This work investigates the dry turning of polyoxymethylene reinforced with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(25\%\)</EquationSource> </InlineEquation> fibreglass (POM C GF25%) with a carbide insert, following an orthogonal Taguchi <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(L_{32}\)</EquationSource> </InlineEquation> experimental plan. The influence of the four cutting parameters (nose radius r, cutting speed <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(V_c\)</EquationSource> </InlineEquation>, depth of cut <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(a_p\)</EquationSource> </InlineEquation> and feed f) on surface roughness (<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(R_a\)</EquationSource> </InlineEquation>), tangential cutting force (<InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(F_z\)</EquationSource> </InlineEquation>), cutting power (<InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(P_c\)</EquationSource> </InlineEquation>) and cutting energy (<InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(E_c\)</EquationSource> </InlineEquation>) is first quantified by ANOVA and described by linear-with-interactions RSM polynomials, which serve as an interpretable parameter-effect analysis. Five machine-learning regressors (SVR, Random Forest, Gradient Boosting, XGBoost and a Deep Neural Network) are then trained on the same <InlineEquation ID="IEq9"> <EquationSource Format="TEX">\(L_{32}\)</EquationSource> </InlineEquation> data and compared by five-fold cross-validation; Gradient Boosting and XGBoost emerge as the best generalisers and are retained as per-output surrogates (<InlineEquation ID="IEq10"> <EquationSource Format="TEX">\(R_a\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq11"> <EquationSource Format="TEX">\(P_c\)</EquationSource> </InlineEquation>: GB; <InlineEquation ID="IEq12"> <EquationSource Format="TEX">\(F_z\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq13"> <EquationSource Format="TEX">\(E_c\)</EquationSource> </InlineEquation>: XGBoost). A Gaussian-process smoothing of these surrogates is used as the objective function of four metaheuristic algorithms (MOALO, MOGWO, MODA and MOGOA) to construct three bi-objective Pareto fronts: (<InlineEquation ID="IEq14"> <EquationSource Format="TEX">\(R_a\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq15"> <EquationSource Format="TEX">\(F_z\)</EquationSource> </InlineEquation>), (<InlineEquation ID="IEq16"> <EquationSource Format="TEX">\(R_a\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq17"> <EquationSource Format="TEX">\(P_c\)</EquationSource> </InlineEquation>) and (<InlineEquation ID="IEq18"> <EquationSource Format="TEX">\(R_a\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq19"> <EquationSource Format="TEX">\(E_c\)</EquationSource> </InlineEquation>). The four algorithms produce essentially overlapping fronts. The two main contributions of the study are (i) the experimental characterisation and modelling of POM C GF25% machining and (ii) the use of cross-validated ML surrogates, in place of the conventional RSM polynomials, as the objective function of the multi-algorithm metaheuristic search.

Machine-learning surrogates for the multi-objective optimisation of cutting parameters in dry turning of POM C GF25%

期刊论文
发表日期：

2026年6月25日

查看全文

Nahla Djouambi,
Mohamed Athmane Yallese,
Mounia Kaddeche,
Abderraouf Maoudj,
Ahmed Khellaf,
Mourad Nouioua

摘要

This work investigates the dry turning of polyoxymethylene reinforced with \(25\%\) fibreglass (POM C GF25%) with a carbide insert, following an orthogonal Taguchi \(L_{32}\) experimental plan. The influence of the four cutting parameters (nose radius r, cutting speed \(V_c\) , depth of cut \(a_p\) and feed f) on surface roughness ( \(R_a\) ), tangential cutting force ( \(F_z\) ), cutting power ( \(P_c\) ) and cutting energy ( \(E_c\) ) is first quantified by ANOVA and described by linear-with-interactions RSM polynomials, which serve as an interpretable parameter-effect analysis. Five machine-learning regressors (SVR, Random Forest, Gradient Boosting, XGBoost and a Deep Neural Network) are then trained on the same \(L_{32}\) data and compared by five-fold cross-validation; Gradient Boosting and XGBoost emerge as the best generalisers and are retained as per-output surrogates ( \(R_a\) , \(P_c\) : GB; \(F_z\) , \(E_c\) : XGBoost). A Gaussian-process smoothing of these surrogates is used as the objective function of four metaheuristic algorithms (MOALO, MOGWO, MODA and MOGOA) to construct three bi-objective Pareto fronts: ( \(R_a\) , \(F_z\) ), ( \(R_a\) , \(P_c\) ) and ( \(R_a\) , \(E_c\) ). The four algorithms produce essentially overlapping fronts. The two main contributions of the study are (i) the experimental characterisation and modelling of POM C GF25% machining and (ii) the use of cross-validated ML surrogates, in place of the conventional RSM polynomials, as the objective function of the multi-algorithm metaheuristic search.