SIA OpenIR  > 机器人学研究室
基于肌电分析的人机交互关键技术研究
Alternative TitleThe Key Technologies in EMG-based Human-Robot Interaction
丁其川1,2
Department机器人学研究室
Thesis Advisor韩建达
ClassificationR318.04
Keyword人机交互 肌电信号 仿生假肢 动作分类 运动估计
Call NumberR318.04/D58/2014
Pages123页
Degree Discipline模式识别与智能系统
Degree Name博士
2014-12-05
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract表面肌电信号是人体自身的资源,蕴含着与人体运动相关的丰富信息,用它作为交互信息媒介以构建人机交互系统有着天然的优势,而通过肌电信号识别人体运动意图是实现人机自然交互过程的关键环节。由于肌电信号具有个体特异性,并且易受皮肤汗液、肌肉疲劳、身体疾病等因素的影响,因此由肌电信号精确地识别出人的运动意图是一项极具挑战性的任务。 基于肌电信号的运动识别通常涉及两类问题,一类是利用模式分类方法识别人体离散动作模态,另一类是利用生理学模型或模式回归方法估计人体关节连续运动量。目前大多数研究集中于前者,遴选肌电信号特征及改进分类算法,以提高动作识别准确率是多数研究者关注的焦点,但在面向实际应用时,如何保证分类系统在传感器出现故障等非理想状况下的鲁棒性却未见有研究成果,而这一问题直接关系到系统的实用性与安全性;相比离散动作模态分类,关节连续运动估计更有理论意义与应用价值,特别是在构建人机合一系统中,如仿生假肢、外骨骼机器人、医疗康复机器人等,人机运动过程连续匹配是保证使用者安全并实现机器人辅助功能的前提,但是这类问题的研究较少,而且现有的连续运动模型都是“开环”结构形式,无法校正预测误差的缺点使得模型的估计精度受限。本文针对以上问题展开深入研究,主要内容包括动作分类系统的容错机制构建,以及连续运动模型的“闭环”结构设计。具体内容如下: 首先,针对肌电分类系统实际应用中,出现因电极脱落、传感器损坏等故障造成的肌电信号部分缺失问题,提出基于高斯混合模型(Gaussian Mixture Model, GMM)的边缘化及条件均值归错容错分类方法。为了规避直接使用高维样本训练GMM时遇到的运算量大、模型精度低及稳定性差等缺点,提出通过降维模型扩展全维的方法实现高维GMM的间接拟合,仿真验证了扩展全维GMM可以有效近似真实GMM。利用提出的容错分类方法离线分析了不同通道肌电信号丢失时动作识别效果,与传统零归错或均值归错方法相比,提出的边缘化方法性能优越。随后将边缘化容错分类机制融合到肌电假手控制系统,并通过在线实验验证了肌电假手鲁棒性的提高。 其次,为了建立可以揭示运动产生过程的关节连续运动模型,合理简化Hill肌肉力模型,并结合关节前向动力学,构建直接映射肌电信号到关节角速度/角度等运动量的状态方程;然后提出利用肌电信号特征拟合量测方程,用以校正建模误差导致的预测偏差;联合两方程构成状态空间方程,并利用非线性滤波方法递归估计运动状态。随后,考察了人体外部负载扰动对模型估计的影响,并提出基于负载的合成最大归一化方法,使得在一种负载条件下辨识的模型对其它负载具有通适性。实验结果验证了所提出方法的有效性,而控制机械臂模仿人体肘关节运动的实验更显示了构建的模型应用于辅助康复系统的潜能。 最后,从模式回归的角度建立联系多肌肉表面肌电与多关节连续运动的模型。针对传统开环模型预测偏差无法评估修正的缺点,提出一种处理多维序列数据的闭环估计框架。将原始每个输入样本向量分割成两个向量,一个作为新的输入向量,另一个作为量测输出,而原始输出样本则作为中间状态向量,然后分别拟合前向预测模型和后向校正模型,两模型联合构成闭环结构模型,于是可以根据模型的函数形式及噪声特性,选择合适的滤波算法递归估计中间状态。随后,利用提出的方法建立了上肢多关节连续运动模型,并在视觉运动捕获系统的辅助下,进行实验验证,大量实验结果以及与开环估计结果的比较显示了所提出方法的有效性及性能提高。该项研究将应用于构建上肢康复机器人自主控制系统。 本文工作拓展了基于肌电分析的人机交互技术研究内容,对提升肌电系统的鲁棒性与可靠性进行了深入探索,为后续实现辅助机器人的自然运动控制提供了理论基础和技术支撑。另外,本文提出的高维GMM建模方法与闭环估计框架也为解决其它模式识别问题提供借鉴。
Other AbstractSurface electromyogram (sEMG) signals are human’s own resources, which contain a wealth of information associated with human motion. Thus, there is a natural advantage to utilize sEMG signals as the command media to construct human-robot interaction (HRI) systems. The key technology of realizing a natural HRI is to identify human’s motion intentions by using the sEMG signals. Because sEMG signals have individual specificity, and are easily influenced by skin sweat, muscle fatigue and body disease, it is a challenging task to accurately identify the motion intentions from sEMG signals. EMG-based motion recognition usually involves two kinds of problems. One is identifying discrete motion modes by using pattern classification algorithms. The other is estimating continuous movements of joints by using a physiology model or regression model. Most previous studies focused on the former. Selecting proper sEMG features and improving classification algorithms to enhance motion recognition accuracy have drawn many researchers’ attentions. However, there are no research results on how to ensure classification systems’ robustness under non-ideal conditions during practical application phase, such as sensors’ failure or electrode disconnecting; even through the problem is directly related to the practicality and safety of myoelectric systems. Compared to classifying discrete motion modes, estimating continuous movements of joints has more theoretical significance and application value, especially in designing some human-robot hybrid systems, such as bionic prosthetics, exoskeleton robots and rehabilitation robots. In such a system, the robot’s motion should continuously match human’s motion to ensure user’s safety and then provide efficient assistant. Only a few studies have focused on the estimation of joints’ continuous movements. The existing models for continuous motion estimation are of “open-loop” structures, in which the prediction deviation cannot be corrected, so the models’ estimation accuracies are limited. In order to solve the existing problems, this paper will investigate how to construct a fault tolerance mechanism for motion classification systems, and design a “closed-loop” structure for continuous motion estimation. Details are as follows: Firstly, this paper proposes the marginalization/conditional-mean imputation method based on Gaussian Mixture Model (GMM) to achieve a fault-tolerant classification with respect to the sEMG signals with missing data caused by electrodes disconnecting or sensors failuar in practical applications. To avoid the disadvantages of complex computation, low model accuracy and poor stability caused due to fitting GMM by directly using high-dimensional training samples, a dimension-expansion method is proposed to build GMM by extending a reduced-dimensional GMM to the full-dimensional one. Simulations show that the extended full-dimensional GMM can effectively approximate the true GMM. Then, the proposed fault-tolerant classification method is utilized to offline analyze the performance of motion recognition when sEMG signals are missing from different channels. Experimental results show that the proposed marginalization method is superior to the normal zero imputation and mean imputation. Afterward, the fault-tolerant classification mechanism is involved into a multiple-DOFs myoelectric hand, and the online experiments show the proposed method can effectively improve the robustness of the myoelectric hand. Secondly, a continuous movement model that can reveal the generation process of a joint’s movement is built by simplifying the Hill-type muscle model and then involving the joint forward dynamics. The built state equation maps sEMG signals to the joint angers and angular velocities. Moreover, sEMG features are extracted to build the measurement equation, which will be employed to correct the prediction error caused by the model uncertainty. After combining the state equation and the measurement equation, a state-space equation is obtained and then the movement states can be estimated by using nonlinear filtering algorithms. Furthermore, the paper investigates the impact of external loads on the movement estimation and proposes a synthesized-maximum normalization method to make a “reference” model identified by one load suitable to other different loads. Experimental results demonstrate the effectiveness of the proposed methods, and a robotic arm is commanded to follow the human elbow movement estimated by the proposed model, which shows the potentials of EMG-based robot-assisted rehabilitation. Thirdly, a model for estimating the continuous movement of multi-joints with sEMG signals is built by using pattern regression methods. Because traditional “open-loop” models cannot correct the prediction error, a closed-loop estimation framework is proposed to handle multi-dimensional data. In the framework, every original input vector is divided into two vectors. One is used as the new input, and the other is used as measurement output. Besides, the original output vector is considered as the middle state. Then, a foreword prediction model and a backward correction model are fitted respectively, so a state-space model can be obtained by combining the two models. A proper filtering algorithm can be chosen to estimate the states recursively according to the model’s expressions and noise characteristics. Afterward, the proposed closed-loop estimation framework is ustilized to build the motion model of human upper limb. The performance of the built model is verified by experiments under the assistance of a visual motion capture system. Comprehensive experimental results and comparisons with the “open-loop” model show the effectiveness and performance improvement of the built model. The proposed method will be applied to construct the autonomous control system of a robot for the upper limb rehabilitation. This paper expands the researches on EMG-based HRI and deeply investigates the methods to enhance the robustness and reliability of myoelectirc systems. Thus, the study provides a theoretical basis and technical support for achieving a natural control of assistant robots in the future. In addition, the proposed method of fitting high-dimensional GMM and closed-loop estimation framework can also be employed to solve other pattern recognition problems.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/16793
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
丁其川. 基于肌电分析的人机交互关键技术研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2014.
Files in This Item:
File Name/Size DocType Version Access License
基于肌电分析的人机交互关键技术研究.pd(12057KB) 开放获取CC BYApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[丁其川]'s Articles
Baidu academic
Similar articles in Baidu academic
[丁其川]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[丁其川]'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.