Research On Granger Causal Brain Network Analysis And Emotion Recognition Of EEG Signals

Inspired by the global brain science research program,it has become one of the mainstream directions of contemporary scientific development to explore the information processing mechanisms and cognitive behavior mechanisms of human intention,emotion,and other physiological activities.In the field of brain science and artificial intelligence,electroencephalogram(EEG)emotion recognition is considered to be a frontier research field based on brain network analysis methods.Cognitive neuroscience has demonstrated a complex causal relationship between human brain regions involved in emotional cognition,which is related not only to the asymmetry of left and right hemispheres but also to the synergistic interactions among various brain regions under emotional cognition.As a statistical hypothesis testing method,Granger causality(GC)can effectively reveal the causal relationship between two signals.Therefore,the purpose of this thesis is to construct the causal brain network of EEG signals using the GC method,as well as conduct research on GC analysis and emotion recognition of EEG signals under different brain divisions.These studies are important to our understanding of the mechanism of emotion processing in the human brain,improving the performance of emotion recognition systems,and promoting the development of intelligent human-computer interaction.Most of the current researches on causal brain network analysis of emotional EEG signals has mainly focused on the intra-frequency band across whole brain regions,while the causal interactions among the whole brain regions,left and right hemispheres,and multiple brain regions cross frequency bands are not fully considered.To address these limitations,this thesis combines the brain region division strategy with causal relationships of EEG signals,and considers the perspectives of whole brain regions,left and right brain regions,multi-brain regions and whole brain regions with multiple channels.This paper utilizes the progressive causal analysis method from global to local brain regions to investigate the GC brain network constructed based EEG signals,feature extraction,feature fusion and emotion recognition model construction,respectively.It is of great significance for gaining a deeper understanding of the connection between emotions and brain regions,as well as improving the accuracy of emotion recognition.The main work and innovation points are as follows:(1)A multi-band GC brain network analysis method was proposed for analyzing EEG signals from whole brain regions.To comprehensively analyze the causal relationships of EEG signals,this thesis divided the causality of the whole brain EEG signals into two categories: the same-frequency band causality and the cross-frequency band causality based on the frequency band division of EEG signals.On this basis,qualitative and quantitative analysis methods were used to study the EEG emotional recognition performance of cross-frequency bands in the GC brain network.Experimental results demonstrate that both the same-frequency band and the adjacent frequency band EEG signals hold the favorable causal properties,which will provide a theoretical basis for further improving the causal analysis framework of EEG signals.(2)A multi-band GC brain network analysis and feature extraction method for EEG signals in left and right brain regions was proposed.Based on the asymmetric characteristics of the left and right hemispheres of the brain,a hemispheres frequency-Granger causality(HF-GC)measure was proposed to divide the causal relationships of EEG signals into four categories,and the GC brain networks were constructed accordingly.Then,an adaptive two-stage decorrelation feature extraction scheme(ATD)was designed for the adjacency matrix of the GC brain network to effectively remove the redundant connections of the GC brain networks.Finally,a multi-GC feature fusion scheme was designed to balance the recognition accuracy and feature number of each GC feature,which comprehensively considers the influence of recognition accuracy and computational complexity of the four types of GC features.The experiment results have verified the effectiveness of the proposed method.Furthermore,this research will provide a new approach to understanding the asymmetry of the left and right brain regions in emotion processing,extracting effective EEG features,and improving the accuracy of emotion recognition.(3)An emotion recognition model that combines the GC brain networks and the attention mechanism of multi-brain EEG signals was proposed.Based on the characteristics that the different brain regions in the cerebral cortex have different ability to represent emotion,a novel multi-brain EEG electrode arrangement method was developed and the GC brain network was constructed,so that the spatial causality of intra-brain and inter-brain regions of EEG signals was included in the GC adjacency matrix features and can comprehensively reflect the information transmission mechanisms of human brain.Then,a multi-frequency band and region-aware spatial attention mechanism of 2DCNN emotion recognition model was designed by combining the structural characteristics of GC brain network and the frequency characteristics of EEG signals.By adaptively allocating weights to different frequency bands and brain regions using the attention mechanism,the performance of EEG emotion recognition was improved by focusing on the frequency band and brain region relationships that better represent emotional states.This research is of great significance for gaining a deeper understanding of the causal information interaction between different brain regions and the mechanisms involved in emotion processing.(4)A graph feature extraction method for the GC brain network of the whole brain with multiple channels was proposed.Considering the irregular spatial distribution of EEG electrode channels in the cerebral cortex leads to the problem that EEG signals are not grid data.The GC brain network was combined with a graph convolutional neural network in order to construct a new GC graph feature that is more consistent with the cognitive patterns of the brain.Following that,the causal graph features of each frequency band EEG signals in the same electrode channel are weighted fusion according that the different frequency EEG signals contribute differently to emotion recognition,which will facilitate a deeper exploration of the frequency information within each electrode channel that significantly reflects changes in emotional states,and further enhancing the performance of the emotion recognition system.This research provides initial insights into the impact of the whole-brain multi-channel EEG electrode spatial topology on emotion recognition,making the graph feature extraction process more interpretable.