Scholarlink

<p>In intelligent manufacturing, embedding domain knowledge into data-driven models is essential for interpretable and causally grounded process control strategies. This paper proposes a causal learning framework that integrates Directed Acyclic Graphs (DAG) with Double Machine Learning (DML) to uncover causal pathways from process inputs to product quality. Unlike conventional black-box models, the framework explicitly encodes the causal chain “raw material addition→ chemical composition→ crack formation”, enabling both interpretable analysis and actionable optimization. The framework was validated through a case study on longitudinal crack control in the continuous casting of medium-carbon steel. The framework identifies Ca, Als, Ti, and Mn as the key elements influencing crack formation <i>via</i> combined ATE and SHAP analysis. The partial dependence plots of SHAP value quantify their optimal compositional ranges for defect suppression: when the content of Als is 0.014 to 0.019 wt pct, Ca is 0.0001 to 0.0012 wt pct, Ti is 0.031 to 0.040 wt pct, and Mn is 0.23 to 0.49 wt pct. The inverse optimization grounded in DML models translated the identified optimal compositional targets into precise, industrially actionable raw material addition specifications: 0.7769 to 1.0930&#xa0;kg/t pure calcium wire, 0.5304 to 1.1872&#xa0;kg/t calcium carbide, 42.4741 to 49.7741&#xa0;m<sup>3</sup>/t oxygen, 0.1696 to 0.7179&#xa0;kg/t high-carbon ferromanganese, 0.7446 to 0.9263&#xa0;kg/t titanium iron, and 0.8998 to 1.6139&#xa0;kg/t aluminum wire. The framework demonstrates strong generalization capability across both test set and an independent production period, offering a generalizable paradigm for integrating domain knowledge with causal machine learning in intelligent manufacturing processes.</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

Data-Driven Control of Longitudinal Crack in Medium Carbon Steel Slabs: A Causal Learning Framework Integrating DAG and DML

  • 期刊论文

  • 发表日期：

    2026年5月18日

查看全文
  • Zihan Liu,
  • Lejun Zhou,
  • Wanlin Wang,
  • Sibao Zeng,
  • Liwu Zhang,
  • Jianghua Qi,
  • Kui Chen,
  • Peng Liu

摘要

In intelligent manufacturing, embedding domain knowledge into data-driven models is essential for interpretable and causally grounded process control strategies. This paper proposes a causal learning framework that integrates Directed Acyclic Graphs (DAG) with Double Machine Learning (DML) to uncover causal pathways from process inputs to product quality. Unlike conventional black-box models, the framework explicitly encodes the causal chain “raw material addition→ chemical composition→ crack formation”, enabling both interpretable analysis and actionable optimization. The framework was validated through a case study on longitudinal crack control in the continuous casting of medium-carbon steel. The framework identifies Ca, Als, Ti, and Mn as the key elements influencing crack formation via combined ATE and SHAP analysis. The partial dependence plots of SHAP value quantify their optimal compositional ranges for defect suppression: when the content of Als is 0.014 to 0.019 wt pct, Ca is 0.0001 to 0.0012 wt pct, Ti is 0.031 to 0.040 wt pct, and Mn is 0.23 to 0.49 wt pct. The inverse optimization grounded in DML models translated the identified optimal compositional targets into precise, industrially actionable raw material addition specifications: 0.7769 to 1.0930 kg/t pure calcium wire, 0.5304 to 1.1872 kg/t calcium carbide, 42.4741 to 49.7741 m3/t oxygen, 0.1696 to 0.7179 kg/t high-carbon ferromanganese, 0.7446 to 0.9263 kg/t titanium iron, and 0.8998 to 1.6139 kg/t aluminum wire. The framework demonstrates strong generalization capability across both test set and an independent production period, offering a generalizable paradigm for integrating domain knowledge with causal machine learning in intelligent manufacturing processes.