华南理工大学学报(自然科学版) ›› 2023, Vol. 51 ›› Issue (12): 1-8.doi: 10.12141/j.issn.1000-565X.230027

所属专题: 2023年机械工程

• 机械工程 • 上一篇    下一篇

按摩机器人优化示教策略及BPNN-DMPs轨迹学习模型

翟敬梅 路东伟   

  1. 华南理工大学 机械与汽车工程学院,广东 广州 510640
  • 收稿日期:2023-01-20 出版日期:2023-12-25 发布日期:2023-04-24
  • 作者简介:翟敬梅(1967-),女,博士,教授,主要从事机电装备信息化处理与人工智能研究。E-mail: mejmzhai@scut.edu.cn
  • 基金资助:
    广东省自然科学基金资助项目(2023A1515010682)

Research on Optimized Teaching Strategy and BPNN-DMPs Trajectory Learning Model of Massage Robot

ZHAI Jingmei LU Dongwei   

  1. School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou 510640,Guangdong,China
  • Received:2023-01-20 Online:2023-12-25 Published:2023-04-24
  • About author:翟敬梅(1967-),女,博士,教授,主要从事机电装备信息化处理与人工智能研究。E-mail: mejmzhai@scut.edu.cn
  • Supported by:
    the Natural Science Foundation of Guangdong Province(2023A1515010682)

摘要:

针对按摩机器人作业对象面貌个性化差异,引用动态运动基元(DMPs)模型泛化位姿轨迹和力轨迹。为提高DMPs学习精度,提出了优化示教策略,基于按摩区域Mediapipe特征点计算作业对象间相似度,以此优选学习对象;其次,引入高斯混合回归,算法综合多次按摩信息以加强学习能力;最后,构建反向传播神经网络(BPNN)模型拟合DMPs算法的强迫项,从根本上改变了原本模型的局限性。实验表明,不增加运行时间下的BPNN-DMPs模型与DMPs、MDMPs和SADMPs算法相比,位置和姿态平均误差分别减少44.1%和54.5%、44.1%和54.5%、29.7%和46.4%。高斯混合回归能够综合多轨迹规律,优化示教策略实施效果显著,与未做优选对象相比,面部实验中的位置和姿态平均误差降低了52.3%和70.2%,标准差降低了46.3%和71.1%;背部实验中的位置、姿态和力平均误差降低了27.7%、66.7%和24.1%,标准差降低了25.7%、54.4%和44.1%。

关键词: 按摩机器人, 动态运动基元, 优化示教策略, 特征匹配, 高斯混合回归, 反向传播神经网络

Abstract:

For the individual difference of faces which are the operation object of massage robot, the dynamic motion primitives (DMPs) model was used to generalize the posture trajectory and force trajectory. Firstly, in order to improve the learning accuracy of DMPs, the study proposed an optimized teaching strategy. Based on the Mediapipe feature points in the massage area, the similarity between the operating objects was calculated to optimize the learning objects. Secondly, Gaussian mixture regression (GMR) was introduced, and the algorithm integrated multiple massage information to enhance learning ability. Finally, a back-propagation neural network (BPNN) model was constructed to fit the forced term of DMPs algorithm, which fundamentally changes the limitations of the original model. The experiment shows that the average errors of position and attitude of BPNN-DMPs model are reduced by 44.1% and 54.5%, 44.1% and 54.5%, 29.7% and 46.4% respectively, compared with DMPs, MDMPs and SADMPs algorithms without increasing the running time. Gaussian mixture regression can integrate multiple trajectory patterns and the implementation effect of the optimized teaching strategy is significant. Compared with the non-optimized object, the average errors of the position and posture of the face experiment are reduced by 52.3% and 70.2%, and the standard deviation is reduced by 46.3% and 71.1%. The average errors of position, posture and force in the back experiment decrease by 27.7%, 66.7% and 24.1%, and the standard deviation decreases by 25.7%, 54.4% and 44.1%.

Key words: massage robot, dynamic motion primitives, optimized teaching strategy, feature matching, Gaussian mixture re-gression, back-propagation neural network

中图分类号: