中国组织工程研究 ›› 2022, Vol. 26 ›› Issue (12): 1805-1811.doi: 10.12307/2022.499

• 骨与关节生物力学 bone and joint biomechanics •    下一篇

分析5种步态模式状态时下肢表面肌电信号识别人体下肢运动意图

赵祎明,王  婕,张高巍,孙建军,杨  鹏   

  1. 河北工业大学,人工智能与数据科学学院、智能康复装置与检测技术教育部工程研究中心,天津市   300130
  • 收稿日期:2021-05-14 修回日期:2021-05-18 接受日期:2021-07-14 出版日期:2022-04-28 发布日期:2021-12-13
  • 通讯作者: 王婕,博士,副教授,河北工业大学,人工智能与数据科学学院、智能康复装置与检测技术教育部工程研究中心,天津市 300130
  • 作者简介:赵祎明,男,2001年生,河北省保定市人,汉族,河北工业大学在读本科生,主要从事肌电信号分析、外骨骼机器人研究工作。
  • 基金资助:
    河北省自然科学基金(F2019202369),项目负责人:王婕

Recognition of human lower limb movement intention based on surface electromyography of lower limb corresponding to five gaits

Zhao Yiming, Wang Jie, Zhang Gaowei, Sun Jianjun, Yang Peng   

  1. School of Artificial Intelligence & Engineering Research Center of Intelligent Rehabilitation Device and Detection Technology Ministry of Education, Hebei University of Technology, Tianjin 300130, China
  • Received:2021-05-14 Revised:2021-05-18 Accepted:2021-07-14 Online:2022-04-28 Published:2021-12-13
  • Contact: Wang Jie, MD, Associate professor, School of Artificial Intelligence & Engineering Research Center of Intelligent Rehabilitation Device and Detection Technology Ministry of Education, Hebei University of Technology, Tianjin 300130, China
  • About author:Zhao Yiming, School of Artificial Intelligence & Engineering Research Center of Intelligent Rehabilitation Device and Detection Technology Ministry of Education, Hebei University of Technology, Tianjin 300130, China
  • Supported by:
    the Natural Science Foundation of Hebei Province, No. F2019202369 (to WJ)

摘要:

文题释义:
表面肌电信号:是一种生物电信号,是人体肌肉在进行运动时产生的电位在皮肤表面的叠加。表面肌电信号产生于大脑运动意图产生之后,肌肉真正收缩产生之前,是一种十分接近于人体运动初衷的信号,被广泛应用于康复、医疗和体育等研究领域。
概率神经网络:是一种结构简单的神经网络算法,融合了贝叶斯决策理论,能用线性学习算法实现非线性学习算法的功能,属于径向基神经网络的一种,拥有输入层、隐含层、求和层和决策层这4层神经节点,被广泛应用于解决模式分类问题。

背景:下肢外骨骼机器人用于恢复患者的基本行走能力,要实现人-机的高度协同和制定机器人控制策略,就要对包括步态在内的人体下肢运动意图进行识别。
目的:提出一种新式分类器,利用人体表面肌电信号进行步态识别。
方法:首先,采集人体5种步态模式的下肢表面肌电信号;其次,对采集的信号进行降噪和特征提取以获得5种步态模式的数据集,并将其两两结合形成10组数据集;再次,将每组数据集的训练集按照AdaBoost算法规则输入至概率神经网络进行训练;从次,将已训练概率神经网络集成获得10个强分类器;最后,将测试集样本输入到10个强分类器并采取投票机制判断其所述类别。
结果与结论:实验使用改进AdaBoost算法对5种步态模式进行识别,取得了90.2%的平均识别率;将其与仅使用一个概率神经网络进行对比,单独使用概率神经网络时的平均识别率为68.2%,低于改进AdaBoost分类器的平均识别率;将弱分类器更换为支持向量机、BP神经网络、决策树构建强分类器进行对比,当仅使用一个弱分类器时,支持向量机的平均识别率高于概率神经网络,然而将弱分类器使用改进AdaBoost算法集成以后,以概率神经网络为弱分类器时的识别率相较于以支持向量机为弱分类器时84.3%的识别率更高。

https://orcid.org/0000-0003-3522-6213 (赵祎明) 

中国组织工程研究杂志出版内容重点:人工关节;骨植入物;脊柱;骨折;内固定;数字化骨科;组织工程

关键词: 步态识别, 表面肌电信号, 巴特沃斯滤波, 自适应滤波, 特征提取, 概率神经网络, AdaBoost算法, 多分类问题

Abstract: BACKGROUND: The lower limb exoskeleton robot is used to recover the basic walking ability of patients. To realize the high coordination between humans and robots and to design robot control strategies, it is necessary to recognize the movement intention of the lower limb which includes gait.  
OBJECTIVE: To propose a new classifier to recognize gait using surface electromyography.
METHODS:  Firstly, surface electromyography data of lower limbs in five gait patterns were collected. Secondly, the collected signals were denoised and the features were extracted to obtain the data sets of five gait patterns. After that, two of them were combined and 10 groups of data sets were built. Thirdly, the probabilistic neural networks were trained with the AdaBoost algorithm based on each group of data sets. Fourthly, ten strong classifiers were obtained by ensemble of trained probabilistic neural networks. Finally, the testing samples were input into 10 strong classifiers and the categories were judged by voting.  
RESULTS AND CONCLUSION: An average recognition rate of 90.2% was achieved in the experiment by improved AdaBoost algorithm to recognize the five gait patterns. Compared with separate probabilistic neural network, the average recognition rate of separate probabilistic neural network was 68.2%, which was lower than that of improved AdaBoost algorithm. When the weak classifier was replaced by support vector machine, BP neural network or decision tree was used to construct strong classifiers for comparison. When a weak classifier was utilized, the average recognition rate of separate support vector machine was higher than that of separate probabilistic neural network. However, but the average recognition rate of strong classifier based on probabilistic neural network was higher than that 84.3% based on support vector machine with improved AdaBoost algorithm.

Key words: gait recognition, surface electromyogram, butterworth filter, adaptive filtering, feature extraction, probabilistic neural network, AdaBoost algorithm, multi-classification problem

中图分类号: