脑电信号特征的归一化方式与选择方法研究

RESEARCH ON NORMALISATION MANNER AND SELECTION METHOD FOR EEG FEATURES

针对多被试脑电数据存在被试间特征值差异较大的问题,分析了单次归一化的数据范围对分类准确率的影响。实验在情感数据集上采用6种常用的归一化方法,对所有被试的特征、单个被试的所有特征、单个被试的单个属性特征这三种单次归一化的数据范围进行准确率上的比较,证明了单个被试的单个属性特征更适于作为多被试脑电数据单次归一化数据范围。此外,提出方差贡献率与F-score结合的特征选择方法,在不降低准确率的情况下大量减少了特征数量。小波包树结点能量作为变换最少的特征得到的分类结果最好,小波包熵比脑电节律小波熵的分类准确率高。

Electroencephalographic （EEG） feature values of multiple subjects vary greatly between individual subjects. In light of this, we analyse the influence of data range of single normalisation on classification accuracy. This experiment uses six kinds of common normalisation methods on emotion dataset to compare the accuracies from three kinds of data range in regard to single normalisation. They are the features of all subjects, all features of single subject and the single property feature of single subject. The result proves that the single feature property of single subject is more appropriate as the data range of single normalisation of multiple subjects＇ EEG data. In addition, we present the feature selection method which combines the variance contribution rate with F-score, it brings a significant reduction in the number of features without reducing the classification accuracies. Wavelet packet tree node energy gets the best classification accuracies as the changed lest feature. The wavelet packet entropy has higher classification accuracies than the EEG rhythm wavelet entropy.