical signatures of the experimental data. A comparative assessment of four ML algorithms, support vector regression (SVR), AdaBoost, XGBoost, and decision tree (DT), demonstrated that SVR consistently attained greater predictive accuracy. The SVR model produced mean absolute percentage errors (MAPEs) of 1.0% for weld temperature, 3.2% for weld strength, and 1.6% for weld hardness. The model's generalization was further validated with an independent experimental dataset, affirming its reliability for practical industrial applications. The results demonstrate that ML models can effectively elucidate intricate process-property correlations in friction stir welding of naval brass, reducing reliance on experiments and enabling data-driven optimization. The suggested framework facilitates intelligent process monitoring, quality control, and the incorporation of Industry 4.0 methods in advanced welding applications.