SIA OpenIR  > 机器人学研究室
基于肌电分解的智能假肢系统人机交互技术研究
Alternative TitleKey technologies in sEMG-Decomposition-based Human-Prostheses Interaction
李自由1,2
Department机器人学研究室
Thesis Advisor刘光军 ; 赵新刚
Keyword表面肌电信号 智能假肢 人机交互 机器学习 模式识别
Pages121页
Degree Discipline机械电子工程
Degree Name博士
2020-11-27
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract目前我国各类残疾人总数已接近一个亿,其中肢体残疾人数量为最多,约占29%。肢体残疾,尤其是手臂的功能缺失,严重影响残疾人的工作与日常活动。现阶段的大部分假肢仅具有装饰美观或简单的被动抓握功能,使用起来费时费力。而表面肌电信号,相对于其他生理电信号而言,具有信噪比高、信息量丰富与易于开发可穿戴设备等优点,被认为是肢体运动意图估计与假肢交互最具潜力方式之一。但现阶段的肌电假肢系统,仍然面临着功能性弱、识别效果差等问题。因此,基于表面肌电信号准确地、稳定地进行肢体运动意图识别是一项极具研究价值和挑战的任务。位于皮肤表面的肌电信号传感器,其采集到的微弱电信号,主要来源于所募集的诸多运动单元动作电位。这些动作电位反映了不同肌肉的刺激收缩状态与强度,因此,表面肌电信号中蕴含着大量肢体运动、肌肉力等信息,通过一系列信号解码方法,可用于识别肢体运动意图。基于表面肌电信号的肢体运动意图识别,根据其应用目标,可分为两类:1)离散动作分类,2)连续运动估计。针对这两类应用目标,主要采用有监督学习的基本框架,目前大多数的研究工作主要集中在特征提取、降维方法与分类模型等环节,以提高肌电识别的准确性。但在实际的智能假肢交互应用中,基于表面肌电信号的识别方法常常受到各种非理想因素的干扰,如电极偏移、个体性差异和肌肉疲劳等。因此,本论文的研究工作,主要围绕肌电识别系统中的准确性和鲁棒性两个关键科学问题,开展了自适应肌电分解、离散动作识别、连续运动估计,以及非理想肌电识别等方面的研究。具体研究内容与进展如下:首先,针对传统肌电分解方法中涉及人为参数过多、分解结果的一致性差等缺点,提出了一种基于小波变换的自适应肌电分解方法。提出两类高斯小波母版作为运动单元动作电位的形式化函数;通过尺度变换,定义了不同种类的实际运动单元动作电位;基于小波变换中的卷积运算,构造了一种自适应的模板匹配方法;并通过信号重构与逐次信号剥离,实现一种迭代分解过程。实验结果表明,在不同收缩力的实验中,所提出肌电分解方法的结果合理地反映了运动单元种类、数量、发放频率与强度等信息。其次,在基于有监督学习的肌电信号离散动作识别一般框架中,讨论了肌电特征或特征组合与不同降维算法之间的关系;并利用TSFRESH等时序特征库,将传统的肌电信号时频域等特征扩展至上千维,采用集成学习中的随机森林等模型进行离散动作分类研究,取得了较好的识别精度,优于其他相关研究结果。随后,开发了面向多种肌电信号传感器和多款智能假肢系统的人机交互软硬件系统,前述研究工作均进行了大量的在线测试,并成为后续研究的主要测试平台。之后,开展了基于表面肌电信号的下肢连续运动估计研究工作。针对传统基于Hill肌肉模型中涉及大量生理假设与模型简化等问题,该部分主要采用数据驱动方式,侧重于系统的输入与输出;在直接映射模型的基础上,进一步考虑了关节运动过程中的时间序列特点,利用滑动滤波等方法,降低了直接映射模型中的波动干扰,提高了系统在多个指标上的回归性能,实现了基准的下肢多关节角度预测精度。最后,重点关注了非理想肌电中的电极偏移干扰问题。针对肌电交互系统中常见的一类电极旋转偏移问题,提出了一种偏移估计与自适应校正方法。该方法首先建立相对于电极静止的极坐标系,提出了用于估计旋转偏移的活跃极角;进而建立基于偏移角度的线性变换模型,在特征空间内,对电极偏移位置下的肌电特征进行自适应校正。大量的结果表明,所提出方法的识别精度远高于未进行校正的模型识别精度。因此,该方法不仅有效提高了肌电交互系统的鲁棒性,且仅利用单类别肌电数据实现了多类别的动作分类,极大地降低了使用者在多次使用时的训练成本与学习负担。本论文的研究工作深入探究了基于肌电信号的智能假肢系统人机交互中的关键技术,包括自适应肌电分解技术,肌电离散帯连续识别技术,以及非理想肌电抗干扰技术。在肌电识别的准确性与鲁棒性方面积累了大量的理论方法与实验经验,为后续实现自然、准确与稳定的肌电假肢人机交互提供了理论与工程支撑。
Other AbstractAt present, the total number of various types of the disabilities in China has approximated to 100 million. Among them, the number of physically disabled is the largest, accounting for about 29\%. Physical disabilities, especially disfunctions of arm and hand, seriously affect their work and daily activities. But, most of the prostheses at current stage only have beautiful decorations or simple passive grasping functions, which are time-consuming and laborious to use. Surface electromyogram (sEMG) based interaction with intelligent prostheses is regarded as the most promising and potential manners to help the physically disabled, because of high signal-noise ratio of sEMG, rich kinematics and dynamics information of sEMG, and feasibilities to develop wearable devices. However, the sEMG-based prostheses are still faced with weak functions and poor recognition accuracies. Therefore, it is a task of great research value and challenge to study accurately and robustly sEMG-based recognition and interaction with intelligent prostheses. Signals of sEMG, which are acquired by electrodes on the skin, are composed of many motor unit action potentials. These action potentials reflect the status and magnitude of different motor units or muscles. Therefore, plenty of kinematics and dynamics information of sEMG can be decoded by intelligent methods and used to recognize intentions of human body. There are two goals of sEMG-based recognition: 1) sEMG-based classification of discrete gestures, and 2) sEMG-based regression of joints' kinematics and dynamics. For these application goals, most researchers are focusing on improving classification accuracies by the supervised learning based diagram, which consists of feature extraction, dimension reduction and classification methods. However, in real applications of sEMG-based prostheses, sEMG signals are always disturbed by many non-ideal factors, including electrode shifts, individual differences, and muscle fatigue. Therefore, this paper is focusing on two key scientific issues: accuracy and robustness of sEMG-based recognition. A lot of research work of sEMG decomposition, sEMG-based classification and regression, and non-ideal sEMG-based interaction are conducted. Details are as follows: Firstly, for the problems of too many artificial parameters in traditional decomposition methods and poor consistency of decomposition results, an adaptive decomposition method based on the wavelet transform is proposed. Two Gaussian wavelets are first introduced as the formal function of existing decomposed MUAP waveforms. With the help of scaled transformation in wavelet transform theory, actual MUAP waveforms can be defined as scaled Gaussian wavelets. An adaptive template matching method is proposed around spikes in signals and valid action potentials are re-constructed by these scaled Gaussian wavelets. The re-construction can be iteratively conducted. Results in different muscle contraction levels show statistics of types and numbers of different MUAPs, and their firing rates and magnitudes. Secondly, in the general diagram of sEMG-based recognition, study of feature combinations and reduction methods is conducted and discussed. With the help of TSFRESH, thousands of sEMG features are extracted and random forest classifier in ensemble learning is used for sEMG-based classification. Accuracies in NinaPro datasets of classifying more than 50 gestures show advantageous performances. Afterwards, hardware and software human-robot interaction systems are developed and tested for kinds of sEMG acquisition systems and intelligent prostheses. These systems provide valid platforms for further research. Then, sEMG-based regression methods and experiments are conducted for predicting continuous joints' angles from multiple sEMG signals. For the problems of many assumptions and simplifications of Hill-based models, data-driven models are trained for regression prediction, which focus on the inputs and outputs. With the direct mapping data-driven models, sequtiential characteristics of sEMG signals and joint movements are further taken into consideration. Temporally smoothed regression techniques, like sliding filters, reduce rapid fluctuations of sEMG-based predictions, and improve the performance metrics. Finally, we are focusing on the problem of electrode shifts in non-ideal conditions. For the problem of low recognition accuracies interfered by the ring-electrode rotation shifts, an electrode shifts estimation and adaptive correction solution is proposed. The method first estiblishes a polar coordinate system that is stationary relative to the right sEMG acquisition system, and defines an activation polar angle to measure electrode rotation shifts. A linear transformation is built between initial and interfered feature space based on the estimated shifts. Results of plenty of experiments, show that the recognition accuracies of the proposed method are much higher than those of non-corrected models. In conclusion, the proposed method improves the robustness of sEMG-based recognition systems, classifys multiple gestures only by one-label newly acquired data, and reduces the training time and learning burden for users. This paper conducts intensive studies on the critical technologies of sEMG-based human and intelligent prostheses interaction, including adaptive decomposition algorithms, sEMG-based classification and regression technologies, and rubust recognition methods for non-ideal sEMG. Basic theories and engineering experience are obtained in the aspects of sEMG-based accuracy and robustness. These research studies lay a solid foundation and support for further development of intelligent prostheses for the disabled.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/27977
Collection机器人学研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院大学
Recommended Citation
GB/T 7714
李自由. 基于肌电分解的智能假肢系统人机交互技术研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2020.
Files in This Item:
File Name/Size DocType Version Access License
基于肌电分解的智能假肢系统人机交互技术研(13707KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[李自由]'s Articles
Baidu academic
Similar articles in Baidu academic
[李自由]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[李自由]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.