APP下载

Research of Fusion Classification of EEG Features for Multi-Class Motor Imagery*

2016-09-09ZHANGHuanQIAOXiaoyanCollegeofPhysicsandElectronicsEngineeringShanxiUniversityTaiyuan030006China

传感技术学报 2016年6期
关键词:特征提取想象分类

ZHANG Huan,QIAO Xiaoyan(College of Physics and Electronics Engineering,Shanxi University,Taiyuan 030006,China)



Research of Fusion Classification of EEG Features for Multi-Class Motor Imagery*

ZHANG Huan,QIAO Xiaoyan*
(College of Physics and Electronics Engineering,Shanxi University,Taiyuan 030006,China)

In view of the problems of pattern simplification,low accuracy of classification and poor practicability in motor imagery BCI,they improve feature extraction method to common spatial pattern(CSP),and the support vector machine(SVM)is used to carry out multi-class classification,combining with the CSP to classify the feature signal of EEG.Firstly,they select EEG signal in the specific channel to do wavelet decomposition and reconstruction,in or⁃der to remove redundant information;Secondly,they improve the method,by doing subtractions between different characteristic parameters,and obtain obvious characteristics of EEG;Finally,the SVM is used to carry out multitask classification,combining with the CSP to classify the feature signal of EEG.Using BCI competition data,the four kinds of motor imagery tasks of left hand,right hand,tongue and feet are identified based on EEG signals.Ex⁃perimental results show that the correct rate of classifying is 90.9%for maximum,the average accuracy rate is 86.4%,the Kappa coefficient is 0.8867,and the information transmission rate was 0.68bit/trial and the method can extract EEG features effectively and achieve better classification to a multi-task motor imagery of EEG signals.

BCI;motor imagery;feature extraction and classification;wavelet transform;common spatial pattern;support vector machine

脑-机接口(BCI)是在人脑和计算机或其他电子设备之间建立不依赖于常规大脑信息输出通路(外周神经和肌肉组织)的全新对外信息交流和控制技术[1]。BCI技术在助残及康复工程、正常人辅助控制、娱乐等领域有着广泛的应用前景,正受到世界范围内更多的关注和研究。

对BCI中的运动想象脑电信号进行模式识别,国内外已进行了较多研究,但目前对于四模式运动想象脑电识别还存在分类正确率较低且不稳定,所用导联数目多,算法复杂导致实际应用不理想等。Ghaheri H等人[2]采用对每个不同的脑电信号时间段进行共空间模式(CSP)的特征提取以及LDA线性分类器进行分类,可达到80.0%的分类正确率,由于时域分析时间开销大,同时使用了22个脑电导联,增加了应用的复杂性。Luis F Nicolas-Alonso等人[3]对22个脑电导联信号滤波后,用CSP提取特征结合谱回归核判别分析对四种任务运动想象的脑电分类,最好的分类正确率为94%,Kappa值为0.92,但对应的平均正确率却只有73%,Kappa值为0.64。虽然最高的正确率较好,但平均正确率较低,使得实际应用缺乏稳定性。天津大学万柏坤等人[4]采用二维时频分析结合Fisher分析的方法特征提取,使用支持向量机分类处理四种不同肢体部位动作识别,识别率达到85.71%,实验中使用了60个导联上的数据,操作复杂可使用性不强。王瑞敏等人[5]用短时傅里叶变换分解转变成多频段的时频信号,然后采用CSP结合支持向量机(SVM)方法对单个导联上的运动想象脑电信号分类识别,最佳识别准确率为88%,平均准确率只有65%,且在减少导联时,识别正确率大幅下降。……

登录APP查看全文

猜你喜欢

特征提取想象分类
快乐的想象
分类算一算
细观察 多想象 善表达
基于Gazebo仿真环境的ORB特征提取与比对的研究
这些并不是想象,有些人正在为拥抱付费
教你一招:数的分类
那时我们如何想象未来
一种基于LBP 特征提取和稀疏表示的肝病识别算法
基于MED和循环域解调的多故障特征提取
Walsh变换在滚动轴承早期故障特征提取中的应用