Scholarlink

<p>This study addresses the critical demand for high-precision, low-noise gears in humanoid robot joint reducers. Conventional gear shaping processes typically achieve only GB/T 10,095.1-2022 Grade 6-7 precision, which cannot meet the requirements of silent transmission (30–45&#xa0;dB). To solve this problem, an intelligent optimization framework based on Convolutional Deep Double Q-Network (CDDQN) is proposed for multi-error collaborative compensation in small-to-medium module gear shaping. A “machine–tool–workpiece” process system model is established to quantify the mapping relationship between motion errors of three key axes (X, C1, C2) and core accuracy indicators (cumulative pitch deviation <i>F</i><sub><i>p</i></sub>, radial runout <i>F</i><sub><i>r</i></sub>). A high-fidelity digital tooth surface model is reconstructed via reverse engineering, and the dynamic performance of gear pairs is evaluated through ADAMS multi-body dynamics simulation. The proposed CDDQN model adopts a dual-branch input architecture to fuse numerical process parameters and vibration spectrum features, effectively suppressing Q-value overestimation and enhancing feature extraction capability. Extensive experiments on 8 gear specimens and 2 different gear configurations demonstrate that the optimized gears consistently achieve GB Grade 4 precision, with transmission noise reduced by 30–45&#xa0;dB. This method provides a practical solution for high-precision gear manufacturing in advanced equipment.</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

Enhancing gear shaping precision and reducing noise via DRL

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月29日

查看全文
  • Fucong Liu,
  • Luyang Zou,
  • Zhi Liu,
  • Xiubing Jing,
  • Houjun Qi,
  • Kan Xing

摘要

This study addresses the critical demand for high-precision, low-noise gears in humanoid robot joint reducers. Conventional gear shaping processes typically achieve only GB/T 10,095.1-2022 Grade 6-7 precision, which cannot meet the requirements of silent transmission (30–45 dB). To solve this problem, an intelligent optimization framework based on Convolutional Deep Double Q-Network (CDDQN) is proposed for multi-error collaborative compensation in small-to-medium module gear shaping. A “machine–tool–workpiece” process system model is established to quantify the mapping relationship between motion errors of three key axes (X, C1, C2) and core accuracy indicators (cumulative pitch deviation Fp, radial runout Fr). A high-fidelity digital tooth surface model is reconstructed via reverse engineering, and the dynamic performance of gear pairs is evaluated through ADAMS multi-body dynamics simulation. The proposed CDDQN model adopts a dual-branch input architecture to fuse numerical process parameters and vibration spectrum features, effectively suppressing Q-value overestimation and enhancing feature extraction capability. Extensive experiments on 8 gear specimens and 2 different gear configurations demonstrate that the optimized gears consistently achieve GB Grade 4 precision, with transmission noise reduced by 30–45 dB. This method provides a practical solution for high-precision gear manufacturing in advanced equipment.