【作者】 周炜；

【导师】 罗志增；

【作者基本信息】 杭州电子科技大学 ， 控制理论与控制工程， 2009， 硕士

【摘要】 表面肌电信号(surface electromyography (SEMG)蕴涵了很多与肢体运动相关联的信息,用SEMG控制仿生假手已成为假手研究的一项重要内容,因此还原表面肌电信号中蕴含的信息,对假肢研究意义重大。本文结合课题要求,从SEMG信号的拾取、预处理、特征提取、基于SEMG信号的动作模式识别等多个方面进行了理论探索和实践。对信号的盲源分离、基于直方图和频谱分析的特征提取,及先识别后融合的分类方法进行了深入的研究,以此为目标,本文作了以下工作:(1)本文简要概括了肌电假手的研究背景、研究现状和研究意义;总结了SEMG信号处理中常用的特征提取方法,同时选择直方图和频谱分析作为本文特征提取的切入点,为接下去的研究奠定基础和方向,并介绍了几种肌电假手动作识别的常用模式分类方法;(2)本文针对多通道肌电信号采集产生的混迭现象,首次将时频分析和参考累积量算法结合,提出一种改进的盲源分离算法用于消除SEMG混迭现象。算法整合了参考累积量在统计非平稳信号信息的优势和时频分析在反映高阶统计量变化的优势,通过构造参考累积量矩阵作为对照函数,并进行时频分析,利用非正交联合对角化(Non-Orthogonal Joint Diagonalization)该矩阵,最后得到本通道的最优SEMG估计,为接下去的SEMG模式识别打下基础。(3)为有效提取SEMG信号的特征,本文提出了一种基于直方图和频谱分析的特征提取方法。作者经过分析后认为要想更好的表征SEMG信号的特征,应该从不同的角度对其特征参数进行提取,本文利用直方图统计对SEMG信号时域特征参数即两路SEMG最大波幅比值进行提取,同时结合频谱分析提取SEMG频域特征参数—功率谱比值。最后利用BP神经网络对时域和频域的两类特征参数分别进行模式识别,利用其得到的识别结果再进行D-S证据融合,得到最终的识别结果。算法既利用D-S理论来表达和处理不精确的、模糊的信息,又可以发挥神经网络的自学习、自适应和容错能力,使整个识别系统具有很强的鲁棒性,同时也能提高识别率。通过不同类特征参数的相互信息的补充,避免了单类特征参数造成的不确定性,同时也避免了特征参数简单组合可能导致的对多类别特征无法识别的问题。最后通过对本文所提出的信号预处理算法与未进行信号预处理的提取的特征参数进行比较,发现其利用本文所提算法得到的识别率优于其他方法,平均识别率达到94%,且算法的鲁棒性也更强了。更多还原

【Abstract】 Electromyography signal contains plenty information connected with body movements. It is a main task that controlling the muelectric limb by using the surface electromyography signals. So it is very usful to the research of the myelectric limb that recovering the information from the SEMG signal. According to the requirement of the research subject,the paper designs the control system of the myoelectric prosthesis based on the SEMG signal. For this, many theory exploration and practice will be done on collection of SEMG signal, feature extraction, pattern recognition of hand movements and the control of myoelectric limb. The author do some deeply research on Blind Source Separation , the feature extract by the histgram and the frequency analysis, and the classification. In order to implement the goal, the paper makes following work and innovations:(1) Firstly, the paper briefly generalizes the research background, current research situation and research significance of myoelectric prosthesis. Secondly, the paper summarizes methods of feature extraction used in SEMG signal processing, and establish the method of feature extraction is based on the analyzing to the histgram and the frequency. And the last introduce the method of pattern recognition applied in myoelectric limb.(2) In order to eliminate the signal retraction of multi-channel Surface Electromyography (SEMG), this paper proposes a new separation method based on the referenced cumulant on the time-frequency. The method combined the merits of them. First construct the time-scale cumulant matrice, and then do the time-frequency analysis, at last, the estimation of the SEMG can be performed by the non-orthogonal joint diagonalization. It is a good base for the next patten-recgonization.(3) In order to extract effectively the feature of SEMG signal, the paper firstly offers a method of feature extraction based on the histogram of the SEMG and the Time-frequency analysis. In order to token the features of SEMG signal correctly, the Feature Parameter need to be extracted from different aspects. So the paper extracted the amplitude ratio and the power spectrum ratio from the time domain and frequency domain. At last, we used the BP neural network to recognize the patten respectively and get the final result from the the D-S evidence theory. This method use the D-S evidence theory to deal with the imprecise and vague information and also to play neural network’s merits like as self-learning, adaptive and fault-tolerance capabilities, so it is a Robustness’arithmetic,and it can also improve the ratio of the patten recgonization. The method use the different paeampers to supply the more information, and it can avoid the unsure from the one feature parameter ,and also it avoid the unrecgonize because of the simple feature combine. The paper has taken a compare between the result of using a feature character and the two combine, it shows the method that the paper offered has a great advantage, its ratio even reach 94%. Off course, the paper has also another compare from the different processing methods, it also shows that the recognition effect is better from the processing with methods which the paper offered.更多还原