Emotion EEG Recognition Based On Convolutional Sparse Autoencoder

Emotion plays an important role in our life.Bad emotion may lead to many serious consequences.Therefore,identifying a person's emotional state is of great significance to individuals and society.In recent years,with the rapid development of artificial intelligence technology,adding machine learning and deep learning algorithm to emotion recognition based on EEG can achieve higher recognition accuracy than traditional signal analysis methods.However,due to a large number of noise signals in EEG signals,the accuracy of emotion recognition obtained in many studies is not high.At the same time,a large number of studies can only identify whether the subjects have an emotional state,and can not identify the specific reaction degree of the emotional state.Based on this,the main research contents of this thesis are as follows:(1)Firstly,during feature selection,select the EEG acquisition channel related to emotional EEG,calculate the power characteristics of each frequency band of the specified channel through Fast Fourier Transform(FFT),and use the power characteristics to represent the emotional EEG signal.Secondly,this thesis has done a number of ten kinds of experiments,which are carried out around the identification of ten reaction degrees of Valence(pleasure-unpleasure)and ten reaction degrees of Arousal(excitement-unexcitement).(2)Two models of Convolutional Sparse Auto Encoder(CSAE)and Long Short-Term Memory(LSTM)are constructed to identify the response degree of Valence(pleasure-unpleasure)and Arousal(excitement-unexcitement).The experimental results show that the average accuracy of CSAE model is 90.2% and 90.8% respectively in identifying the response degree of Arousal(excitement-unexcitement)and Valence(pleasure-unpleasure).The average accuracy of LSTM model is 85.5% and 81.5% respectively in identifying the response degree of Arousal(excitement-unexcitement)and Valence(pleasure-unpleasure).The average accuracy of CSAE model is higher than that of LSTM network model,however,when using CSAE model for classification,there is over fitting phenomenon,while LSTM network model does not have over fitting phenomenon.(3)In order to improve the classification accuracy and avoid the phenomenon of over fitting,this thesis combines CSAE model and LSTM network model to construct CSAELSTM model.After the training of CSAE model learning input data features,the decoder part of CSAE is replaced by LSTM network model,and the final emotional EEG recognition is completed by LSTM network model.The experimental results show that the average accuracy of CSAE-LSTM model is 90.78% and 90.12% respectively when identifying the response degree of Arousal(excitement-unexcitement)and Valence(pleasure-unpleasure).The accuracy of CSAE-LSTM model is higher than that of LSTM network model,which is not different from that of CSAE model,and there is no fitting phenomenon.