Scholarlink

<p>Friction stir welding (FSW) has been widely applied in aerospace and transportation owing to its solid-state nature. However, accurate prediction of the nugget zone temperature in medium-to-thick gauge plates remains difficult due to steep thermal gradients and limited sensor access. This study proposes a hybrid approach integrating finite element simulation and machine learning. A three-dimensional transient thermal model was developed in Abaqus using a combined surface-volume heat source to generate temperature data, which were experimentally validated via thermocouples on 12-mm-thick AA 6061-T6 aluminum alloy plates, showing a maximum relative error below 6.37%. Using these data, a stacking ensemble model was constructed, incorporating Back Propagation Neural Network, Support Vector Regression, Extreme Gradient Boosting, and Random Forest as base learners, with Light Gradient Boosting Machine as the meta-learner. Hyperparameters were optimized via grid search and fivefold cross-validation. Results demonstrate that the Stacking model outperforms all single models, achieving a mean absolute error of 4.0086&#xa0;°C, mean squared error of 46.96, and determination (<i>R</i><sup>2</sup>) of 0.9911. SHAP analysis indicates that rotational speed, welding speed, and surface temperature are the most critical features. The proposed method effectively integrates physical modeling with data-driven learning. Through this hybrid framework, precise nugget temperature prediction from surface data is achieved for the investigated 12-mm aluminum alloy. Consequently, it provides a technical basis for process monitoring within the defined parameter range.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Stacking Ensemble Learning Method for Nugget Zone Temperature Prediction in Friction Stir Welding

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Yixian Zhu,
  • Qingpo Xu,
  • Shaofei Meng,
  • Haitao Liu,
  • Xianlei Shan

摘要

Friction stir welding (FSW) has been widely applied in aerospace and transportation owing to its solid-state nature. However, accurate prediction of the nugget zone temperature in medium-to-thick gauge plates remains difficult due to steep thermal gradients and limited sensor access. This study proposes a hybrid approach integrating finite element simulation and machine learning. A three-dimensional transient thermal model was developed in Abaqus using a combined surface-volume heat source to generate temperature data, which were experimentally validated via thermocouples on 12-mm-thick AA 6061-T6 aluminum alloy plates, showing a maximum relative error below 6.37%. Using these data, a stacking ensemble model was constructed, incorporating Back Propagation Neural Network, Support Vector Regression, Extreme Gradient Boosting, and Random Forest as base learners, with Light Gradient Boosting Machine as the meta-learner. Hyperparameters were optimized via grid search and fivefold cross-validation. Results demonstrate that the Stacking model outperforms all single models, achieving a mean absolute error of 4.0086 °C, mean squared error of 46.96, and determination (R2) of 0.9911. SHAP analysis indicates that rotational speed, welding speed, and surface temperature are the most critical features. The proposed method effectively integrates physical modeling with data-driven learning. Through this hybrid framework, precise nugget temperature prediction from surface data is achieved for the investigated 12-mm aluminum alloy. Consequently, it provides a technical basis for process monitoring within the defined parameter range.