多种混杂因素下鲁棒式肌电模式识别方法研究

发布时间：2018-05-02 22:08

本文选题：肌电信号 + 模式识别　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：肌电控制受到诸多因素的干扰,如电极位置窜动、手臂姿势变化、肌肉收缩力变化、个体差异、长期时变等,从而导致实际应用中肌电控制的成功率较低。针对上述干扰因素,本文分别从特征提取方法、分类器泛化能力、自适应学习策略等方面进行研究。主要内容包括:基于粒子群优化(Particle Swarm Optimization,PSO)算法的特征阈值优化方法,基于离散傅里叶变换(Discrete Fourier Transform,DFT)、小波变换(Wavelet Transform,WT)及小波包变换(Wavelet Packet Transform,WPT)的特征提取方法,基于支持向量机(Support Vector Machine,SVM)核函数的学习策略,基于代表样本更新的在线无监督学习策略等。本文首先综述了国内外肌电模式识别的研究现状,发现了目前的研究所存在的一些问题,并确定本文的主要研究内容。为了减少电极位置窜动、手臂姿势及肌肉收缩力变化等混杂因素的干扰,本文首先从肌电模式特征提取方面进行研究。提出一种基于PSO算法的特征阈值优化方法,相比于传统的基于经验选择的方法,简化了过零点数(Zero Crossing,ZC)、脉冲百分率(Myopulse Percentage Rate,MYOP)、Willison幅值(Willison Amplitude,WAMP)、斜率符号变化(Slope Sign Change,SSC)等特征的参数选择过程,识别正确率平均提升10.2%;此外,本文提出将绝对均值(Mean Absolute Value,MAV)、均方根(Root Mean Square,RMS)等传统常用特征与DFT、WT、WPT相结合的复合式特征提取方法,该方法能够明显提高肌电模式识别的鲁棒性,分别将识别正确率提升30.5%、25.4%、22.9%。针对优势手/非优势手互换、手臂姿势及肌肉收缩力变化等混杂因素的干扰,本文从提升分类器的泛化能力方面进行研究。首先引入概率神经网络(Probabilistic Neural Networks,PNN)作为肌电模式识别的分类器,发现其泛化能力比线性判别分析分类器(Linear Discriminant Analysis,LDA)更强。然后研究了SVM的核函数,提出一种多核学习的方式,以提升SVM的泛化能力。实验证明基于高斯核的多尺度核函数能够取得最高的模式识别成功率,相比于高斯核,成功率平均提升了1.5%。针对长期时变、手臂姿势变化等混杂因素导致的模式识别成功率下降问题,本文提出一种基于代表样本的在线学习策略,能够从训练集中选择最能代表类别信息的样本。实验证明该方法不仅能够缓解电极长期佩戴过程中肌电模式识别成功率的退化,也能提升电极位置窜动、肌肉收缩力变化等更复杂因素干扰下的识别成功率。
[Abstract]:The EMG control is disturbed by many factors, such as the movement of the electrode position, the change of arm posture, the change of muscle contractile force, the individual difference, the long time change and so on, which leads to the low success rate of the electromyography control in the practical application. The main contents include: the feature threshold optimization method based on Particle Swarm Optimization (PSO) algorithm, the feature extraction method based on discrete Fourier transform (Discrete Fourier Transform, DFT), wavelet transform (Wavelet Transform, WT) and wavelet packet transform (Wavelet), based on support The learning strategy of the Support Vector Machine (SVM) kernel function is based on the online unsupervised learning strategy, which represents the update of the sample. This paper first summarizes the research status of the EMG pattern recognition at home and abroad, and finds some problems in the present research, and determines the main contents of this paper. In this paper, a new method of feature threshold optimization based on PSO algorithm is proposed. Compared with the traditional method based on experiential selection, the number of zero crossing points (Zero Crossing, ZC) and pulse percentage (Myopulse Percentage) are simplified. Rate, MYOP), Willison amplitude (Willison Amplitude, WAMP), slope symbol change (Slope Sign Change, SSC) and other characteristics of the parameter selection process, the recognition accuracy is improved by an average of 10.2%. Furthermore, this paper puts forward the combination of the traditional common features such as absolute mean (Mean Absolute), mean square root and other common features. Combined feature extraction method, this method can obviously improve the robustness of EMG pattern recognition. The recognition accuracy is increased by 30.5%, 25.4%, 22.9%. for the interference of the mixed factors such as the hand / non dominant hand exchange, the arm posture and the changes of the muscle contractile force. This paper first introduces the generalization ability of the lifting classifier. Probabilistic Neural Networks (PNN), as a classifier for EMG pattern recognition, finds that its generalization ability is stronger than that of linear discriminant analysis classifier (Linear Discriminant Analysis, LDA). Then, the kernel function of SVM is studied and a multi kernel learning method is proposed to improve the generalization ability of SVM. The experiment is based on Gauss. The kernel's multi-scale kernel function can achieve the highest success rate of pattern recognition. Compared with the Gauss kernel, the success rate increases the success rate of pattern recognition in 1.5%. for long time variation and arm posture change. This paper proposes an online learning strategy based on representative sample, which can choose the most from the training center. The experiment shows that the method can not only alleviate the degradation of the success rate of the electromyographic pattern recognition during the long-term wear of the electrode, but also improve the recognition success rate under the interference of more complex factors such as the change of the electrode position and the changes of the muscle contractile force.

【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R496;TP391.4

【参考文献】