基于EEG和fNIRS的多模态脑—机接口的特征提取与分类方法研究

发布时间：2018-04-11 16:00

本文选题：脑-机接口 + 多模态　；参考：《杭州电子科技大学》2017年硕士论文

【摘要】：脑-机接口(Brain-Computer Interfaces,BCI)是一种直接从中枢神经系统提取信息,采用人的思维直接操作外围设备的新技术,是未来人机交互的最高形态。脑-机接口不仅广泛应用于医学康复领域,在军事、休闲娱乐、心理卫生和人工智能等多个领域都具有较高的研究意义和价值。本文首先介绍了脑-机接口的相关研究背景,针对传统基于单一模态脑电(electroencephalography,EEG)脑-机接口易受环境噪声干扰、分类精度低等问题,在EEG脑-机接口的研究基础上,引入功能近红外成像(function Near Infrared Spectroscopy,fNIRS)技术,设计并简化了基于握拳动作的EEG-fNIRS多模态脑-机接口的实验范式,研究最重要的特征提取与分类环节。首先,利用实验室EEG、fNIRS采集与分析系统对三名受试者进行EEG-fNIRS的同步采集实验,并对原始信号进行滤波去噪、基线校正等预处理。根据握拳动作诱发的EEG信号的事件相关去同步、事件相关同步现象及其时频特性,提取了EEG信号的频带能量、AR模型系数和小波系数特征;同时根据握拳动作引起的血液动力学响应的特点,提取含氧血红蛋白浓度不同时段的均值及斜率特征。对特征向量进行归一化处理之后,采用线性判别分析(Linear Differential Analysis,LDA)及支持向量机(Support Vector Machines,SVM)对不同类型特征分类并进行8次5折交叉验证。结果表明,EEG的小波系数特征分类效果要好于频带能量与AR模型参数结合的特征;fNIRS的斜率特征分类效果要好于均值特征,其中斜率特征分类正确率最高的时间段在执行动作任务之后的3~5s。其次,根据单模态特征分类结果,提出了基于EEG小波系数和fNIRS斜率结合的融合特征,并对结合特征使用主成分分析(Principal Component Analysis,PCA)。然后采用LDA、SVM分类,并进行8次5折交叉验证,对比多模态信号与单模态信号的分类正确率。结果表明,经过特征融合的握拳动作任务平均识别率比单独的EEG特征和fNIRS特征提高3~9%。表明fNIRS能够显著增强基于EEG的脑-机接口性能,利用多模态脑信号能够提高传统脑-机接口系统的性能,以及对实验范式的简化,对提高EEG-fNIRS多模态BCI的应用有一定的意义。
[Abstract]:Brain-Computer Interface (BCI) is a new technology for extracting information directly from the central nervous system and directly manipulating peripheral equipment by human thinking. It is the highest form of human-computer interaction in the future.Brain-computer interface is not only widely used in the field of medical rehabilitation, but also has high significance and value in military, recreational, mental health, artificial intelligence and other fields.This paper first introduces the related research background of brain-computer interface, aiming at the problems that the traditional brain-computer interface based on single mode electroencephalography (EEG) is vulnerable to environmental noise interference and low classification accuracy, based on the research of EEG brain-computer interface.The functional near infrared imaging function Near Infrared spectroscope of NIR is introduced to design and simplify the experimental paradigm of EEG-fNIRS multimodal brain-computer interface based on grip movement, and the most important feature extraction and classification are studied.Firstly, three subjects were collected synchronously by using the EEGFNIRS acquisition and analysis system in the laboratory, and the original signals were filtered and de-noised, and the baseline correction was performed.According to the event correlation desynchronization, event correlation synchronization and time-frequency characteristics of EEG signal induced by fist grip, the coefficients of AR model and wavelet coefficients of EEG signal are extracted.At the same time, according to the characteristics of hemodynamic response caused by the grip movement, the mean and slope characteristics of hemoglobin concentration in different periods were extracted.After normalization of feature vectors, linear discriminant analysis (LDAs) and support Vector machines (SVM) are used to classify different types of features and perform 5 fold cross validation for 8 times.The results show that the feature classification effect of wavelet coefficients of EEG is better than that of slope feature classification of fNIRS which combines band energy with AR model parameters.Among them, the time period with the highest accuracy of slope feature classification is 3 / 5 s after performing the action task.Secondly, according to the classification results of single mode features, a fusion feature based on EEG wavelet coefficients and fNIRS slope is proposed, and principal component analysis (PCA) is used to analyze the combined features.Then, LDA-SVM is used to classify, and 8 times 5 fold cross validation is carried out to compare the classification accuracy between multimodal signal and single mode signal.The results show that the average recognition rate of grip movement task after feature fusion is 3% higher than that of EEG feature and fNIRS feature.It is shown that fNIRS can significantly enhance the performance of BCI based on EEG, and the performance of traditional BCI system can be improved by using multimodal brain signals, and the simplification of experimental paradigm is of certain significance to improve the application of EEG-fNIRS multimodal BCI.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R318;TN911.7

【参考文献】