Research On EEG Emotion Recognition Based On Transfer Learning

Emotion is closely related to People’s Daily life.With the development of artificial intelligence technology,emotion recognition research has received more and more attention.By collecting people’s external behavior and physiological parameters in different emotional responses,computers can automatically identify the changes of human emotion.And widely used in human-computer interaction,traffic safety,medical assistance,distance education and emotional counseling and other scenarios,emotional recognition research has important significance and great application value.Traditional non-physiological signals such as voice and facial expressions were used for emotion recognition,which was prone to errors in emotion recognition caused by the user’s intention to disguise the external expression.Electroencephalogram(EEG)of physiological signals can objectively reflect changes in emotional states.It has the advantages of low cost,high temporal resolution,high portability and simplicity,etc.,and has attracted wide attention.It is of great significance to promote the research on emotion recognition based on EEG.However,EEG signals are usually weak,easy to be doped with noise and non-stationary characteristics,leading to great differences in EEG signals between different subjects and even the EEG signals of the same subject in a long time span.The training set and test set used in traditional EEG emotion recognition come from the same subject and the same session.When the training set and test set come from different subjects or different sessions,the effect of emotion recognition is not ideal.Transfer learning can transfer knowledge from one domain to another to help build a better model for another domain.Adding transfer learning method to EEG emotion recognition can effectively improve the accuracy of EEG emotion recognition for different subjects and different sessions.This paper mainly focuses on subject-to-subject and session-to-session EEG emotion recognition,and designs an emotion recognition model based on sample transfer and feature transfer.The main contents are as follows:(1)As the source domain contains both data that are beneficial to target domain recognition and data that have a negative effect on target domain recognition,which leads to negative migration,this paper designs a sample selection method based on K-means based on the characteristics of sample distribution.The method of target domain data by k means clustering,clustering center to get more,is selected in the source close to the clustering center distance of point as the new source domain data,control the number of new source domain data and make all kinds of sample to keep balance,finally using screening new source domain data model trained classifier,classification of target domain.Because the new source domain data eliminates some sample points that are easy to cause negative migration of the recognition model,the classifier model trained by the new source domain data has a better classification and recognition result for the target domain than the model trained by the unfiltered source domain data.(2)Due to the difference of feature probability distribution between source and target domains,the classification recognition model trained by unaligned feature data in source domain has poor recognition effect on target domain.To solve this problem,this paper proposes a loss function that can align the probability distributions of two domains from the perspective of feature-based transfer learning.The Maximum Mean Discrepancy(MMD)is used to measure the Mean difference in the Reproducing Kernel Hilbert Space(RKHS)between the source domain and the target domain.By minimizing the Maximum Mean Discrepancy in the two domains,You can align the marginal probability distributions of these two domains.The local maximum mean Discrepancy(LMMD)is used to measure the difference of conditional probability distributions in the two domains.The maximum mean difference is calculated for the data of the same class in the two domains,and then the mean value of their squares is calculated.By adding MMD and LMMD with the traditional cross entropy loss by a certain coefficient,the loss function that can align the source domain and target domain data in the training model is obtained,and the loss function is used in the EEG emotion recognition model of stack autoencoder to improve the recognition performance of target domain and the stability of the model.The recognition model based on stack autoencoder can be pretrained layer by layer to make the model convergence easier.Stack from the encoder to depth model,join in the loss function of improving the learning elements migration,through the depth model and migration study together,using the deep learning ability to learn fully the characteristics of alignment of two fields,to achieve the purpose of cross domain identification,thus improve the classifier on the target domain of emotion recognition accuracy.In this paper,a subject-to-subject and session-to-session EEG emotion recognition experiment was carried out on SEED database.The results show that the sample selection method based on K-means and the emotion recognition model based on stack autoencoder both have good recognition performance for different subjects’ emotional EEG and different sessions’ emotional EEG.