Emotion Recognition Of EEG Signals Based On Self-attention Dynamic Graph Neural Network

In recent years,emotion recognition has become an important research field.Electroencephalogram(EEG)signals are one of the most robust cues for emotion recognition and reasoning due to their objectivity.The current popular EEG emotion recognition methods mainly extract the Spatio-temporal feature representation of a single channel through deep learning models such as convolutional neural networks(CNNs)and long short-term memory recurrent neural networks(LSTM-RNNs),and use multi-channel data fusion method for emotion modeling.However,CNN is not suitable for processing time-series data,LSTM-RNN is very time-consuming and it is difficult to avoid the problem of gradient explosion/vanishing during training.In addition,the various channels of EEG signals have complex spatial and temporal relationships in emotional expression.When extracting features of EEG signals,it is necessary to consider the spatial distribution information of EEG channels and their mutual influence.Therefore,Therefore,considering multi-channel Spatio-temporal information fusion of EEG feature representation plays an important role in improving EEG emotion recognition.Based on this,this paper uses multi-channel EEG signals to construct a brain network,and proposes a brain network representation learning method that employs self-attention dynamic graph neural networks(BNRL-SDGNNs)to obtain the feature representation of the spatial and temporal dimensions of EEG signals.First,a spectrogram representation of each channel's raw EEG signal is generated to capture its time and frequency information.Second,using a self-attention dynamic graph neural network,robust feature representations containing global brain network information are automatically learned.Finally,long-term dynamic temporal features of emotional responses are automatically acquired using a long short-term memory network(LSTM).In addition,this paper also uses a temporal convolutional neural networks(TCNs)to replace LSTM-RNN to increase training speed and avoid gradient explosion/vanishing.The experimental results show that the training speed of TCN is significantly faster and the gradient is more stable.The BNRL-SDGNNs model proposed in this paper has achieved good results,which are higher than the results of bidirectional LSTM and multi-channel fusion in each emotional dimension.This shows that the various channels of EEG signals influence each other,and when exploring changes in emotions,we cannot only consider a single channel.BNRL-SDGNNs fully consider various information in the feature extraction of EEG signals.