【作者】 孙静；

【导师】 王海贤；

【作者基本信息】 东南大学 ， 生物医学工程， 2022， 硕士

【摘要】 在基于脑电图(Electroencephalogram,EEG)的脑-机接口(Brain–computer interfaces,BCI)领域,共同空间模式(Common spatial patterns,CSP)技术是EEG信号特征提取领域中广泛使用的一种方法。受到余弦距离可以扩大不同类别样本之间的距离这一事实的启发,本文提出了欧拉CSP(Euler common spatial patterns,e-CSP)用于脑电信号的特征提取,然后将提取的特征用于脑电分类。e-CSP本质上是融合欧拉表示的CSP。它主要包括以下两个步骤:每个样本值首先通过使用欧拉表示映射到一个复数的空间,然后在欧拉空间中构建经典的CSP方法。因此,e-CSP相当于将欧拉表示作为核函数应用于CSP的输入。它在计算上与CSP一样简单,但是它能够从EEG信号中提取了更多的鉴别性特征,从而提高识别的准确率。本文对此进行了大量的实验,实验结果也证明了e-CSP的识别能力。欧拉共同空间模式从距离度量的角度优化了CSP方法,对于CSP的目标函数而言,协方差矩阵的计算是十分关键的步骤,这是本文考虑的另一个优化角度。跨频耦合(Cross-frequency coupling,CFC)表示不同频段之间的相互作用,比单一频段更能控制复杂的大脑网络,为脑电信号的研究提供了一个新的思路。本文应用振幅-振幅耦合(Amplitude–amplitude coupling,AAC)的方式来重新表述e-CSP中的协方差矩阵;由此,提出了基于AAC修饰的e-CSP,称为CFC-e CSP。通过所提出的方法,提取的特征更加细致化,在后续任务识别上更具优势。所提出的方法在Cho数据集上得到了验证,实验结果说明了所提方法的鉴别能力。为了进一步地提升方法的性能,本文受到了微状态分析理论的启发,提出基于微状态的动态空间模式的构建。在这一章中研究了基于运动想象的微状态的分析,基于分析结果分别开展了各个微状态标签下的分类实验。实验是基于Cho数据集进行的,实验结果表明,在微状态的框架下,CFC-e CSP方法提取了更有利于运动想象任务识别的空间模式,分类准确率有所提升。本文主要基于欧拉表示与跨频耦合两个方面对CSP算法进行改进,在BCI竞赛的公开数据集以及Cho数据集上对三种方法都做了验证,算法的分类准确率得到了提升。本文从信号的特性出发,考虑了与运动想象任务的相关性,在对CSP算法优化进行深入研究的同时,也使得方法提取的特征更加细致化,更具识别性。更多还原

【Abstract】 In the field of electroencephalogram(EEG)-based brain–computer interfaces(BCIs),the technique of common spatial patterns(CSP)is a widely used method in the field of feature extraction of electroencephalogram(EEG)signals.Motivated by the fact that a cosine distance can enlarge the distance between samples of different classes,we propose the Euler CSP(e-CSP)for the feature extraction of EEG signals,and it is then used for EEG classification.The e-CSP is essentially the conventional CSP with the Euler representation.It includes the following two stages: each sample value is first mapped into a complex space by using the Euler representation,and then the conventional CSP is performed in the Euler space.Thus,the e-CSP is equivalent to applying the Euler representation as a kernel function to the input of the CSP.It is computationally as straightforward as the CSP.However,it extracts more discriminative features from the EEG signals.Extensive experimental results illustrate the discrimination ability of the e-CSP.The Euler common spatial patterns optimizes the CSP method from the aspect of distance metric.For the objective function of CSP,the calculation of covariance matrix is a very critical step,which is another optimization aspect we consider.Cross-frequency coupling(CFC)represents the interaction between different frequency bands,which can better control the complex brain network than a single frequency band and provides a new idea for research on EEG signals.In this paper,we apply amplitude–amplitude coupling(AAC)to reformulate the covariance matrices in e-CSP;as a result,the AAC-modulated e-CSP is proposed.With the proposed method,the extracted features are more detailed and more advantageous for subsequent task recognition.The proposed method is validated based on the Cho’s dataset.The experimental results illustrate the discrimination ability of the proposed method.In order to further improve the performance of the method,inspired by the theory of microstate analysis and proposes the construction of a microstate-based dynamic spatial model.In this section,the analysis of microstates based on motor imagery(MI)is investigated,and the classification experiments under each microstate label are carried out separately based on the analysis results.The experiments are all conducted based on the Cho dataset,and the experimental results show that the classification accuracy of the CFC-e CSP method is improved in the framework of microstates,and extracted spatial patterns that were more beneficial for the recognition of motor imagery tasks.We improve the CSP algorithm mainly based on two aspects,Euler representation and cross-frequency coupling,and validate all three methods on the public dataset of the BCI competition as well as the Cho dataset,and the classification accuracy of the algorithm is improved.We consider the relevance to the motor imagery task in terms of signal characteristics,and make the features extracted by the methods more detailed and recognizable while conducting intensive research on the CSP algorithm improvements.更多还原