Research On Emotion Recognition Of EEG Signal Based On Deep Learning Algorithm

Objective: With the development of Artificial Intelligence and Brain Computer Interface(BCI)technology,computer recognition of human emotions has become an inevitable requirement for the development of education,medical treatment,traffic safety and other fields.Electroencephalography(EEG)is the most objective and realistic manifestation of human emotions,and is an important data source for emotion recognition research.The deep learning algorithm can mine emotional information from massive EEG data,eliminate the traditional feature extraction process relying on prior knowledge and improve the accuracy of emotion recognition by relying on end-to-end learning.Thus,this thesis focuses on the research of emotion recognition based on deep learning and EEG signals.Methods and results:(1)In this study a method of EEG signal preprocessing based on heterogeneous data is proposed.The method first Downsampling the data with higher sampling frequency to the same frequency as the data with lower sampling frequency,so as to unify the data frequency of heterogeneous data sets.Secondly,the original one-dimensional chain data is converted into a two-dimensional spatial layout matrix according to the electrode placement standards used in the datasets.Then normalize the non zero elements in each transformed matrix.Finally,based on the sentiment model used in the datasets,the label data of the heterogeneous datasets sentiment model is unified.(2)This thesis proposes a convolution neural network based on attention mechanism for emotional recognition of EEG data.The network uses an improved attention module based on grid to extract the effective information in the click matrix from the spatial dimension,introduces the Inception model to achieve multi-scale feature information extraction,and uses deep separable convolution to reduce the amount of parameters.This thesis demonstrates the batch size and model depth that are most suitable for emotional recognition of EEG signals through experiments.Experiments show that.The accuracy of the network in this thesis has reached 85.632%on heterogeneous data sets and 91.025% on SEED data sets,which is superior to the current mainstream convolution neural network model.(3)In order to fully utilize the spatiotemporal characteristics of EEG signals,this study proposes a hybrid neural network framework composed of convolutional neural networks and recurrent neural networks,and explores the differences in the final emotion recognition performance of the hybrid neural network framework using cascaded or parallel schemes.Experiments were conducted on two heterogeneous datasets,DEAP and SEED,and the hybrid neural network model of the parallel scheme achieved the highest classification accuracy of 91.373%.Through confusion matrix analysis,the classification accuracy of the proposed hybrid neural network model in the three emotions of positive,neutral and negative is improved compared with the traditional methods.Conclusion: In this thesis,a unified preprocessing method based on heterogeneous data of EEG signals has been proposed,which greatly reduces the differences between heterogeneous data sets by Downsampling,converting to two-dimensional matrix,unifying category labels and specifying electrode arrangement order.This study proposes a convolutional neural network model based on attention mechanism and a hybrid network model incorporating recurrent neural networks for emotion recognition of EEG signals,namely SIAUNet and SIAU-GRU.SIAUNet relies on attention mechanism and Inception module to achieve multi-scale spatial feature extraction of EEG signal features,and relies on deep separable convolution to achieve lightweight,improving the accuracy of emotion recognition and the generalization ability of the model.SIAU-GRU introduces GRU for temporal feature extraction and spatiotemporal feature fusion,enhancing the sensitivity of the model to emotional information.