Research On Cross-subject EEG-based Emotion Recognition Method

Emotion plays a vital role in human-to-human interaction,as well as in humancomputer interaction.When machines perceive human emotions,they can conduct more personalized and targeted interactions,and emotion recognition is a key issue.At present,EEG signal has been widely used in emotion recognition tasks due to its high temporal resolution.However,EEG-based emotion recognition is still affected by some factors,which hinder its practical application.Since the individual differences of EEG signals are large,most models are trained for specific subjects,and the generalization is poor when applied to new subjects.To this end,this paper proposes a deep learning model architecture based on multi-branch network and transfer learning and experiments on SJTU Emotion EEG Dataset(SEED).This paper mainly does the following work:(1)Based on the differences in the frequency bands and brain regions of EEG signals,the influence of various data organization forms on the performance of deep learning models is explored.Different brain regions and signal frequency bands contain different information in the emotional EEG signal analysis.This paper combines these two factors to determine the optimal data structure for subsequent experimental studies.On the single-subject dataset,compared with the original EEG signal,the accuracy of the adjusted data after input into the model training improved by 13.80%.(2)Aiming at the characteristics of task-state EEG signals,we focused on how to effectively extract the inherent background features and task awareness features of cross-subjects EEG signals.In the current work,researchers often ignore the background information and blindly extract the task information,which leads to the model overfitting in the cross-subject EEG-based emotion recognition task.In this paper,a multi-branch network(MBN)model is proposed,and a training strategy of label orthogonalization is designed to extract the above two features,thus improving the representation ability of the model for the cross-subject EEG signals.The MBN model achieves an average accuracy of 79.57% in the cross-subject EEG-based emotion recognition task without using new-subject data,which is higher than the performance of the optimal model in the existing literature.(3)Aiming at the problem of subdomain distribution deviation in the feature transfer of EEG signals cross-subjects,a multi-subdomain adversarial network(MSAN)model is proposed for cross-subject EEG-based emotion recognition.The MSAN uses adversarial training to model the differences in the subdomain data with different labels,and the recognition performance is further improved by reducing the intra-class distance and enlarging the inter-class distance.Finally,through the organic integration of the MBN model and the MSAN model,optimal recognition accuracy of 87.94% is achieved.Compared with the existing transfer learning algorithms,the proposed model achieves high recognition performance while performing more stable on different subjects.