EEG Emotion Recognition Based On Group-Feature Correlation

Emotion is the emotional and psychological state produced by people’s activities,which affects the normal life of individuals and the stable development of society.With the continuous enhancement of social individual consciousness and the maturity of braincomputer Interface(BCI),human computer interaction(HCI)and other technologies.Electroencephalogram(EEG)has been widely used in medical diagnosis or treatment and engineering research,as a physiological signal that can intuitively and objectively express emotional states.Emotional computing has also become a hot task in the field of artificial intelligence.Based on the analysis of EEG signals,this paper focuses on emotion pattern classification from two aspects: feature extraction and feature selection.The innovative research results are summarized as follows:(1)According to the characteristics of EEG signal,such as weak,non-stationary,multichannel and easy to be disturbed,this paper analyzes the linear and nonlinear features of EEG signal in time domain,frequency domain and time-frequency domain,and preliminarily finds that the frequency domain signal has the characteristics of strong representation and low time complexity.On this basis,combined with the relevant knowledge of entropy theory,a new feature called Relative Intensity Ratio Entropy(RIRE)is proposed.(2)On account of the problem that complexity of EEG mechanism and richness of information,the accuracy of feature selection is greatly reduced by traditional dimensionality reduction methods,this paper combines greedy algorithm,recursive idea and max relevance and min redundancy algorithm(m RMR),a feature selection method named Greedy-m RMR is proposed,which ensures the accuracy of the results while reducing the dimensions.(3)Considering the correlation between features,it is easy to select a large number of centralized and localized features one by one.This paper combines the idea of clustering,adopts "clustering + selection" strategy and designed a feature selection algorithm called Kmeans-group,fully explore the joint effect between features.Therefore,the mechanism of "clustering before selection" is adopted to select features in each category.In view of this,Kmeans-group algorithm can select the dominant features evenly distributed in the whole space,and give consideration to the effect of dimensionality reduction and recognition accuracy.(4)Experimental verification is carried out on DEAP data set in this paper and results show that using RIRE feature and support vector machines(SVM)classification,relying on the SAM model,the accuracy in Valence and Arousal reaches 90.79% and 91.18%,respectively,which is superior to most of the current traditional features in time domain,frequency domain and time-frequency domain.By Greedy-m RMR algorithm,the classification accuracy can reach 91.16% and 91.70% respectively while the feature dimensions is reduced by 14% and 16%.By Kmeans-group algorithm,the classification accuracy can still reach87.29% and 88.40% respectively while the feature dimension is reduced by 69.4%,which is higher than that of m RMR algorithm in the same dimensions.The results of this paper can be used for reference for affective computing based on EEG signals.The proposed features and algorithms have good expansion effect,and can be extended to the field of machine learning and HCI.