<p>This study aims to enhance the accuracy of flank wear prediction, which is essential for extending tool life and improving machining efficiency, especially in nickel-based high-temperature alloy milling where wear behavior is complex and processing conditions vary. Features strongly correlated to tool wear with minimal sensitivity to process parameters are selected based on comprehensive evaluation indicators. A multi-head self-attention one-dimensional convolutional long short-term memory (MCL) model has been developed for predicting tool wear. To enhance generalization across various machining conditions, a meta-learning approach is employed. The prediction framework integrates data-driven statistical features with physically derived milling coefficients and tool life features to enhance model accuracy. The proposed method has been validated using both the NASA dataset and a self-built dataset. Experimental results show that the MCL model achieves an <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="170_2025_15432_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> score of 0.9369, outperforming the multi-head self-attention bidirectional long short-term memory (MB) model at 0.9057 and significantly exceeding the XGBoost model, which scored 0.5241. Incorporating features based on physical models enhances predictive accuracy, as removing the tool life (<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="170_2025_15432_Article_IEq2.gif" Format="GIF" Height="19" Rendition="HTML" Resolution="72" Type="Linedraw" Width="37" /> </InlineMediaObject> <EquationSource Format="TEX">$X_{life}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>X</mi> <mrow> <mi mathvariant="italic">life</mi> </mrow> </msub> </math></EquationSource> </InlineEquation>), milling coefficient (<InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="170_2025_15432_Article_IEq3.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="33" /> </InlineMediaObject> <EquationSource Format="TEX">$X_{cut}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>X</mi> <mrow> <mi mathvariant="italic">cut</mi> </mrow> </msub> </math></EquationSource> </InlineEquation>), and traditional statistical (<InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="170_2025_15432_Article_IEq4.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="40" /> </InlineMediaObject> <EquationSource Format="TEX">$X_{trad}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>X</mi> <mrow> <mi mathvariant="italic">trad</mi> </mrow> </msub> </math></EquationSource> </InlineEquation>) features results in a reduction of the <InlineEquation ID="IEq5"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="170_2025_15432_Article_IEq5.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> score by 43.70%, 18.10%, and 29.57%, respectively. Additionally, applying meta-learning further improves performance, increasing the <InlineEquation ID="IEq6"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="170_2025_15432_Article_IEq6.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> score from 0.9369 to 0.9565, which represents a 2.09% improvement. The results demonstrate that integrating physical and statistical features enhances the accuracy and robustness of tool wear prediction. The proposed MCL model outperforms conventional approaches, with meta-learning further improving performance.</p>

Methods to improve wear prediction accuracy of nickel-based superalloy milling cutters under variable process conditions

期刊论文
发表日期：

2025年4月9日

查看全文

Liang Xie,
Bingran Li,
Zixuan Wang,
Peiqing Ye,
Hui Zhang

摘要

This study aims to enhance the accuracy of flank wear prediction, which is essential for extending tool life and improving machining efficiency, especially in nickel-based high-temperature alloy milling where wear behavior is complex and processing conditions vary. Features strongly correlated to tool wear with minimal sensitivity to process parameters are selected based on comprehensive evaluation indicators. A multi-head self-attention one-dimensional convolutional long short-term memory (MCL) model has been developed for predicting tool wear. To enhance generalization across various machining conditions, a meta-learning approach is employed. The prediction framework integrates data-driven statistical features with physically derived milling coefficients and tool life features to enhance model accuracy. The proposed method has been validated using both the NASA dataset and a self-built dataset. Experimental results show that the MCL model achieves an $R^2$ $R^{2}$ score of 0.9369, outperforming the multi-head self-attention bidirectional long short-term memory (MB) model at 0.9057 and significantly exceeding the XGBoost model, which scored 0.5241. Incorporating features based on physical models enhances predictive accuracy, as removing the tool life ( $X_{life}$ $X_{life}$ ), milling coefficient ( $X_{cut}$ $X_{cut}$ ), and traditional statistical ( $X_{trad}$ $X_{trad}$ ) features results in a reduction of the $R^2$ $R^{2}$ score by 43.70%, 18.10%, and 29.57%, respectively. Additionally, applying meta-learning further improves performance, increasing the $R^2$ $R^{2}$ score from 0.9369 to 0.9565, which represents a 2.09% improvement. The results demonstrate that integrating physical and statistical features enhances the accuracy and robustness of tool wear prediction. The proposed MCL model outperforms conventional approaches, with meta-learning further improving performance.