EEG（electroencephalographic）源信号同时具有高时间和高空间分辨率的特性，已被广泛用于脑认知功能和疾病等相关研究中。目前针对EEG源信号的分析方法研究主要有：EEG源信号的估计，即利用采集到的sensor-EEG估计出EEG源信号，可称为EEG源定位技术；基于EEG源信号的脑网络构建方法研究，即利用EEG源信号构建不同脑网络的方法，如功能连接网络，有向功能连接网络。围绕EEG源信号，本论文主要包括以下几个方面的内容：

（1）结合fMRI（functional magnetic resonance imaging）提供的空间先验信息，提出一种新的EEG源定位算法。fMRI具有较高的空间分辨率，与EEG的高时间分辨率特性形成互补，可为EEG源定位提供具有真实生理学意义的空间先验信息，帮助提高源定位的准确性。但是目前结合fMRI的EEG源定位方法没有考虑源信号的时间过程特性，以及脑信号的非平稳性特点，甚至一些还表现出对fMRI先验信息的过度依赖。针对这些问题，本研究提出了一种新的结合fMRI的时空约束EEG源定位方法FIST（fMRI-informed spatio-temporal unifying tomography）。该方法将fMRI的空间先验信息与基于时频分解系数的混合范数约束项相结合，同时对源信号的时间和空间进行约束。分别在多种模拟数据和真实EEG数据下验证了FIST的有效性。结果表明，与没有引入fMRI信息的源定位方法如STOUT（spatio-temporal unifying tomography），LORETA（low resolution tomography）以及wMNE（weighted minimum-norm estimate）相比，FIST的空间和时间估计准确性在EMD（earth mover's distance），AUC（area under the ROC curve）以及MSE（mean square error）三个指标上都有提高。与引入fMRI信息的方法dSPM（dynamic statistical parametric mapping）相比，FIST表现出对fMRI先验具有一定的选择性，且未完全依赖。

（2）基于多变量自回归（multivariate autoregressive，MVAR）模型，利用全局皮层因子信号，提出一种新的脑有向功能连接（brain directed functional connectivity，BDFC）估计的方法。基于MVAR模型系数的PDC（partial directed coherence）是频域有向功能连接估计的常见方法之一。然而由于源信号的高空间分辨率，直接将其应用于MVAR模型中进行模型系数的估计非常困难，需要估计的系数个数高达百万级。另外目前利用PDC表征的有向功能连接，其得到的频域连接与EEG源信号的频域特性不符，这也使PDC的生理可解释性较低。针对以上问题，我们提出基于全局皮层因子信号的MVAR方法（global cortex facor-based MVAR，GCF-MVAR）。对EEG源信号，采用特征值-特征向量分解进行降维，将得到的低维因子信号利用MVAR模型进行估计。此外，通过引入能反映EEG源信号真实频域特征信息的功率谱，将EEG源信号的功率谱与PDC相乘，得到加权PDC（weighted PDC，wPDC）用于表征有向功能连接。分别利用多种模拟数据和真实EEG数据验证了GCF-MVAR方法的有效性。结果表明，与FMVAR （factor MVAR），ROI-MVAR以及MVAR方法相比，提出的GCF-MVAR对有向功能连接估计的准确性最高，且生理可解释性最强。

（3）针对熟悉人脸识别中的两个ERP（Event-related potential）成分N170和N250，它们到底分别是由哪些脑区产生，他们的时空特性和频域特性又是什么，以及是否在200 ms之前已有脑区参与熟悉信息加工这些问题仍存在争议。本研究利用尝试FIST源定位方法和S-Transform时频分析方法来探索上述问题。结果表明，在对熟悉人脸进行加工时，枕叶-颞叶区即检测到N170成分的脑区，频域活动主要为4-10 Hz，其中6-10 Hz活动较强，且在100 – 200 ms达到最大，并持续到200 ms之后。另外对于N170成分，右侧梭状回，右侧颞下回-枕叶外侧-顶下小叶，右侧舌回，右侧顶上小叶，右侧楔状叶，以及右侧楔前叶均被激活，是N170的产生源，负责对面部结构的识别。而对N250成分，左侧和右侧梭状回，左侧颞下回-枕叶外侧-顶下小叶，右侧舌回，右枕叶外侧，以及左侧颞中回后部是其产生源，负责对熟悉信息的加工。大脑对熟悉信息的加工呈动态性的，左侧和右侧梭状回、右侧舌回、左侧和右侧颞下回-枕叶外侧-顶下小叶，左侧颞中回后部，以及右侧颞下回中部在刺激呈现后的200 ms前已参与到对熟悉信息的加工，其中右侧颞下回-枕叶外侧-顶下小叶以及右侧颞下回中部在200 ms后不再参与。

（4）利用本论文提出的GCF-MVAR方法，构建情绪相关的BDFC网络，用于情绪识别的研究中。目前关于情绪识别的研究中使用的不同时域或者频域EEG均是基于sensor-EEG信号提取的，反映了大脑不同位置的神经元信号活动。但是这些特征无法反应与情绪有关的神经元活动的具体脑区位置信息，以及情绪加工过程中不同脑区间的信息的传播和相互作用。本研究设计了一个基于EEG源信号的BDFC网络情绪识别研究框架，利用高密度EEG信号估计得到EEG源信号后，使用GCF-MVAR方法提取BDFC网络特征，用于对不同情绪状态的分类。分别基于高密度EEG信号和常用的低密度EEG信号提取了传统基于sensor的特征，如PSD（power spectral density），RASM（rational asymmetry），DASM（differential asymmetry），ASM（asymmetry），以及DCAU（differential caudality），并将其分类性能与BDFC特征行了比较分析。结果表明基于EEG源信号得到的BDFC特征的分类率最高，可达85.00%。将BDFC特征分别与基于高密度sensor-EEG提取的特征和低密度sensor-EEG提取的特征进行融合后，其分类效果也相较于单独使用sensor-EEG特征时得到提高。基于高密度EEG信号提取的传统特征的分类率高于低密度EEG的分类率。

综上，本文围绕EEG源信号进行了EEG源定位方法和基于EEG源信号的BDFC网络构建方法研究，并将提出的方法分别应用于熟悉人脸识别和情绪识别的脑机制研究中。这些结果为基于EEG源信号进行认知功能和脑疾病等相关研究提供了新的思路和方法。

EEG (electroencephalographic) source signal which has high temporal and spatial resolution has been widely used in cognitive and brain disorders studies. At present, the analysis methods for EEG source signals mainly include: EEG source localization, that is, estimate EEG source signals based on recorded sensor-EEG; brain network construction based on EEG source signals, that is, a method for constructing different brain state-related networks by using EEG source signals, such as functional connectivity network method and directed functional connectivity network method. Focusing on the EEG source signals, the main contents of this paper are as follows:

(1) We propose a new EEG source localization method which use functional magnetic resonance imaging (fMRI) signals to provide prior source locations. Complementary with EEG, fMRI which has high spatial resolution is powerful at providing prior source locations based on actual brain physiology. It hereby can help improve the accuracy of EEG source localization. However, most of the current methods which use fMRI to provide prior information do not account for the temporal interrelated and non-stationary features of electromagnetic brain signals, and some are too much dependent on the fMRI prior. Here, we propose a new fMRI informed EEG source localization method which termed fMRI-informed spatio-temporal unifying tomography (FIST). It combines the fMRI prior with a mixed norm constraint based on time-frequency decomposition coefficients, so that the temporal and spatial of source signals can be simultaneously constrained. Both simulated and real EEG data are applied to assess the performance of the proposed method. Compared with methods that do not introduce fMRI information, such as STOUT (spatio-temporal unifying tomography), LORETA (low resolution tomography), and wMNE (weighted minimum-norm estimate), the temporal and spatial estimation accuracy of FIST is improved in three metrics: EMD (earth mover's distance), AUC (area under the ROC curve) and MSE (mean square error). In addition, FIST shows good ability to select the fMRI priors to get a better estimation without totally depending on the prior, when comparing with dSPM (dynamic statistical parametric mapping) method which also has fMRI prior information.

(2) We propose a new brain directed functional connectivity (BDFC) method based on multivariate autoregressive (MVAR) model using global cortex factor signals which derived from EEG source signals. The partial directed coherence (PDC) computed based on MVAR model coefficient is increasingly being used to study BDFC in the frequency domain based on EEG source signals. However, due to the high spatial resolution of EEG source signals, it is very difficult to directly fitting an MVAR model with EEG source signals, the number of coefficients to be estimated is up to millions. In addition, the frequency domain BDFC which obtained with PDC does not match the frequency domain characteristic of EEG source signals, it makes the PDC physiologically hard to interpret. To solve these methodological issues, we propose a new BDFC method termed global cortex factor-based MVAR (GCF-MVAR). GCF-MVAR use low-dimensional global cortex factor signals, which derived from EEG source signals by eigenvalue-eigenvector decomposition, to fit MVAR model. In addition, the weighted PDC (wPDC) is proposed to measure BDFC, which is obtained by multiplying the source spectral power (SP) by PDC. SP can reflect the true frequency activity in a source region. GCF-MVAR is applied to both simulated and real EEG data to assess its performance. The results are compared with those obtained using other methods including FMVAR (factor MVAR), ROI-MVAR, and MVAR. It shows that GCF-MVAR had the lowest estimation error. By virtue of using the source SP to weight the PDC, GCF-MVAR improves the physiological interpretation of the source connectivity.

(3) For the two event-related potential (ERP) components N170 and N250 in familiar face recognition, it’s still under debate about which brain regions are generators of N170 and N250, what’s the spatio-temporal and frequency domain characteristics, and whether there are brain regions which have been involved in familiarity information processing before 200 ms after the stimulus onset. The current study explores the above questions using FIST source localization method and time-frequency analysis method S-Transform. The results indicate that brain activities of 4-10 Hz at occipito-temporal sit (EEG065) were observed, activities at 6-10 Hz are stronger, reach the maximum during 100-200 ms, and continue to 200 ms later. In addition, for the N170, right fusiform, the right inferior temporal-lateral occipital- inferior parietal, the right lingual, the right superior parietal lobe, the right cuneus and precuneus are activated, which may be the generator of N170, and are related with structural face processing. For the N250, the left and right fusiform, the left inferior temporal-lateral occipital- inferior parietal, the right lingual, the right lateral occipital, and the left posterior of middle temporal are the generators, they modulate familiarity. The brain is dynamic in the processing of familiarity information, the left and right fusiform, the right lingual, the left and right inferior temporal-lateral occipital-inferior parietal, the left posterior of middle temporal and the right middle of inferior temporal are involved in the processing of familiarity information before 200 ms after the stimulus onset, while the right inferior temporal-lateral occipital- inferior parietal and the right middle of inferior temporal are not involved 200 ms later.

(4) We use the proposed BDFC method GCF-MVAR to construct emotion-related BDFC network, and use it as features for emotion recognition to explore the brain mechanism of emotions. Currently, a variety of frequency and time domain features extracted from sensor EEG signals are used for emotion recognition. However, they cannot unambiguously describe the location of emotions associated with neural activities and information propagation or the interaction between brain regions. In this study, we designed a BDFC network-based framework using EEG source signals to investigate emotion recognition. First high-density sensor-EEG signals is used to estimate EEG source signals, then GCF-MVAR method is utilized to extract emotion-related BDFC features for emotion recognition. In addition, traditional sensor-EEG based features such as PSD (power spectral density), RASM (rational asymmetry), DASM (differential asymmetry), ASM (asymmetry), and DCAU (differential caudality), are also extracted using both low-density EEG signals and high-density EEG signals, respectively, and compared their classification performance with BDFC features. The results reveal that BDFC features facilitated the highest classification rate of up to 85.00 %. When combine the BDFC with PSD and DE which derived from low-density sensor-EEG signals and high-density sensor-EEG signals, respectively, the classification accuracies are also improved compared with that only use sensor-EEG features. The sensor features derived from high-density EEG signals also exhibited higher classification accuracy compared to low-density EEG signals.

In conclusion, this paper investigates the EEG source localization method and the BDFC method based on EEG source signals, and applies the proposed methods to study the brain mechanism of familiar face recognition and emotion recognition, respectively. These results provide new ideas and methods for cognitive and brain disorders-related studies based on EEG source signals.