脑机接口中一种多类运动想象任务识别新方法

A novel recognition method of multi-class motor imagery tasks in brain computer interfaces

目的针对脑机接口中三类运动想象任务,提出一种最小二乘法自适应滤波结合独立成分分析以及样本熵(RLS-ICA-Samp En)、多类共同空间模式(CSP)、增量式支持向量机(ISVM)相结合的脑电识别新方法,以解决脑机接口中多类运动想象正确率低的问题。方法首先采用ICA将EEG分离,然后利用样本熵自动识别分离后的噪声,再采用RLS对识别出来的噪声进行滤波,最后进行信号重构,得到去除噪声的脑电信号。多类CSP采用"一对一"CSP与多频段滤波相结合,对去噪后的脑电信号进行特征提取。通过"一对多"方式的ISVM对三类运动想象脑电信号获取的特征向量进行分类。为检验新方法的有效性,将本文方法与多类CSP+ISVM(方法 1)及RLS-ICA+多类CSP+ISVM(方法 2)进行比较。结果对三类想象任务而言,本文方法识别正确率与方法 1和2相比均高8%左右。结论与方法1和2比较,RLS-ICA-Samp En、多类CSP、ISVM相结合的脑电识别新方法能更好地适用于多类运动想象任务识别。

Objective For multi-class motor imagery tasks in brain computer interface （ BCI ） , this paper presents a novel recognition method of electroencephalography （ EEG） by combining RLS-ICA-SampEn [ RLS （ recursive least-squares ） , ICA （ independent component analysis ） , SampEn （ sample entropy ） ] , multi-class CSP （common spatial patterns） and ISVM （incremental support vector machine ）.Methods In the RLS-ICA-SampEn, Firstly, the ICA is used to decompose the contaminated EEG signals into independent components （IC）.Then, the sample entropy is used to automatically identify the noise signal in the IC .Next, the RLS adaptive filters are applied to the identified noise in IC to remove noise further .Finally, the processed ICs are then projected back to reconstruct the noise-free EEG signals.The RLS-ICA-SampEn is used to preprocess EEG signals to get pure EEG signals , in which some noise signals can be removed .The multi-class CSP combines the CSP and the multi-band filtering technology , in which the CSP uses the ‘one versus one ’ strategy.The multi-class CSP is used to extract featuresfor pure EEG signals.The obtained features are input tothe ISVM for classification.The ‘one versus rest’ strategyis applied to classify three-class EEG signals.In order toverify the effectiveness of the proposed novel method , it iscompared with other two methods including multi CSP ＋ISVM（method 1）, RLS-ICA ＋multi CSP ＋ISVM（method 2）.Results The result shows that the recognitionaccuracy obtained by the proposed method is higher about 8% than other two methods.Conclusions Comparedwith method 1 and 2, the proposed method is better suited for the recognition of multi -class motor imagery tasksin BCI.