Research On Data Augmentation And Emotion Recognition Methods Based On EEG Signals

Emotion is a physiological and psychological state of people,which affects people’s thinking and behavior and plays a vital role in daily life.It has external manifestations such as facial expression and voice,and internal manifestations such as EEG and ECG.With the wide application of artificial intelligence in various fields,there are more and more studies on the internal manifestations of emotions.Among them,EEG signals are captured from the cerebral cortex,and physiological and psychological studies have also confirmed that the brain is related to human emotions.Therefore,emotion recognition based on EEG has become a hot research topic in recent years.From the perspective of human-computer interaction,the addition of emotion assessment and emotion recognition will make the machine more intelligent and provide a more humanized interaction.However,the data scarcity of EEG emotion and low accuracy of emotion recognition also hinders its development.In order to achieve the purpose of increasing the amount of emotion data and improving the accuracy of emotion recognition,this paper makes the following research:(1)Aiming at the problem of the small sample size of the EEG in emotion recognition dataset and the model training is difficult.This paper proposed an EEG emotion data augmentation model c VGDDNet.The model combines c VAE and c GAN and uses two discriminators and introduces a distance regularization term in the loss function to learn the feature distribution of real data and generate artificial samples for training set expansion.The PSD and DE feature maps of EEG are input into the model,the encoder learns the distribution of the EEG feature map in the latent space by using the latent vector,the distance regularization term improved the generator’s ability to generate various artificial samples,and the double discriminator control the authenticity of artificial samples.In order to verify the effectiveness of artificial samples,this paper designs a CNN-based classifier to conduct emotion recognition experiments in two datasets respectively.After data augmentation,the average accuracy of 82.56% and83.43% were achieved for Valence and Arousal dimension binary classification tasks of the DEAP dataset,and the average accuracy of 85.78% was achieved in the three-classification task of the SEED dataset.The experimental results show that the artificial samples generated by the c VGDDNet data augmentation model can effectively improve the accuracy of the emotion recognition.(2)Aiming at the problem of insufficient time and frequency feature mining of EEG.This paper proposed an EEG emotion recognition model 4D-CARNN based on CNN,LSTM and attention mechanism.Firstly,the EEG data is converted into a 4D representation of spatial-frequency-temporal,the CNN module with the frequency self-attention layer is used to extract spatial and frequency features,and the bidirectional LSTM integrated with the attention mechanism is used to obtain its temporal information.Finally,the emotion recognition and classification is completed.The model was experimented on two datasets,it achieved 88.16% and 88.74% average accuracy for Valence and Arousal dimension binary classification tasks of the DEAP dataset,and 89.36% average accuracy on the SEED dataset three-classification task,which outperforms some existing models.These experimental results show that the method helps to improve the extraction and recognition rate of EEG emotional frequency and time information.