【作者】 魏庆国；

【导师】 高上凯；

【作者基本信息】 清华大学 ， 生物医学工程， 2006， 博士

【摘要】 脑-机接口(BCI)的分类性能取决于用户控制自己脑状态的能力、数据记录方法和分类算法。本文使用头皮脑电图(EEG)和皮层脑电图(ECoG)两种数据记录,对基于运动想象的BCI的分类算法进行了研究。5个受试参加了一个在线反馈左/右手运动想象BCI实验,记录的EEG数据用于离线分析。全局场强、全局场变化频率和空间复杂性这三个线性描述符能够定征特定脑区的总体状态。这三种测量被单独和联合地施加于与运动皮层相关的三个电极子集,提取的特征矢量用于对这些受试的实验数据进行分类。对于来自7导和11导的8个特征矢量,5个受试的最好分类识别率在85%和99.9%之间,而平均分类识别率在89%和93.5%之间。幅度和相位耦合测量,分别由非线性回归系数和相位锁定值量化,为BCI特征提取提供了一种新的思路。将这两种测量分别施加于由三种耦合方法决定的少量感兴趣的电极,获得的6个特征矢量用于对这5个受试的实验数据进行分类。结果表明,耦合测量对运动想象数据具有好的可分性,而且耦合特征和AR特征的结合能够明显地改进分类识别率。为了促进基于ECoG的BCI的开发及实用化,第三届BCI数据分析竞赛提供了一个左手小指/舌头运动想象ECoG数据,训练集和测试集记录于不同的时间,要求分类算法具有在两组数据之间进行转换的能力。本文提出了一个多特征结合的方法对这个数据进行分类。基于运动相关电位和事件相关去同步的三个特征矢量,分别由共空域子空间分解算法和波形均值提取,然后由Fisher判别分析降为一维,并将它们连接成一个三维的特征矢量,最后使用线性支持向量机进行分类。这个算法具有分类识别率高、稳健性好和泛化能力强等特点,因而在测试集数据取得了91%的高分类识别率。此项结果在27份参赛报告中名列第一。为了简化上述竞赛算法的复杂性,本文提出了一种基于特征子集选择的分类算法。根据两类数据的绝对平均功率差选择10个最优导联,这些导联之间的关联信息由非线性回归系数在0-3Hz和8-30Hz两个频带提取,并将它们连接成一个200维的特征矢量。为了去除冗余信息和保留对分类有意义的特征,由遗传算法并结合支持向量机进行特征选择。一个包含29个特征的最优子集被选出,在测试集数据取得了87%的分类识别率。这个算法的分类识别率与竞赛算法具有可比性,并且具有更高的运算速度。更多还原

【Abstract】 The classification performance of a brain-computer interface (BCI) depends on the user’s ability to control his/her brain state, data recording methods and classification algorithms. Studies on classification algorithms of motor imagery-based BCIs, based on two types of data recordings, electroencephalogram (EEG) and electrocorticogram (ECoG), are presented in this dissertation.Five subjects participated in an on-line BCI experiment during which they were asked to imagine either left or right hand movement. The EEG recordings from all subjects were analysized off-line. The three multichannel linear descriptors, global field strength, global frequency of field changing and spatial complexity, could characterize the overall state of the brain. The three measures were applied alone and together to three electrode subsets determined by neurophysiological a priori knowledge, and the resulting feature vectors were used for classifying the data from the five subjects. For the eight feature vectors derived from 7 and 11 electrodes, the best and averaged classification accuracies of five subjects range from 85% to 99.9% and from 89% to 93.5% respectively.Amplitude and phase coupling measures, quantified by nonlinear regressive (NLR) coefficients and phase locking values (PLV) respectively, provide a new route for BCI feature extraction. The two measures were separately appied to two coupling methods decided by neurophysiological a priori knowledge and a small number of electrodes of interest, the resulting 6 feature vectors were used for classifying the data from the five subjects. Results indicated that coupling measures have good separability for the motor imagery data, and the combination of coupling features and AR features could significantly improve classification accuracy.To promote the development and practicality of ECoG-based BCIs, the organisors of BCI Competition III provided an ECoG data of imagined movement of left small finger or tongue in which the training set and test set were recorded in two different sessions, and required the contributing algorithms to have the ability of session-to-session transfer. We presented an algorithm of feature combination for classifying the data set. Three feature vectors, based on movement-related potentials (MRP) and event-related desynchronization (ERD), were extracted by common spatial subspace decomposition and waveform mean, then they were reduced to one dimension by Fisher discriminant analysis and concatenated into a three-dimensional feature vector, and finally a linear support vector machine was used for classification. The algorithm posseses the characteristics of high classification accuracy, good robustness and strong generalization ability, and thus achieves the high classification accuracy of 91% on test set. The result ranks first among 27 contributions from the whole world.To simplify the complexity of the competition algorithm and increase its operating speed, a feature subset selection-based algorithm was presented for classifying the competition data. 10 optimal leads are chosen according to the averaged band power difference between the two classes of signals, then the association information among these leads is extracted by nonlinear regressional coefficients in two frequency bands 0-3Hz and 8-30Hz and concatenated into a 200-dimensional feature vectors. To eliminate redundant information and retain meaningful features for classification, an optimal subset of 29 features is picked out by combining a genetic algorithm for feature selection with a support vector machine for their evaluation. The algorithm achieves a classification accuracy of 87% on test set that is comparable with the competition algorithm, and has higher operating speed.更多还原