【作者】 张鹏；

【导师】 文冬； 郑岳；

【作者基本信息】 燕山大学 ， 计算机技术（专业学位）， 2021， 硕士

【摘要】 随着人工智能的兴起,以脑机接口(Brain Computer Interface,BCI)为桥梁的脑神经科学研究正迅速展开。其中对虚拟现实场景下空间认知脑电信号的分析成为该领域研究热点,通过空间认知训练前后脑电信号变化可以有效评估认知能力训练效果。目前,在空间认知脑电信号研究方面取得了不少进展,主要包括脑电信号的特征提取以及后续的数据分类。但是仍然存在着不足,主要体现在计算不同通道之间的耦合特征强度时没有考虑到其他通道空间位置的影响。另外,探索出一种鲁棒性更强、性能更优的脑电特征数据分类模型也是该领域的一大难点。基于此,本文结合国内外研究现状,从特征提取与分类方法入手提出了新的脑电耦合特征提取方法和脑电信号分类模型。首先,本文结合互信息理论,提出了基于多维条件互信息共空间模式(Multivariate Conditional Mutual Information Common Spatial Pattern,MCMICSP)的脑电信号特征提取算法。由于传统的CSP算法输入的脑电信号要求表现出严格的线性相关,并不适用于本实验的脑电数据;因此本文将传统共空间算法中的协方差矩阵替换为多维条件互信息耦合矩阵。这样不仅考虑其他脑电通道对耦合特征的影响,而且根据脑电信号的线性相关程度来构建空域特征滤波器,进一步弥补了现有特征提取方法的不足。其次,本文提出了多尺度密集融合卷积神经网络(Multiple Scale Dense Fusion Convolutional Neural Network,MSDFCNN)。针对传统CNN卷积核单一、容易缺失有效特征信息等问题,引入不同尺度卷积核来有效减少脑电信号特征丢失的现象。针对本实验研究样本相对较少的问题,通过引入密集网络方法策略,实现特征重用,同时减少梯度弥散;并使用自适应梯度随机下降算法对提出的分类算法进行过程优化。最后,对上述内容进行实验与结果分析。为了分别验证所提两种算法的有效性,本研究选取了虚拟现实场景下空间认知训练前后的脑电数据作为数据集,并与原有的特征提取和分类算法开展对比实验。结果表明,本研究提出的特征提取算法能够更加有效地评估空间认知训练效果;提出的分类算法在脑电耦合特征分类方面有更高的准确度与更强的泛化能力。更多还原

【Abstract】 With the rise of artificial intelligence,brain neuroscience research using Brain Computer Interface(BCI)as a bridge is rapidly unfolding.Among them,the analysis of spatial cognitive EEG signals in virtual reality has become a research hotspot in this field.The changes of EEG signals before and after spatial cognitive training can effectively evaluate the effect of cognitive ability training.At present,a lot of progress has been made in the research of spatial cognitive EEG signals,mainly including the feature extraction of EEG signals and subsequent data classification.However,there are still some shortcomings,which mainly reflected in the calculation of the coupling feature strength between different channels without considering the influence of the spatial position of other channels.In addition,it is also a major difficulty in this field to explore a classification model of EEG feature data with stronger robustness and better performance.Based on this,this article proposed a new EEG coupling feature extraction method and EEG signal classification model combining the current research status at home and abroad.Firstly,this article proposed an EEG signal feature extraction algorithm based on the Multivariate Conditional Mutual Information Common Spatial Pattern(MCMICSP),combining the theory of mutual information.This is due to the fact that the traditional CSP algorithm input EEG signal is required to show strict linear correlation,it is not suitable for the EEG data of this experiment.Therefore,the covariance matrix in the traditional co-space algorithm is replaced by the multi-dimensional conditional mutual information coupling matrix,which not only considers the influence of other EEG channels on the coupling feature,but also constructs the spatial feature filter is according to the linear correlation degree of EEG signal,which further makes up for the shortcomings of the existing feature extraction methods.Secondly,this paper proposed a multiple scale dense fusion convolutional neural network(Multiple Scale Dense Fusion Convolutional Neural Network,MSDFCNN).This is due to the fact that traditional CNN convolution kernel is single and easy to lose effective feature information,different scale convolution kernels are introduced to effectively reduce the feature loss of EEG signal.In view of the relatively small number of samples studied in this experiment,the dense network method is introduced to achieve feature reuse while reducing gradient dispersion;and the adaptive gradient stochastic descent algorithm is used to optimize the process of the proposed classification algorithm.Finally,the experiment and result analysis of the above content are carried out.In order to verify the effectiveness of the two proposed algorithms,this study selected the EEG data before and after spatial cognitive training in the virtual reality as the data set,and carried out comparative experiments with the original feature extraction and classification algorithms.The results show that the feature extraction algorithm proposed in this study can more effectively evaluate the effect of spatial cognition training;the proposed classification algorithm has higher accuracy and stronger generalization ability in the classification of EEG coupling features.更多还原