Scholarlink

<p>Machine learning has been widely applied in the field of materials science and has achieved notable successes. However, due to limitations such as data quality, it is challenging for model prediction accuracy to reach 100%. This study focuses on predicting the quality of linear friction welding (LFW) joints using a hybrid machine learning model. A total of 149 LFW test sets were collected, including data from titanium alloys, high-temperature alloys, and rail steel alloys, with 129 sets specifically related to titanium alloys. Eight machine learning models—artificial neural networks, random forests, decision trees, support vector machines, AdaBoost, Naive Bayes, gradient boosting, and TabPFN—were trained and tested to compare their predictive accuracies. The results indicate that the overall predictive accuracy of these models for LFW joint quality across different materials is generally low. However, when focusing exclusively on titanium alloy LFW tests, the machine learning models of random forest and gradient boost achieved a maximum accuracy of 84.6%, with decision trees showing the lowest accuracy at 61.5%. By removing data identified as low quality through hybrid machine learning algorithms and to optimize the data quality, the prediction accuracy further improved, with the Gradient Boosting model achieving a perfect prediction accuracy of 100%. The linear friction welding of the Ti60-TC17 material pair using the welding parameters provided by the hybrid model resulted in a defect-free joint with good performance, indicating the model's effectiveness.</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

Prediction of Linear Friction Welding Joint Quality Based on Hybrid Machine Learning Models and Data Quality Optimization

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Zhang Shuxin,
  • Xie Faqin,
  • Wu Xiangqing,
  • Yan Xi,
  • Luo Jinheng

摘要

Machine learning has been widely applied in the field of materials science and has achieved notable successes. However, due to limitations such as data quality, it is challenging for model prediction accuracy to reach 100%. This study focuses on predicting the quality of linear friction welding (LFW) joints using a hybrid machine learning model. A total of 149 LFW test sets were collected, including data from titanium alloys, high-temperature alloys, and rail steel alloys, with 129 sets specifically related to titanium alloys. Eight machine learning models—artificial neural networks, random forests, decision trees, support vector machines, AdaBoost, Naive Bayes, gradient boosting, and TabPFN—were trained and tested to compare their predictive accuracies. The results indicate that the overall predictive accuracy of these models for LFW joint quality across different materials is generally low. However, when focusing exclusively on titanium alloy LFW tests, the machine learning models of random forest and gradient boost achieved a maximum accuracy of 84.6%, with decision trees showing the lowest accuracy at 61.5%. By removing data identified as low quality through hybrid machine learning algorithms and to optimize the data quality, the prediction accuracy further improved, with the Gradient Boosting model achieving a perfect prediction accuracy of 100%. The linear friction welding of the Ti60-TC17 material pair using the welding parameters provided by the hybrid model resulted in a defect-free joint with good performance, indicating the model's effectiveness.