Scholarlink

<p>In medium-speed wire electrical discharge machining (MS-WEDM), multi-pass cutting is widely used to balance surface quality, geometric accuracy, and machining efficiency. This study proposes a general machine learning framework for multi-objective modeling and parameter optimization that simultaneously targets surface roughness (Ra), material removal rate (MRR), and drum shape error (DSE). Eight process variables are considered: peak current (I<sub>p1</sub>), pulse-on time (T<sub>on1</sub>), and duty factor (DF<sub>1</sub>) in rough cutting; peak current (I<sub>p2</sub>), pulse-on time (T<sub>on2</sub>), duty factor (DF<sub>2</sub>), and wire offset (WO) in trim cutting; and servo feed rate (FR). Four regression models are benchmarked, namely support vector regression (SVR), random forest (RF), gradient boosting regression trees (GBRT), and artificial neural networks (ANN). Hyperparameters are tuned by Bayesian optimization with five-fold cross-validation. ANN delivers the best predictive performance for Ra, MRR, and DSE, with test R² values of 0.94, 0.99, and 0.93, respectively. To interpret the ANN, Shapley additive explanations (SHAP) quantify the importance and response patterns of process parameters for each objective. Using ANN as the fitness evaluator, the non-dominated sorting genetic algorithm II (NSGA-II) solves the tri-objective optimization problem and yields a well-distributed and balanced Pareto front. Validation on five representative Pareto-optimal solutions shows low RMSE values of 0.066&#xa0;μm for Ra, 1.421&#xa0;mm²/min for MRR, and 0.43&#xa0;μm for DSE. These results confirm that the ANN plus NSGA-II strategy provides reliable guidance for high-performance WEDM and is transferable to other multi-objective manufacturing optimization tasks.</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

A generalizable machine learning framework for multi-objective optimization of multi-pass WEDM

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Qinglang Zhou,
  • Xiaodong Yang,
  • Zhen Qiu,
  • Xianming Liu,
  • Xiaodong Xing,
  • Xiaoming Duan

摘要

In medium-speed wire electrical discharge machining (MS-WEDM), multi-pass cutting is widely used to balance surface quality, geometric accuracy, and machining efficiency. This study proposes a general machine learning framework for multi-objective modeling and parameter optimization that simultaneously targets surface roughness (Ra), material removal rate (MRR), and drum shape error (DSE). Eight process variables are considered: peak current (Ip1), pulse-on time (Ton1), and duty factor (DF1) in rough cutting; peak current (Ip2), pulse-on time (Ton2), duty factor (DF2), and wire offset (WO) in trim cutting; and servo feed rate (FR). Four regression models are benchmarked, namely support vector regression (SVR), random forest (RF), gradient boosting regression trees (GBRT), and artificial neural networks (ANN). Hyperparameters are tuned by Bayesian optimization with five-fold cross-validation. ANN delivers the best predictive performance for Ra, MRR, and DSE, with test R² values of 0.94, 0.99, and 0.93, respectively. To interpret the ANN, Shapley additive explanations (SHAP) quantify the importance and response patterns of process parameters for each objective. Using ANN as the fitness evaluator, the non-dominated sorting genetic algorithm II (NSGA-II) solves the tri-objective optimization problem and yields a well-distributed and balanced Pareto front. Validation on five representative Pareto-optimal solutions shows low RMSE values of 0.066 μm for Ra, 1.421 mm²/min for MRR, and 0.43 μm for DSE. These results confirm that the ANN plus NSGA-II strategy provides reliable guidance for high-performance WEDM and is transferable to other multi-objective manufacturing optimization tasks.