Research On Emotional Feature Extraction And Classification Of EEG

Emotion is an important way of expressing external stimuli in the process of human interaction with the external environment,which affects all aspects of our life at all times.Recognizing human's emotional state accurately and applying it further in artificial intelligence could assist human life better.Therefore,in recent years,researches related to emotion recognition have captured an army of scholars' attention who are in the field of artificial intelligence.Most of researchers recognize emotion state through human's external behavior,such as facial expression,gestures and so on.Compared with external behavior,physiological feature could reflect changes of human's emotion more objectively.Human's physiological signals mainly include electroencephalogram(EEG),electromyogram(EMG),and electrooculogram(EOG),among which,EEG is related with emotion most closely.Related researches on emotion recognition based on EEG signals aim at extracting more effective feature and proposing more effective classification algorithms.However,the recognition accuracy of the current research is too low.Aiming at this problem,this thesis studies the emotion recognition based on EEG signals from three aspects: preprocessing,feature extraction and classification models.In addition,in order to ensure the real-time transmission of data and reduce development costs of hardware,this thesis would also study the channel selection.The specific work and achievements of this thesis are summarized as follows: 1.In terms of preprocessing,the EEG signals collected by the equipment are extremely weak,which also contain some artifacts and other noise.In this study,the collected signals are firstly removed through preprocessing,so as to obtain pure signal relatively.2.In terms of feature extraction,the problem of low recognition accuracy caused by the incomplete time-domain feature and time-frequency energy calculation method and the problem of unstable EEG signals are addressed,in this study,a feature extraction method of EEG signals based on the combination of short-term amplitude and equivalent substitute energy with asymmetric differential entropy are proposed,which reduces the large error caused by the energy calculation method,highlights the signal feature of the right brain,and thus improves the accuracy emotion recognition.3.In terms of classification algorithm,we adopt Light GBM,an ensemble learning model with excellent performance in many data mining competitions to carry out the final emotion recognition.Experiments on DEAP data set show that the accuracy of emotion recognition in this study reaches 98.16%,improves greatly compared with some previous studies.4.In terms of channel selection,EEG data is mostly multi-channel acquisition.Considering that not all channels are related to emotion recognition,and that too many channels would affect the transmission of data,we propose an improved channel selection based on Relief F algorithm.Experiments on the DEAP data set show that,on the premise of ensuring the accuracy of recognition,we select out 16 channels finally,halving the number of channel in the condition that there is no visibly decrease in recognition accuracy,and ensuring the real-time transmission of data.