多类运动想象脑电信号特征提取与分类

Feature Extraction and Classification of Multi-class Motor Imagery EEG Data

摘要:

针对多类运动想象情况下存在的脑电信号识别正确率比较低的问题,提出了一种基于小波包特定频段的小波包方差,小波包熵和共同空间模式相结合的脑电信号特征提取的方法,并将特征向量输入到支持向量机中达到分类的目的；首先选择重要导联的脑电信号,进行特定频段的小波包去噪和分解；其次对通道优化的重要导联的每个通道信号计算小波包方差和小波包熵值作为特征向量；然后对所有重要导联的分解系数重构并进行共同空间模式特征提取；最后结合2种不同导联方式所获取的特征向量作为分类器的输入进行分类；采用BCI2005desc_IIIa中l1b数据进行验证,该算法的分类正确率最高达到88.75%,相对2种单一的提取方法分别提高28.27%和6.55%；结果表明该算法能够有效提取特征向量,进而改善多类识别正确率较低的问题。

关键词:脑电信号;小波包方差;小波包熵;共同空间模式;特征提取;支持向量机

Abstract:

Due to EEG recognition accuracy was relatively low in the case of multi-class problem of motor imagery,this paper presents a method that a new combination about wavelet packet variance (WPV), wavelet EEG feature package entropy (wavelet packet entropy, WPE) and common spatial patterns (CSP) extract features based on wavelet packet specific frequency bands, which input into support vector machine(SVM ) classifier achieve resultant classification. Firstly, selecting the EEG of important channels make wavelet packet de-noising and decomposition(wavelet packet decomposition,WPD) of specific frequency bands; Secondly, optimization of important channels calculate the wavelet packet variance (WPV) and wavelet packet entropy(WPE) as feature vectors; then, the three sub-band coefficients for each channel EEG signal of important channels are reconstructed and feature extraction carried by common space pattern(CSP); Finally, two kinds of feature vectors from different ways that feed into a classifier and achieve classification. The highest classification accuracy rate of 88.75%, comparing with the relative two kinds of single-extraction method increased 28.27% and 6.55% by l1b from BCI2005desc_IIIa.The results show that the algorithm can effectively extract the feature vectors, thereby improving the lower classification accuracy problems.

Key words:EEG; wavelet packet variance(WPV); wavelet packet entropy(WPE); common space pattern(CSP); feature extraction; support vector machine(SVM)