Scholarlink

<p>In this investigation, we present a two-step model integrating machine learning and statistical techniques to correlate friction stir processing (FSP) parameters with ultrasonic testing waveforms and grain size. The average grain size is determined through electron microscopy, and the mean absolute voltage of ultrasonic testing waveforms is obtained from non-destructive evaluation of steady-state processed regions. A random forest model is selected to predict mean absolute voltage of the ultrasonic data, and prediction intervals are constructed. The accuracy of the mean absolute voltage prediction model is validated against experimental data, aligning within standard deviations. The predictions are then aggregated and mapped to grain size predictions using a statistical model, incorporating a Bonferroni adjustment to manage uncertainty propagation. This work novelly presents an FSP case study at the intersection of machine learning prediction for ultrasonic non-destructive evaluation with extensions to microstructural feature prediction. This pipeline lays the groundwork for future development of control systems aimed at in-line quality prediction and anomaly detection during friction stir processing, underscoring its potential for enhancing manufacturing processes.</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

Machine Learning Statistical Ultrasonic Non-destructive Prediction for Friction Stir Processed Metal

  • 期刊论文

  • 发表日期：

    2026年1月28日

查看全文
  • Luke Durell,
  • Yanming Guo,
  • David Garcia,
  • Mayur Pole,
  • Tianhao Wang,
  • Hrishikesh Das,
  • Donald Todd,
  • Kenneth Ross,
  • Erin Barker,
  • Eric Smith,
  • Keerti Kappagantula

摘要

In this investigation, we present a two-step model integrating machine learning and statistical techniques to correlate friction stir processing (FSP) parameters with ultrasonic testing waveforms and grain size. The average grain size is determined through electron microscopy, and the mean absolute voltage of ultrasonic testing waveforms is obtained from non-destructive evaluation of steady-state processed regions. A random forest model is selected to predict mean absolute voltage of the ultrasonic data, and prediction intervals are constructed. The accuracy of the mean absolute voltage prediction model is validated against experimental data, aligning within standard deviations. The predictions are then aggregated and mapped to grain size predictions using a statistical model, incorporating a Bonferroni adjustment to manage uncertainty propagation. This work novelly presents an FSP case study at the intersection of machine learning prediction for ultrasonic non-destructive evaluation with extensions to microstructural feature prediction. This pipeline lays the groundwork for future development of control systems aimed at in-line quality prediction and anomaly detection during friction stir processing, underscoring its potential for enhancing manufacturing processes.