Scholarlink

<p>To increase the reliability of welded structures, it is essential to accurately predict the tensile strength of friction stir welded joints. Traditional artificial neural networks (ANNs) have difficulties in effective capturing the complicated nonlinear interactions between welding parameters, resulting in limited prediction accuracy. This limitation with ANNs may lead to unreliable strength estimations in critical applications. In order to address these challenges, a hybrid model was proposed that integrates dense layers and Gated Recurrent Units (GRU). This model benefits from the GRU's ability to detect dependencies and sequential patterns in the input welding parameters, including shoulder diameter (SD), tilt angle (TA), welding speed (WS), and tool rotational speed (TRS). This hybrid approach was assessed in comparison to a traditional ANN model using a dataset consisting of parameters related to welding and tensile strength measurements. The proposed model was found to have a higher coefficient of determination (<i>R</i><sup>2</sup> = 0.9922) and a lower mean squared error (MSE) than the ANN (<i>R</i><sup>2</sup> = 0.9643), suggesting a stronger correlation between experimental and predicted results. The analysis of loss convergence revealed a smaller gap between training and validation losses (2.58 versus 12.65 for ANN), indicating enhanced generalization, as well as reduced overfitting. The model demonstrated improved consistency, regardless of outliers, and maintained accuracy even on previously unreported data, demonstrating its suitability for practical welding conditions.</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

Development of a Hybrid Deep Learning Approach for Predictive Modeling of Tensile Strength in Friction Stir Welding

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • Renangi Sandeep,
  • Dara Sudhakara,
  • M. Janardhana Raju

摘要

To increase the reliability of welded structures, it is essential to accurately predict the tensile strength of friction stir welded joints. Traditional artificial neural networks (ANNs) have difficulties in effective capturing the complicated nonlinear interactions between welding parameters, resulting in limited prediction accuracy. This limitation with ANNs may lead to unreliable strength estimations in critical applications. In order to address these challenges, a hybrid model was proposed that integrates dense layers and Gated Recurrent Units (GRU). This model benefits from the GRU's ability to detect dependencies and sequential patterns in the input welding parameters, including shoulder diameter (SD), tilt angle (TA), welding speed (WS), and tool rotational speed (TRS). This hybrid approach was assessed in comparison to a traditional ANN model using a dataset consisting of parameters related to welding and tensile strength measurements. The proposed model was found to have a higher coefficient of determination (R2 = 0.9922) and a lower mean squared error (MSE) than the ANN (R2 = 0.9643), suggesting a stronger correlation between experimental and predicted results. The analysis of loss convergence revealed a smaller gap between training and validation losses (2.58 versus 12.65 for ANN), indicating enhanced generalization, as well as reduced overfitting. The model demonstrated improved consistency, regardless of outliers, and maintained accuracy even on previously unreported data, demonstrating its suitability for practical welding conditions.