单次样本对的CSP滤波器设计及其在脑电训练样本优化中的应用

发布时间：2018-05-18 02:11

本文选题：脑-机接口 + 共空间模式　；参考：《信号处理》2017年07期

【摘要】：在运动想象脑-机接口(Motor imagery brain-computer interface,MI-BCI)系统研究中,共空间模式(Common spatial pattern,CSP)作为一种有监督空域滤波设计方法,已被广泛应用于运动想象脑电信号(Electroencephalography,EEG)的特征提取。但是EEG训练样本的采集过程不仅会受到各种噪声伪迹干扰,也会受到受试者分心和疲劳等因素的影响,因此,训练集中难免出现"低质量"的异常单次试验数据。如果不加选择地将所有的单次样本用于CSP滤波器设计和分类器训练,会给所建BCI系统的性能带来较严重的负面影响。针对这一问题,本文提出一种新颖而实用的EEG训练样本筛选方法。方法的基本步骤是,先依次选择单次EEG样本对进行CSP滤波器设计,并结合零训练分类器构造相应的CSP-BCI测试系统。然后以所建CSP-BCI系统的交叉验证识别率为指标,剔除低识别率对应的单次训练数据,以实现对训练样本集的优化。基于所提方法,论文对6位受试者在不同时间采集的75组两类运动想象EEG数据进行了优化筛选和测试。实验结果表明,相比传统方法设计的CSP-BCI系统,基于训练样本优化方法的CSP-BCI系统性能得到明显改善,针对六位受试者测试集的平均识别率分别提高了5.04%、6.42%、13.15%、15.51%、1.94%和8.26%。
[Abstract]:In the research of motor imagery brain-computer interface MI-BCII system, Common spatial pattern (CSP), as a supervised spatial filtering method, has been widely used in feature extraction of motor image-EEG electroencephalography (EGG). However, the acquisition process of EEG training samples is affected not only by noise artifacts, but also by the distractions and fatigue of the subjects. Therefore, "low quality" abnormal single test data are inevitable in the training concentration. If all the single samples are used in the CSP filter design and classifier training, the performance of the BCI system will be seriously affected. To solve this problem, a novel and practical EEG training sample selection method is proposed in this paper. The basic steps of the method are to select a single EEG sample pair in turn to design the CSP filter, and to construct the corresponding CSP-BCI test system combined with the zero training classifier. Then, the single training data corresponding to the low recognition rate is eliminated based on the cross validation recognition rate of the CSP-BCI system, so as to optimize the training sample set. Based on the proposed method, 75 groups of motion imagination EEG data collected by 6 subjects at different time were optimized and tested. The experimental results show that compared with the traditional CSP-BCI system, the performance of the CSP-BCI system based on the training sample optimization method is obviously improved. The average recognition rate of the test set for the six subjects is increased by 5.04% and 6.42%, respectively, by 1.94% and 8.26%, respectively.
【作者单位】：安徽大学计算机科学与技术学院;安徽大学计算机学院;安徽大学计算智能与信号处理教育部重点实验室;安徽大学信息保障技术协同创新中心;
【基金】：国家自然科学基金(6127132,61401002) 安徽省自然科学基金(1408085QF125) 安徽省自然科学研究项目(KJ2016A043)
【分类号】：R318;TN713

【相似文献】