Emotional EEG Recognition Research Based On Wavelet Packet And Hilbert-Huang Transform

As the most direct connection physiological signals with the brain activity, EEG (electroencephalographic) is a hot topic in fields of emotion research and human-computer interaction. Most modern interactive systems can’t effectively recognize human emotional states and drive machines to perform the correct action. The purpose of emotion study is more accurately and timely detecting a particular emotional state during the process of human-computer interaction and implementing corresponding valuable applications.Domain analysis method, frequency domain analysis method, time-frequency analysis method and nonlinear dynamics analysis method are all EEG analysis methods. Among them time-frequency analysis method becomes a powerful tool for non-stationary signals because that it provides the joint distribution information of the time domain and frequency domain. As time-frequency analysis method Hilbert-Huang Transform (HHT) and wavelet packet transform are widely used because of its unique charm. In order to improve emotion EEG recognition accuracy, this paper uses a public multi-modal emotional data sets named DEAP which is provided by Sander Koelstra as the research object. In this paper, the emotional EEG feature’s normalization manner is firstly studied, and then feature selection method which combines variance contribution rate with the F-score are studied. Finally, HHT and wavelet packet transform are used to extract EEG features to recognize valence, and these features’extracting time and classification accuracies are compared.The main tasks are as follows:(1) Emotion EEG feature values vary greatly between individual subjects. The influence of once normalization’s data range on classification accuracy is analyzed. This experiment uses six kinds of common normalization methods to compare the accuracies from three kinds of data range. The three kinds of data range of once normalization are all subjects’features, all features of single subject and single subject’s single feature. The results prove that single subject’s single feature is more appropriate as the once normalization’s data range of multiple subjects’EEG data.(2) Taking into account the problem that wavelet packet decomposition tree node energy has a large number of features, the feature selection method combining variance contribution rate with F-score are proposed. This method brings a significant reduction of the number of features without reducing the classification accuracies.(3) As time-frequency analysis method HHT and wavelet packet transform are more suitable for the analysis of the nonlinear and non-stationary EEG, multiple features are extracted using HHT and wavelet packet transform from emotional EEG data to recognizing valence, comparing them on the aspects of feature extracting time and classification accuracy. The results show that wavelet packet transform needs less time for extracting features than HHT. After the procedure of feature selection using the feature selection method combining variance contribution rate with F-score to reducing the number of features, IMF(Intrinsic Mode Function) energy moment percentage and IMF energy percentage can obtain the highest average classification accuracy up to84.38%, wavelet packet decomposition tree’s fifth layer’s nodes’energy get a highest classification accuracy of69.79%. The experiment shows that HHT has the advantage of recognition accuracy and disadvantage of computation time than wavelet packet transform.