【作者】 任小梅；

【导师】 王志中；

【作者基本信息】 上海交通大学 ， 生物医学工程， 2006， 博士

【摘要】 针电极肌电信号是以针电极和肌电仪为检测手段,获取肌肉收缩时被激活的所有运动单元所发放的动作电位序列,以及检测仪器噪声和周围环境产生的噪声迭加后所产生的信号。它包括电极检测范围内的运动单元动作电位(MUAP)、远离电极的运动单元产生的动作电位,以及仪器产生的随机干扰和工频干扰等。肌电信号分解是对检测信号中的噪声进行抑制、提取有意义的MUAP、按照发放它们的运动单元的不同对其识别和分类的过程。肌电信号分解包括四个必要步骤,即:去除噪声和提取有用的MUAP、选择和提取MUAP波形的特征、对MUAP聚类,以及对已分类MUAP进行有监督分类。本文针对这些步骤,在对目前已有肌电信号分解方法进行分析的基础之上,基于对现有先进的模式分类和信号处理技术的改进方法,提出了一些自动的肌电信号分解系统改进方案。第一,在数据的预处理过程中,本文首先采用对数据归一化、再利用小波阈值估计技术抑制随机噪声、然后利用小波滤波技术滤除背景噪声、最后采用幅度滤波法提取MUAP波形的方法。不仅克服了以往方法中信号对阈值敏感所带来的阈值计算方法难以确定的不足,而且跟以往研究方法相比,滤波过程运行时间缩短、滤除噪声效果更好。其次,针对肌电信号中偶尔混入的工频干扰,采用基于独立成分分析方法和小波滤波技术相结合的方法,可以把工频成分有效滤除,并克服了以往方法中引入时移和总是需要输入一个可能与实际工频信号不完全吻合的参考输入信号的缺陷。第二,在对MUAP波形进行特征提取过程中,本文提出了基于原始数据和形态特征参数相结合的方法、小波系数和形态特征参数相结合的方法,以及基于线性判别分析和模糊集合的面向分类目的的最优小波包系数特征提取方法。跟现有基于原始信号数据特征方法、波形形态特征方法、小波系数特征方法相比,在保持了其良好的分解效果和运行速度的情况下,小波包系数特征方法明显具有占用内存小的特点。第三,在MUAP波形的分类过程中,本文在借鉴了传统多次运用单链聚类技术和最小分类器对MUAP波形进行聚类和有监督分类方法结果的可靠性和可检验性,采用了基于模式对所有类别隶属度的模糊C均值聚类技术,对波形聚类和有监督分类方法进行了优化。结果表明,比现有没有进行分类优化的肌电信号分类方法具有更好的分类效果,有效提高了识别正确率。第四,在对肌电信号分解结果的定量检验过程中,本文按照符合肌电信号的生理特点的大小原理和发放模式等规则的信号模拟方法来分配发放模式,与现有的随机分配各个MU的发放模式相比,使得模拟的信号更加符合肌电信号的生理特性。其次,采用先进的模式识别和信号处理技术,设计开发了一种新的手工分解肌电信号新方法,克服了传统的利用目测和手工测量的方法中客观性不好、准确度不高的缺点,得到更为理想的分解结果,并将此结果作为本文肌电信号自动分解结果的检验标准。虽然本论文基于肌电信号分解基本步骤,提出并实现了一系列改进措施,对模拟和真实肌电信号进行了分解,获得了较好的肌电信号分解结果,但是,通过分析本文分解结果发现,分解方法还存在不完善和不完整之处,需进一步改进和完善,对此在本论文最后一章的展望部分提出了一些初步想法。更多还原

【Abstract】 A myoelectric signal (MES) detected with needle electrodes is a sum of the motor unit action potentials (MUAPs) trains of all recruited motor units (MUs) and additive noise, including background instrumental noise, Gaussian white noise, and sometimes power line interference (PLI). Electromyograph (EMG) decomposition is the process of resolving a composite MES into its constituent MUAP trains. Diognostic application and the commercial success of these techniques have been lagging behind despite the enthusiasm of researchers and the number, quality and significance of scientific publication. Decomposition algorithm are being developed and published.The presented algorithm for EMG decomposition includes the following steps: 1) Removing noises; 2) MUAPs extraction and EMG signal segmentation ; 3) MUAP clustering; 4) MUAP classification. Our work mainly includes the following points:Firstly, removing noise and effective MUAP peak detection is the first important step in EMG decomposition. We first combined independent component analysis and wavelet filtering to remove power line interference, and then applied a wavelet filtering method and threshold estimation calculated using wavelet transform to suppress background noise and Gaussian white noise. In contrast to existing methods based on amplitude single-threshold filtering of the original myoelectric signal or a conventional digitally filtered signal, our technique is fast and robust.Secondly, in this thesis we first take the original data, the morphological feature data and the wavelet coefficients at third to sixth level after wavelet transform at sixth level as feature space. Otherwise, we utilized the wavelet packet coefficients of the optimal decomposition based on linear discriminative analysis and fuzzy clustering. Among all these feature extraction methods, the wavelet coefficients is the most consuming, and the optimal wavelet packet coefficients is the most effective in reducing the dimensionality.Thirdly, we arranged these single MUAPs to their constituent MUAPTs based on the single-linkage hierarchical clustering algorithm. The supervised classifier we used is based on the minimum distance classifier to classify non-overlapping but un-classified active segments by clustering. Morever, we used the fuzzy K-means clustering to optimize the classification program and to improve the classification accuracy.Finally, a very important aspect in this field is the evaluation of performance. In order to obtain a reference decomposition result, different methods have been proposed: 1) synthetic signals; 2) real signal decomposed manually; 3) recordings from the same MU at different locations and the results compared. Our methods are focused on the former two methods. The EMG signals were generated on the basis of a model of a normal intramuscular EMG recording, proposed by the former researcher. All of our artificial recordings were corrupted with random white noise at various signal-to-noise ratios (SNR) and with a PLI signal. The generation of the firing pattern was based on three statistical characteristics of the pattern: regular firing, double-discharge firing, and random firing. A regular firing pattern was introduced by use of the mean inter-pulse interval (IPI), which follows a size principle. Otherwise, we developed a new EMG manually decomposition program based on the new pattern classification and digital signal processing technique.The technique of our EMG signal decomposition is fast and robust, which has been evaluated through synthetic EMG signals and real EMG signals. Because the speed of our EMG decomposition program was improved greatly and the performance of our method is very robust our technique may be suitable for on-line analysis. Therefore, our decomposition technique only clustered and classified single MUAPs. Certainly, the decomposition results are incomplete because of the restriction of time and the limitation of our data collection. Currently, in order to finish the complete EMG decomposition, we’re going to do the further research and investigation related to new EMG signal decomposition method. The next study will include but nor limited points as follows: (1) extracting new effective MUAP features such as non-linear feature parameters; (2) trying to finish complete EMG signal decomposition through resolving superimposed action potentials.更多还原