Research And Implementation Of EEG Emotion Recognition Algorithm Based On Deep Learning

Electroencephalogram（EEG）can objectively reflect the internal affection of human beings through the basic activity of the brain.Whereas,there are artifacts appearing in-evitably in the process of EEG acquisition,which will lead to crummy reception of EEG.Moreover the existing feature selection methods can not extract all the emotion-related channels or will extract emotion-irrelevant channels,that will lead to incorrect emotional signals of EEG.This thesis will take SEED dataset and DEAP datasete as research ob-jectives,then introduce the sparseness theory into EEG emotion classification,and finally analyze how the EEG reflect internal affection through the feature extraction and selection as well as classification algorithms.The main work of this thesis is as follows:（1）A sparse causality brain network based on L_p（p=0.5）regularization will be constructed,and a sparse EEG feature extraction method based on emotion weight will be proposed.Based on the logistic regression theory of machine learning,this thesis intro-duces L_p（p=0.5）sparse regularization for feature extraction of EEG signals for the first time.Experimental results show that using L_p（p=0.5）regularization is better than using L₁regularization and L₂regularization to select EEG signal channels and ignore irrelevant channels.Based on the above experimental results,a directed brain network is developed in this thesis.Brain networks can describe the connectivity of brain neurons and the flow of information.In this thesis,Granger causality theory is used to obtain the edge of the brain network by estimating the connectivity and information flow between channels.At the same time,a sparse theory based on L_p（p=0.5）regularization is introduced to delete invalid connections.In order to quantify the direct correlation between channels and emo-tions,this thesis added emotion weights to the Granger causality brain network analysis model to enhance the feature selection ability of the model.Experimental results show that the sparse Granger causality analysis model based on channel correlation proposed in this thesis has the best accuracy of emotion classification of EEG signals（86.58%）compared with the existing algorithms.（2）Taking convolutional neural network as the basic model,the structure of the net-work is modified to establish an algorithm for emotion recognition of EEG signals that integrates sparse theory and convolutional neural network.The function of sparsity is di-mensionality reduction.Convolutional neural network is commonly used in the field of image recognition,and the existing network model is not suitable for emotion classifi-cation of EEG signals.Although the EEG signal is relatively complex and the original data has a higher dimension,in fact,its effective information can be concentrated in a low-dimensional space.Therefore,a sparse convolutional neural network hybrid classifi-cation model（HCNNS）is proposed in this thesis.Experiments show that this model can not only optimize the network structure and reduce the number of parameters,but also greatly reduce the running time and improve the classification accuracy.（3）The online control robotic arm system based on emotional EEG recognition and classification is designed and implemented.The EEG signal recognition and classification were combined with the robot,and the emotional EEG signals were collected in real time through the EEG acquisition equipment,and then the emotion recognition and classifica-tion were carried out.Finally,the results of classification are transformed into instruction transmission to control the robotic arm and make it act accordingly.