【作者】 李涛；

【导师】 胡春海； 陈琛；

【作者基本信息】 燕山大学 ， 工程硕士（专业学位）， 2017， 硕士

【摘要】 脑机接口是一种脑和外界进行信息交互的通信接口,在越来越多的领域得到了广泛的应用,对其相关算法的研究已经成为重要的研究课题。运动想象脑电信号是一种自发性的信号,本文将运动想象脑电信号作为研究对象,研究了局部均值分解和差分进化算法的不足之处及改进方法,并研究了运动想象脑电信号的特征提取识别和最优频带选择相关算法。首先,分析了局部均值分解中存在端点效应问题的原因,用自延拓的方法进行改善。通过仿真实验验证,说明了该方法能够有效的减小端点问题带来的影响。针对微弱脑电信号难以提取到有效特征的问题,提出多种熵值的运动想象脑电信号特征提取方法。采用改进的局部均值对运动想象脑电信号分解,得到调幅-调频乘积函数分量,计算有效PF分量的能量熵、模糊熵、多尺度熵。以三种熵值作为特征向量,最后使用学习矢量量化神经网络进行模式识别,实现对想象运动的分类。通过仿真实验研究,取得了较好的分类识别率,说明了该方法能够很好的提取出微弱的运动想象脑电信号特征。然后,针对差分进化算法容易出现过早收敛的问题、以及难以设置最佳参数的问题,提出了基于凸二次函数的多策略变异算子的变异方法和simgod函数的交叉因子取值方法。通过实验研究分析,验证了提出方法的有效性。最后,针对个体差异难以自适应的确定个体最优频带的问题,采用了改进的差分进化算法进行自适应的频带优选。在频带选择系统中,使用共空间模式提取脑电特征,用线性判别分析计算分类正确率作为为适应度,再由差分进化算法搜索最佳频带。使用3个数据集进行了实验,实验结果表明,该方法快速有效的选择出了最优频带,且该方法还适用于最优时间选择和最佳通道组合优选。更多还原

【Abstract】 Brain computer interface is a kind of interactive mode of information between brain and the outside world,which has been widely used in more and more fields.Research on the related algorithm has become an important topic.Motor-imagery EEG is a spontaneous signal.By taking a typical motor imagery EEG as the research object,this paper analyzed the shortcomings and improved methods of local mean decomposition and differential evolution.Then researched the feature extraction and recognition algorithms,optimal band selection algorithms also studied.Firstly,we analyzes the reason of the end effect problem in the local mean decomposition,and improves the method by using the method of self extension.The simulation results show that the proposed method can effectively reduce the impact of end effect.Aiming at the problem that the weak EEG signal is difficult to extract the effective features,this paper proposes a method of feature extraction based on multi entropy.The method processes motor imagery EEG with local mean decomposition,then obtained product function components.Calculating of the effective PF component’s fuzzy entropy,multi-scale entropy,energy entropy.Three kinds of entropy as feature vector,and finally using learning vector quantization neural network for pattern recognition to realize the classification of motor imagery.The simulation experiment shows a better classification recognition rate,which proves that the method can extract weak motor-imagery EEG features.Secondly,the differential evolution algorithm may easily appear premature convergence,and difficultly set the best parameters.To resolve the above two problems,we proposed multi strategy method for mutation operator based on convex quadratic function and crossover exchange factor based on simgod function.By experimental analysis,we can verify the effectiveness of the proposed method.Finally,we used the improved differential evolution algorithm to obtain the optimize the frequency band to resolve the problem that the individual difference is difficult to self-adaption determine the optimal frequency band.In the frequency selection system,theEEG feature is extracted by the common spatial pattern,the classification accuracy is calculated by linear discriminant analysis,and use the differential evolution algorithm to search the best band.Experiments were carried out by using 3 datasets,and the results show that the proposed method is efficient and effective to select the optimal frequency band.And the method is also suitable for the optimal time selection and the optimal channel combination selection.更多还原