Scholarlink

<p>This study presents a systematic comparison of six machine learning algorithms for surface roughness (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\varvec{Ra}\)</EquationSource> </InlineEquation>) prediction in dry turning of Ti-6Al-4V across two nose radii (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\varvec{r}_{\varvec{\varepsilon }} \varvec{= 0.4}\)</EquationSource> </InlineEquation>&#xa0;mm and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\varvec{r}_{\varvec{\varepsilon }} \varvec{= 0.8}\)</EquationSource> </InlineEquation>&#xa0;mm) using a central composite design with 38 experiments. The algorithms evaluated are Support Vector Regression, Random Forest, XGBoost, Gaussian Process Regression, Extreme Learning Machine, and a Lichtenberg-Optimised ELM in which the Lichtenberg Algorithm replaces random weight initialisation with a structured fractal search. All models are trained on <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\varvec{\log (Ra)}\)</EquationSource> </InlineEquation> and validated by leave-one-out cross-validation under two feature configurations: process parameters only and process parameters augmented with cutting force measurements. GPR achieves the highest accuracy (<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\varvec{R}^{\varvec{2}} \varvec{= 0.922}\)</EquationSource> </InlineEquation>); under the present conditions, process parameters alone prove sufficient — force signals degrade four of six models due to collinearity with feed rate, confirmed by variance inflation factors exceeding 10 for <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\varvec{\bar{F}}_{\varvec{f}}\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(\varvec{\bar{F}}_{\varvec{p}}\)</EquationSource> </InlineEquation>. A polynomial regression baseline reaches <InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(\varvec{R}^{\varvec{2}} \varvec{= 0.920}\)</EquationSource> </InlineEquation>, matching GPR; ML advantages lie in uncertainty quantification, localised nonlinearity capture, and robustness to broader conditions. LA-ELM records the largest force-induced gain (<InlineEquation ID="IEq9"> <EquationSource Format="TEX">\(\varvec{\Delta R}^{\varvec{2}} \varvec{= +0.074}\)</EquationSource> </InlineEquation>), reaching <InlineEquation ID="IEq10"> <EquationSource Format="TEX">\(\varvec{R}^{\varvec{2}} \varvec{= 0.895}\)</EquationSource> </InlineEquation> and surpassing Random Forest and XGBoost. Results confirm that the Lichtenberg Algorithm is an effective weight optimisation strategy for ELM in small-dataset manufacturing regression.</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

Surface roughness prediction in dry turning of Ti-6Al-4V: benchmarking six machine learning algorithms under cutting force integration with a Lichtenberg-optimised ELM

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月15日

查看全文
  • Francielle da Silva Leandro,
  • Alex Fernandes de Souza,
  • Paulo Henrique da Silva Campos,
  • Filipe Alves Neto Verri,
  • Pedro Paulo Balestrassi

摘要

This study presents a systematic comparison of six machine learning algorithms for surface roughness ( \(\varvec{Ra}\) ) prediction in dry turning of Ti-6Al-4V across two nose radii ( \(\varvec{r}_{\varvec{\varepsilon }} \varvec{= 0.4}\)  mm and \(\varvec{r}_{\varvec{\varepsilon }} \varvec{= 0.8}\)  mm) using a central composite design with 38 experiments. The algorithms evaluated are Support Vector Regression, Random Forest, XGBoost, Gaussian Process Regression, Extreme Learning Machine, and a Lichtenberg-Optimised ELM in which the Lichtenberg Algorithm replaces random weight initialisation with a structured fractal search. All models are trained on \(\varvec{\log (Ra)}\) and validated by leave-one-out cross-validation under two feature configurations: process parameters only and process parameters augmented with cutting force measurements. GPR achieves the highest accuracy ( \(\varvec{R}^{\varvec{2}} \varvec{= 0.922}\) ); under the present conditions, process parameters alone prove sufficient — force signals degrade four of six models due to collinearity with feed rate, confirmed by variance inflation factors exceeding 10 for \(\varvec{\bar{F}}_{\varvec{f}}\) and \(\varvec{\bar{F}}_{\varvec{p}}\) . A polynomial regression baseline reaches \(\varvec{R}^{\varvec{2}} \varvec{= 0.920}\) , matching GPR; ML advantages lie in uncertainty quantification, localised nonlinearity capture, and robustness to broader conditions. LA-ELM records the largest force-induced gain ( \(\varvec{\Delta R}^{\varvec{2}} \varvec{= +0.074}\) ), reaching \(\varvec{R}^{\varvec{2}} \varvec{= 0.895}\) and surpassing Random Forest and XGBoost. Results confirm that the Lichtenberg Algorithm is an effective weight optimisation strategy for ELM in small-dataset manufacturing regression.