特征提取对通道选择方法的影响研究

Research on the influence of feature extraction on channel selection method

通道选择可以有效地去除冗余信息,减少特征维数,避免维数灾难和过拟合,因此在运动想象脑电(EEG)信号解码中非常重要。现有的文献主要研究通道选择方法的改进,忽略了不同特征对通道选择方法的影响。主要研究特征提取对通道选择方法的影响。首先,对预处理之后的EEG信号提取方差、自回归(AR)系数、带通功率和小波包能量4种特征,研究单一特征中哪个特征对通道选择方法最有效。另外,计算4种特征的融合特征,研究单一特征和融合特征哪个对通道选择最有效。采用一个公开的脑机接口(BCI)竞赛数据集进行实验,研究不同特征提取方法在Fisher判别准则(FDC)、基于支持向量机的递归通道剔除(SVM-RCE)、最小绝对值收缩和选择算子(LASSO)和组LASSO(gLASSO)4种通道选择方法中的分类结果。实验结果表明,在单一特征中,小波包能量获得了较好的分类结果,其中在SVM-RCE通道选择方法中获得了76.15%的最高平均分类准确率。融合特征的分类结果均优于单一特征,其中在gLASSO通道选择方法中获得了78.6%的最高平均分类准确率。融合特征更能表征复杂的脑电成分,形成信息互补,对脑电任务的分类识别更有效。

Channel selection can effectively remove redundant information,reduce feature dimensionality,avoid dimensionality disasters and overfitting.So it is very important in the decoding of motor imagery electroencephalogram(EEG)signals.The existing literature mainly studies the improvement of the channel selection method,ignoring the influence of different features on the channel selection method.This paper mainly studies the influence of feature extraction on channel selection methods.First,the four features of variance,autoregressive(AR)coefficient,band power and wavelet packet energy from the preprocessed EEG signal are extracted,and then we study which of the single features is the most effective for the channel selection method.In addition,the fusion feature of the four features is calculated,and then we study whether the single feature or the fusion feature is the most effective for channel selection.A public brain-computer interface(BCI)competition data set is used to conduct experiments to study the classification results of different feature extraction methods with four channel selection methods of Fisher discriminant criteria(FDC),support vector machinebased recursive channel elimination(SVM-RCE),least absolute value shrinkage and selection(LASSO)and group LASSO(gLASSO).The experimental results show that,for a single feature,the wavelet packet energy obtains a better classification result,and the highest average classification accuracy of 76.15%is obtained in the SVM-RCE channel selection method.The classification results of the fusion features are better than the single feature,and the highest average classification accuracy rate of 78.6%is obtained in the gLASSO channel selection method.Fusion features can better characterize the complex EEG components,form complementary information,and are more effective in classifying and identifying EEG tasks.