分层向量自回归的多通道脑电信号的特征提取研究

有效的特征提取方法能提高脑机接口（Brain-computer interface，BCI）系统对脑电（Electroencephalogram，EEG）信号的识别率.因脑电信号都是多通道的，本文将分层向量自回归（Hierarchical vector autoregression，HVAR）模型用于脑电信号的特征提取，并结合传统的线性支持向量机（Support vector machine，SVM）用于脑电信号识别.该模型不仅克服了自回归（Autoregression，AR）模型只能用来提取单通道特征的局限性，而且不再采用传统VAR（Vector autoregression）模型所有通道共用一个时滞的处理方法.创新之处在于在传统的VAR模型基础上添加正则化思想，有效地压缩参数空间，实现合理的分层结构.本文首次将HVAR模型用于由Keirn等采集并整理的脑电数据中.实验结果证明HVAR模型在阶数较小的情况下（2阶）与阶数较大（6阶）的AR模型效果相当，可见低阶的HVAR能很好地刻画脑电信号的时空关联关系，这说明HVAR可能是刻画EEG信号的一种新颖的方法，这对其他多通道时间序列分析都有借鉴意义.

Multi-channel EEG Feature Extraction Using Hierarchical Vector Autoregression

Feature extraction and classification of electroencephalogram (EEG) signals is a core part of brain-computer interface (BCI). For multi-channel EEG signal and high dimension of feature vector of BCI system, a novel EEG signal recognition method called hierarchical vector autoregression (HVAR) is presented, which extracts EEG feature using regression coefficient of HVAR model and linear support vector machine (SVM). It overcomes the limitations of the autoregression (AR) model that can be used to extract the single channel EEG only, and effectively avoids the vector autoregression (VAR) model sharing a same delay for all channels. Our contribution is that regularization is added on the traditional VAR model and a reasonable hierarchical structure is adopted. It effectively compresses parameter space of VAR model. In this paper, HVAR model is used for EEG data classification for the first time. Experimental results show that the recognition accuracy of extracted feature of HVAR model using a 2 lag order multi-channel is higher than that of AR model of 6 lag order. So low-level HVAR model can describe the portrayed temporal relationship of EEG well. This shows HVAR may be a novel method to portray EEG signal, which has reference significance to other multi-channel time-series.