Research Of EEG Emotion Recognition Based On Deep Feature Representation And Its Application

With the development of computer science and neuroscience,the Brain Computer Interface(BCI),which establishes the signaling pathway between brain and computer,has attracted more and more researchers’ attention.Due to the important influence of emotion on human life activities,affective computing is added into the BCI to form the Affect Brain Computer Interface(a BCI).Affect brain-computer interfaces have great application prospects in safe driving,medical treatment,education and other aspects.In this thesis,"EEG emotion classification method based on deep feature representation" is taken as the core of the problem,and an EEG emotion recognition framework based on brain functional network and convolutional neural network is proposed.A cross-subject emotion recognition method based on domain self-adaptation is proposed,and an emotional EEG database is constructed.The main research is as follows:1.EEG-based Emotion Recognition Using Convolutional Neural Network with Functional Connections.Due to recent advances in the field of neuroscience,the study of the network perspective of the human brain has been greatly improved.Brain network forms a fully connected network by forming functional connections from multi-channel EEG,thus obtaining the functional characteristics of EEG in emotional state.In this thesis,the EEG signals were divided into 5 frequency bands,and then the phase synchronization index of each channel was calculated respectively.The phase synchronization index included the information of whether the physiological performance of the channels was synchronized.Then an emotion recognition model is constructed based on the ability of convolutional neural network to mine spatial information.By comparing with the other four recent EEG analytical models on DEAP and SEED,the proposed method achieves the highest recognition accuracy and F1-score.2.A cross subject emotion recognition method based on domain adaptation.Due to the nonstationary characteristics and individual differences of EEG,the generalization ability of existing models among subjects is difficult to reach the ideal level,which affects the promotion of EEG emotion recognition in practical application.In this thesis,the labeled EEG data is regarded as the source domain,and the unlabeled EEG data is regarded as the target domain.While minimizing the classification errors of the source domain,the dynamic domain adaptive method and the generative confrontation network training method in transfer learning are used to reduce the potential representation differences between the source domain and the target domain.Compared with other methods,the proposed method achieves the optimal results.3.An emotional EEG database was constructed by designing EEG emotional stimulation trials.On this dataset,the performance of the domain-basepd adaptive crosssubject emotion recognition method proposed in this thesis was validated.