Scholarlink

<p>Modern manufacturing demands the integration of intelligence into machining process to achieve autonomous operations. While CAD/CAM-integrated CNC machine tools have automated substantial portions of milling operations, the critical decisions regarding cutting parameter optimization remain empirically driven. Current optimization approaches encounter three challenges: single-objective orientation, poor generalizability across tool-workpiece combinations, and inadequate dynamic adaptability. This study intends to bridge these gaps by developing a deep reinforcement learning (DRL) framework for autonomous multi-objective optimization. The framework employs a cutting tool as an intelligent agent interacting with a physics-informed simulation environment, aiming at optimizing cutting parameters via a multi-criteria reward function (e.g. torque stabilization, material removal rate, tool longevity). Two types of DRL schemes, discrete and continuous action spaces, were developed based on Deep Q-Network (DQN) and Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithms, respectively, for cutting torque stabilization in milling process. Case studies showed that the framework has high potential in self-optimization of feed rate to achieve the objective. Compared to the conventional operator-defined parameters, DRL-derived parameters significantly reduced torque variance by 80% while shortening machining time by 50%. Comparative analysis of discrete and continuous schemes demonstrated the superiority of the TD3-based continuous action scheme, achieving 85.4 ~ 97.3% torque stabilization success (± 5% target) versus DQN-based discrete action scheme’s 74.1 ~ 91.0%, with higher reward stability and faster optimization. This work establishes a foundation for intelligent manufacturing systems capable of autonomous process refinement, transcending human expertise limitations in multi-objective optimization.</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

Self-optimization of machining parameters for intelligent manufacturing: an offline multi-objective deep reinforcement learning framework

  • 期刊论文

  • 发表日期：

    2026年4月8日

查看全文
  • Libo Zhou,
  • Toshihiro Komatsu,
  • Yusuke Morishita,
  • Hirotaka Ojima,
  • Han Huang,
  • Dekui Mu,
  • Jun Zhao,
  • Wei Hang,
  • Huapan Xiao,
  • Jiaming Zhan

摘要

Modern manufacturing demands the integration of intelligence into machining process to achieve autonomous operations. While CAD/CAM-integrated CNC machine tools have automated substantial portions of milling operations, the critical decisions regarding cutting parameter optimization remain empirically driven. Current optimization approaches encounter three challenges: single-objective orientation, poor generalizability across tool-workpiece combinations, and inadequate dynamic adaptability. This study intends to bridge these gaps by developing a deep reinforcement learning (DRL) framework for autonomous multi-objective optimization. The framework employs a cutting tool as an intelligent agent interacting with a physics-informed simulation environment, aiming at optimizing cutting parameters via a multi-criteria reward function (e.g. torque stabilization, material removal rate, tool longevity). Two types of DRL schemes, discrete and continuous action spaces, were developed based on Deep Q-Network (DQN) and Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithms, respectively, for cutting torque stabilization in milling process. Case studies showed that the framework has high potential in self-optimization of feed rate to achieve the objective. Compared to the conventional operator-defined parameters, DRL-derived parameters significantly reduced torque variance by 80% while shortening machining time by 50%. Comparative analysis of discrete and continuous schemes demonstrated the superiority of the TD3-based continuous action scheme, achieving 85.4 ~ 97.3% torque stabilization success (± 5% target) versus DQN-based discrete action scheme’s 74.1 ~ 91.0%, with higher reward stability and faster optimization. This work establishes a foundation for intelligent manufacturing systems capable of autonomous process refinement, transcending human expertise limitations in multi-objective optimization.