Research Methods On EEG Emotion Recognition Based On Graph Convolutional Neural Networks

That machines recognize and respond to human emotion is a research hotspot in the artificial intelligence field.Emotion recognition is inseparable from the carrier signals of emotion,which is generally divided into two types: non-physiological signals(voice,expression,action,etc.)and physiological signals(electroencephalogram,skin electricity,electrocardiogram,etc.).Physiological signals are real and cannot be disguised,so using physiological signals for emotion recognition can effectively improve the reliability of recognition and reduce misjudgment of human intentions.Neuroscience studies have shown that brain regions such as the frontal cortex and amygdala are closely related to the expression of emotion,so the Electroencephalogram(EEG)signals obtained from the scalp of the human brain can directly and accurately reflect the emotional activity state of the human brain.In addition,the EEG signal acquisition equipment is low in cost,easy to operate,and non-invasive to the human body.Obviously,using EEG signals for emotion recognition has patent advantages among physiological signals.Considering that the measurement of the adjacency relationship between EEG electrodes does not necessarily follow the Euclidean distance metric,the dissertation uses Graph Convolutional Neural Network(GCN)to model the EEG signals.GCN-based EEG modeling methods comprehensively consider EEG feature information and electrode topology,and learn better graph representations through feature propagation and aggregation.Focusing on the EEG emotion recognition problem,the dissertation conducts research on the recognition methods based on GCN,and builds an EEG emotion recognition system based on these methods.The main work of this dissertation are as follows:(1)A self-attention spatio-temporal graph convolutional neural network model is proposed for EEG emotion recognition task.In this model,a graph structure is used to effectively capture the spatio-temporal dependencies of EEG electrodes.At present,the understanding of the emotional neural activity mechanism of the brain is not completely clear,and it is difficult to use the existing prior knowledge to construct the emotional functional connections between EEG electrodes.Considering that the emotional functional connections between EEG electrodes play an important role in emotion recognition,the dissertation constructs the correlation between electrodes based on self-attention,and captures temporal dependencies of EEG signals by constructing temporal graphs,to extract emotional information of EEG signals from spatial and temporal domains,respectively.The proposed model achieves state-of-the-art performance on two public EEG datasets.In addition,the effectiveness of the self-attention graph generation method is demonstrated by comparing with other graph generation methods.(2)A spatial multiscale graph convolutional neural network model is proposed to handle the EEG emotion recognition task.Research in brain science has shown that certain neural nuclei such as the amygdala,hippocampus,and striatum are responsible for the primary processing and perception of emotion,while brain regions such as the frontal lobe are responsible for higher-order emotion regulation(eg,cognitive regulation).From the scale of the cerebral hemisphere,there are lateral differences in the perception and expression of different types of emotion between the left and right hemispheres,and different levels of brain spatial structure have differences in the ability to perceive,process and express emotion.In order to effectively utilize the differences of emotional activity in different brain regions,the dissertation proposes to use multiscale graphs for modeling EEG signals to simulate hierarchical emotional expression of the brain from local to global,and constructs graphs based on the selection of EEG electrodes to achieve multiscale graph representation from local to global.Experiments on different datasets show that the method of multiscale graph representation from local to global can effectively improve the accuracies of emotion recognition.(3)Based on the above research,a set of EEG emotion recognition system is built.The system can realize the functions such as loading EEG signals,extracting EEG features from different frequency bands,and classifying emotions,and has the abilities to visualize extracted EEG features and display the results of emotion recognition.