脑电信号的特征提取和分类研究

发布时间：2018-09-16 20:41

【摘要】：脑电信号是亿万个神经元活动在大脑皮层的综合反映，含有丰富的有用信息，并且受内分泌系统和神经系统支配，，能够更客观的反映人的生理状态。近年来，脑-机接口的技术也逐渐被人们熟知，而脑-机接口依赖于脑电信号的模式识别，所以脑电信号的识别研究也受到了国内外科学工作者们的广泛关注。由于脑电信号是一种非平稳并且产生机理非常复杂的随机信号，所以，如何更有效地提取脑电信号中有用信息并进行分类，对于脑电信号的识别研究非常的重要。对脑电信号的识别研究一般分为两方面：分别是主动执行任务产生的脑电和被动接收刺激产生的诱发脑电的信号识别研究。本文对这两方面的脑电的识别都进行了研究，具体的工作如下：在主动执行任务产生的脑电信号识别方面，本文对面向运动想象的脑电识别进行了研究。针对常用的主成分分析(Principle Component Analysis, PCA)脑电特征提取方法没有有效利用样本的类别信息，因而对分类贡献低的问题，本论文采用主成分分析和线性判别分析(Liner Discriminate Analysis,LDA)相结合方法进行特征提取，并将提取的特征送入基于投票机制的极速学习机(Voting based extreme learning machine, V-ELM)分类器中。在国际脑-机接口标准数据集BCI2003dataset Ia上的实验结果显示：本文采用的特征提取方法与分类算法的识别率达到了93.52%，该结果与目前在本数据集上最高的识别率92.15%相比，有明显提升。采用结合主成分分析与线性判别分析的特征提取方法，与脑电识别领域使用最广泛的支持向量机分类算法相比，V-ELM的识别率提升了1.43%，且降低了其大约90%的耗时。在被动接收刺激产生的诱发脑电信号识别方面，本文对不同类型视频节目诱发的情感脑电信号识别问题进行了研究。论文第四章给出了实验设计和脑电信号采集过程；并依据心理学对视听刺激所影响的脑区方面的研究成果，在64个电极中选取了与视听相关的FP1、FP2、P1、P2共4个电极进行分析；接着，根据情感脑电本身携带的有效信息主要集中在和β波所属频率范围，对原始脑电信号进行了滤波；最后，将小波包方差作为特征提取方法、支持向量机作为分类方法进行视频诱发的脑电识别。实验结果表明：本文采用的特征提取和分类方法是有效的，两类平均识别率在71.44%~82.94%，三类识别率达到了66.07%。
[Abstract]:EEG is a comprehensive reflection of the activity of hundreds of millions of neurons in the cerebral cortex. It contains abundant useful information and is controlled by the endocrine system and nervous system, which can reflect more objectively the physiological state of human beings. In recent years, the brain-computer interface technology is also gradually known, and the brain-computer interface depends on the pattern recognition of EEG, so the research of EEG recognition has been widely concerned by scientists at home and abroad. Because EEG signal is a kind of non-stationary and complex random signal, it is very important to study how to extract useful information from EEG signal and classify it more effectively. The recognition of EEG signals is generally divided into two aspects: one is the EEG generated by active tasks and the other is the signal recognition of evoked EEG produced by passive receiving stimuli. In this paper, the two aspects of EEG recognition are studied, the specific work is as follows: in the field of EEG recognition generated by active tasks, this paper studies EEG recognition oriented to motion imagination. In view of the problem that common principal component analysis (PCA) (Principle Component Analysis, PCA) EEG feature extraction methods do not make effective use of the classification information of the samples, so they have a low contribution to the classification. In this paper, the principal component analysis (PCA) and linear discriminant analysis (Liner Discriminate Analysis,LDA) are used to extract the features, and the extracted features are fed into the (Voting based extreme learning machine, V-ELM classifier based on the voting mechanism. The experimental results on BCI2003dataset Ia, the standard data set of BCI, show that the recognition rate of the feature extraction method and the classification algorithm is 93.52, which is obviously higher than the highest recognition rate of 92.15% in the present data set. By combining principal component analysis (PCA) with linear discriminant analysis (LDA), the recognition rate of VELM is 1.43% higher than that of the most widely used SVM classification algorithm in EEG recognition field, and the time consuming is reduced by 90%. In the aspect of the recognition of evoked EEG signals generated by passive receiving stimuli, the recognition of emotional EEG signals induced by different types of video programs is studied in this paper. The fourth chapter gives the experimental design and EEG acquisition process, and according to the psychological research on the brain area affected by audio-visual stimulation, we select four FP1,FP2,P1,P2 electrodes from 64 electrodes to analyze. According to the effective information carried by emotional EEG itself mainly in the frequency range of 尾 wave, the original EEG signal is filtered. Finally, wavelet packet variance is used as feature extraction method. Support vector machine (SVM) is used as a classification method for video induced EEG recognition. The experimental results show that the feature extraction and classification methods used in this paper are effective. The average recognition rate of the two categories is 71.44 and 82.940.The recognition rate of the three types reaches 66.07.
【学位授予单位】：北京工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.7

【相似文献】