The Research Of EEG Emotion Recognition And Reconstruction Based On Neural Networks

The existing artificial intelligence system has considerable "logic intelligence" based on the strong computing power and storage structure,but lacks performance in "emotional intelligence".To achieve more advanced and comprehensive intelligence,researchers have begun to pay more attention to the emotional intelligence.Specifically,the emotion recognition is a critical part of emotional intelligence.At present,there are many ways to recognize emotions,and the EEG signal of the central nervous system is becoming an important means for emotion recognition and exploring the brain's information processing mechanism.Based on the public database named DEAP and the EMOT database collected by our laboratory,this thesis utilizes the GRU(Gated Recurrent Units)and Variational Auto-Encoder(VAE)to conduct emotion recognition and emotion reconstruction research.The main work includes as follows:(1)Since the EEG data is a typical non-stationary dynamic time series,this thesis proposed an EEG emotion recognition method based on the Bidirectional GRU(Bi GRU)recurrent neural network.The GRU network can learn important Time Dynamic Features(TDF)in EEG sequences due to its cyclical structure in the time dimension.At the same time,to use the information of EEG sequence as much as possible,we proposed to apply the bidirectional GRU structure to model the EEG data to improve the accuracy of emotion classification.The experimental results show that: 1)we obtained 74.14% average accuracy on the EMOT database with the common GRU network,and the accuracy in the Valence and Arousal dimensions of the DEAP database are 73.24% and 74.67% respectively;2)The Bi GRU network can make up for the information loss of the GRU in EEG sequence and effectively improve the accuracy of emotion recognition.(2)Considering the noise and high dimensionality of GRU in feature extraction,this thesis proposed an EEG emotion recognition method based on hybrid network of the GRU and VAE.It mainly uses VAE to execute feature extraction and nonlinear dimensionality reduction on TDF.After that,we used the SVM to classify features.The experimental results show that the hybrid network based on GRU-VAE can remove the noise of features and improve the accuracy of emotion recognition to some extent,and the low-dimensional feature vector representation of VAE learning can be more effective than other dimensionality reduction methods such as LDA,KPCA and ISOMAP.(3)To explore the problem of decoding and visualizing emotions from brain activity,this thesis proposed a method of EEG emotional reconstruction based on Conditional VAE(Conditional Variational Auto-Encoder,CVAE),which aims to generate the pictures in the human brain with different emotions.VAE model can fit the distribution of data without supervision,but to reconstruct the emotional content of human brain from EEG,we need to introduce the conditionalization of VAE.Based on the EEG evoked by emotional image,the power spectrum feature is firstly extracted as the specific condition of emotional reconstruction,and then sent to CVAE together with emotional image for modeling.Then,the encoder of CVAE is used to learn the brain's potential representation of different emotional images.Finally,through the sampling of latent space,the decoder of CVAE can generate the corresponding emotional image.The experimental results verified the possibility of reconstructing emotions from EEG.We conducted all the experiments on the EEG database evoked by video and image,and the results show that the GRU and VAE can be effectively applied to emotion recognition tasks.The power spectrum feature of EEG can reflect the brain's clues to the processing of emotional information.At the same time,the conditional VAE is also an effective information reconstruction method in exploring the brain activity decoding problems.