Scholarlink

<p>Hybrid laser cleaning combines continuous-wave (CW) laser preheating with pulsed laser ablation to remove piston crown carbon deposits, but conventional tuning cannot simultaneously minimize surface roughness (Sa) and carbon residue rate (RC). Four machine learning models—Support Vector Regression (SVR), XGBoost, Random Forest (RF), and Backpropagation Neural Network (BPNN)—are compared. SVR achieves optimal Sa prediction (Test R<sup>2</sup> = 0.9874, bootstrap R<sup>2</sup> = 0.985–0.998). For RC, Bayesian optimization improves SVR from R<sup>2</sup> = 0.5658 to R<sup>2</sup> = 0.8965 (58.45% increase). Feature analysis reveals divergent sensitivities: Sa responds equally to all parameters, while RC is dominated by frequency and CW laser power. NSGA-II generates 66 Pareto-optimal solutions within narrow parameter windows, identifying three operational regimes: low-Sa (0.894μm, RC = 4.3%), balanced (0.958μm, RC = 3%), and low-RC (RC = 1.9%, Sa = 1.209μm). Comprehensive validation—bootstrap resampling, cross-validation, SHAP analysis, and sensitivity testing—confirms model reliability despite limited data. This framework enables intelligent laser cleaning optimization, reducing trial-and-error costs while providing insights into laser–material interactions.</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

Multi-objective optimization of hybrid laser cleaning process parameters for carbon deposits based on bayesian-SVR and NSGA-II

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月6日

查看全文
  • Yishun Su,
  • Yong Hu,
  • Qunli Zhang,
  • Zhehe Yao,
  • Zhijun Chen,
  • Yanyi Yin,
  • Zhenzhen Yang,
  • Jianhua Yao

摘要

Hybrid laser cleaning combines continuous-wave (CW) laser preheating with pulsed laser ablation to remove piston crown carbon deposits, but conventional tuning cannot simultaneously minimize surface roughness (Sa) and carbon residue rate (RC). Four machine learning models—Support Vector Regression (SVR), XGBoost, Random Forest (RF), and Backpropagation Neural Network (BPNN)—are compared. SVR achieves optimal Sa prediction (Test R2 = 0.9874, bootstrap R2 = 0.985–0.998). For RC, Bayesian optimization improves SVR from R2 = 0.5658 to R2 = 0.8965 (58.45% increase). Feature analysis reveals divergent sensitivities: Sa responds equally to all parameters, while RC is dominated by frequency and CW laser power. NSGA-II generates 66 Pareto-optimal solutions within narrow parameter windows, identifying three operational regimes: low-Sa (0.894μm, RC = 4.3%), balanced (0.958μm, RC = 3%), and low-RC (RC = 1.9%, Sa = 1.209μm). Comprehensive validation—bootstrap resampling, cross-validation, SHAP analysis, and sensitivity testing—confirms model reliability despite limited data. This framework enables intelligent laser cleaning optimization, reducing trial-and-error costs while providing insights into laser–material interactions.