Research On EEG Emotion Recognition Method Based On Core Brain Network

Emotions are a part of human social life and play an important role in interpersonal communication.By identifying the emotional factors of daily activities,it is possible to know the true attitude of human beings towards things.In recent years,with the continuous development and progress of life sciences,emotion recognition has become a research hotspot in the fields of cognitive science,computer science,neuroscience and artificial intelligence.With the advantages of high temporal resolution and low acquisition cost,EEG signals have been widely used in the field of emotion recognition.Among them,the method based on brain network analysis is an important method in EEG data analysis,which can analyze the collaboration relationship between different brain regions and study the overall functional state of multiple brain regions.However,the actual collected EEG signals often carry noise and redundant information,resulting in the constructed brain function network containing redundant information or even interference information unrelated to the current task.Therefore,how to extract the core network from the initial brain functional network and effectively use it has become a key problem in brain network analysis.At present,there are some methods to screen the core brain network through empirically given formulas,which have achieved good results,but there are obvious shortcomings.These methods are relatively simple,empirical-based rather than data-driven,and do not consider the importance of rational use of core EEG channels after obtaining the core network.In addition,existing research work has not explored specific implementation options in cross-subject mode.Based on the above problems,this thesis innovatively completes three aspects of work:(1)In order to make full use of the structural characteristics of brain functional networks to extract core brain networks,this thesis proposes an automatic core brain network extraction method(CW-GCN)based on Channel Weighting(CW)and Graph Convolutional Network(GCN)to automatically extract core networks from the initial functional network in a data-driven manner.On the published emotional EEG dataset SEED,the core brain network extracted by the proposed CW-GCN was evaluated using classification experiments of five classifiers,and compared with the core network extracted by two control methods(random method and the previous method).The experimental results show that the core brain network extracted by the proposed method shows a good classification effect on each classifier,and the best result is 6.45% higher than the random method and 1.49% higher than the previous method,which fully reflects the superiority of the method CW-GCN.In addition,in order to further optimize the obtained core network,this thesis combines the CW-GCN method with the empirical method to further improve the experimental effect,and the result is 2.70%higher than that of previous method.(2)In order to make full use of the importance relationship between core EEG channels in core networks and obtain deeper information about features,this thesis proposes a sentiment recognition model(CC-SRM-GCN)based on Channel Convolution(CC)and Style-based Recalibration Module(SRM)and graph convolutional networks.This model will take advantage of the importance relationship of core EEG channels obtained by the CW-GCN model in the previous chapter and use it for the feature integration.The style recalibration module then processes the integrated feature data at a deeper level.The experimental results show that the CCSRM-GCN model in this thesis achieves a recognition accuracy of 94.97% on the triclassification task of the SEED dataset,indicating that the importance relationship of EEG channels can provide rich information and improve the recognition accuracy of the model.(3)In order to solve the problem of individual differences in core brain networks in cross-subject mode,transfer learning is introduced,and on the basis of the two existing models,a core brain network extraction method based on domain adaptation and a classification method based on domain adaptation are proposed.In this thesis,the original two models are optimized,and the gradient inversion layer is used to reduce the feature distribution difference between the source domain and the target domain to realize feature migration.Compared with the original model without any migration method,the accuracy of recognition of SEED dataset in cross-subject mode of the migrated CC-SRM-GCN model is improved by 4.57%.