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

CSP Filter Calculation of Single Training Pairs and Its Application in EEG Training Set Optimization

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%。

In the research of motor imagery brain-computer interface （MI-BCI） system, common spatial pattern （ CSP）, as a supervised spatial filtering method, has been widely used to extract the feature of motor imagery electroencephalogram （EEG）. But in the process of training data acquisition, the EEG signals may be influenced by different kind of interfer- ences as well as subjects＇ distraction and fatigue, thus abnormal training trials with low quality will be generated inevitably. If all training trials are used indiscriminately for CSP filter calculation and classifier training, the performance of BCI system may be badly degraded. To resolve this problem, this paper proposes a novel and practical method for the training trial se- lection. The basic steps of the method are, firstly, a pair of two single-trials （one left and one right）, which are selected from training dataset in turn, are used to calculate the single-trial-based CSP filters （s CSP）. Then, each s_CSP com- bined with the zero training classifier is employed to construct the single-trial-based CSP-BCI system, which will be used for self testing on training dataset, and the recognition rates of self testing are used as the index to assess the quality of the single trials for the purpose of bad trial rejection and training dataset optimization. Seventy-five datasets of two-class MIEEG from six subjects collected during different days are used to test the proposed method. Experimental results reveal that the performance of CSP-BC! system is improved significantly compared to the traditional method. The average recognition rate of the six subjects＇ testing sets are increased by 5. 04% , 6.42%, 13.15%, 15. 51%, 1.94% and 8.26% respectively.