Research On Feature Extraction And Channel Selection Of EEG For Emotion Recognition

In recent years,the field of human-computer interaction has developed rapidly.It has become an important criterion for intelligence that to correctly identify and analyze human emotions.Compared with other physiological signals,EEG has become the emphasis of emotion recognition because of its high temporal resolution,difficulty in disguise and close connection with brain activity.On the one hand,the current research on EEG signals mostly uses traditional linear analysis methods.It is undeniable that they have a certain resolution,but they are weak against noise and insensitive to time signals.Therefore,a non-linear method,multi-scale permutation entropy algorithm,is proposed to explore EEG signals,which greatly improves the recognition of emotion-related EEG signals.On the other hand,researches with emotion recognition based on EEG signals are mostly developed through the analysis of full channel EEG signals.Although good results were achieved,there are many problems including high feature dimensions,correlation between features and redundancy within features in the actual experimental process.These have some bad effects on the results with emotion recognition.To address above issue,this thesis researches the channel optimization selections on emotion recognition based on different EEG features which provides a new idea for portable EEG devices.The main research contents of the thesis are as follows:(1)This thesis proposes a facial expression recognition method based on multi-scale permutation entropy.It is a nonlinear analysis method that combines multi-scale entropy and permutation entropy.This method not only satisfies the requirements for processing non-stationary random signals,but also has strong robustness and obvious distinguishing effects.In the experimental part,we explore the effect of scale factors on the performance of multi-scale permutation entropy.The experimental results show that the analysis of EEG signals based on this feature can effectively distinguish the EEG signals of facial expressions with the scale factor value of 2.(2)This thesis proposes a channel selection method for emotion recognition based on Relief-FGSBS,which combines the Relief algorithm with the Floating Generalized Sequence Backward Selection method.It can reduce the impact of EEG data redundancy on classification accuracy.First,the thesis extracts different types of valid EEG signal features to form preliminary data.Then,we analyze channels performance based on single and combined features.We select the optimal channel set based on the channel selection algorithm and the channel selection frequency.Finally,we use support vector machines to verify the effectiveness of the channel selection method.By comparing the optimal channel set with the random channel,we find that the performance of the optimal channel set is much higher than that of the random channel set.On the self-collected data set and the public data set,the optimal channel set we obtained has extremely high similarity.This shows that the proposed channel selection method is effective.