Emotion Recognition Based On Granger Causality Between EEG Signals

Feature extraction is one of the key points for emotion recognition using EEG signals.Feature extraction generally include time domain,frequency domain,time and frequency domain,liner and nonliner methods.Most of the methods are extracting the global features of brain,but there are few methods involved in the interrelation and interaction between brain regions.In recent years,graph theory and the concept of brain network have attracted much attention in the field of emotion recognition based on EEG In this thesis,the Granger causality between EEG signals was used to describe the dynamic features of the brain network,and the causal information flow relationship between different regions of brain was used as feature for emotion recognition.When human's emotion changed,the EEG signals among different brain regions and frequency bands would interact and form causality information flow,and its feature should be the key point of emotion recognition.Doing binary classification in the three dimensions of valence,arousal and dominance to improve its effect further by using Ghranger causality coefficient differences between EEG signals of channels in different brain regions and different frequency bands as feature.We use the Deap data set which is international emotional database.First,the EEG signals are decomposed into four frequency bands which are theta,alpha,beta and gamma frequency band by wavelet packet transform and then the signals of each frequency band are denoised with double density dual-tree complex wavelet transform,and the causality values between the brain regions in the different frequency bands and band groups are calculated.The causality coefficients of each frequency band and frequency band group are considered as feature which is classified by SVM(support vector machine)respectively,and the recognition accuracies of feature of different frequency bands and groups are compared.According to the experimental data,the frequency band features with high recognition accuracy which is greater than 70%are sent to the classifier,and the emotion classification and recognition objects are independent subjects.The average recognition accuracy of 32 subjects is 96%.The recognition accuracy of four kinds of feature,such as power spectral density(PSD),asymmetric coefficient(AI),energy and entropy,is about 80%.Therefore,the application of Granger causality to the classification and recognition of the emotion based on EEG is an effective way to improve the accuracy of emotion recognition.Dynamic brain network research generally involves the amplitude,phase and frequency or other attributes of EEG signals.Therefore,the causal relationship between the phases of EEG signals in different brain regions was considered as the feature for emotion classification and recognition.The lasso regularization algorithm was used to normalize the Granger causality method.First,we decomposed the EEG signals into four frequency bands and denoised them.The instantaneous phases of EEG were calculated by Hilbert transform and then the causality between different brain regions in different frequency bands was calculated by lasso-Granger method.In order to avoid dimension disaster and be affected by redundant features,mRMR(minimal redundancy maximal relevance)algorithm was used to do feature selection.Finally,we used SVM to train and classify the sample data of individual subject.And the average accuracy rate is about 86%,which is higher than the recognition rate of using feature of PSD,AI,energy and entropy.The experiment results show that the causal relationship between the phases of EEG signals is also an effective method for emotion recognition.