EEG Data Augmentation And Emotion Recognition Based On Generative Adversarial Network

Emotion is the comprehensive state of human feelings,thoughts and behaviors.Emotion recognition refers to the recognition of the corresponding emotional state through human behavior and physiological response.Accurate emotion recognition plays an important role in the field of human-computer interaction and artificial intelligence research.EEG signals are widely used in resear-ch related to emotion recognition because they can not be disguised in acquisition and has the characteristics of objectivity.Based on the publicly available EEG emotion recognition datasets DEAP and SEED,this paper extracts relevant features of EEG and builds generative adversarial network for data augmentation,exploring the influence of different networks on data augmentation.The aim is to improve the accuracy of machine to recognize human emotions,thus improving the interaction between human and machine.The specific research work includes the following three aspects:(1)A method of data augmentation is explored by inputting the extracted EEG features into the constructed conditional Wasserstein generative adversarial network based on gradient punishment,which is represented by cWGAN-GP.Differential entropy(DE)features are extracted from the DEAP dataset,and EEG tables and extracted features are input into the constructed cWGAN-GP network to generate DE features that were close to the distribution of EEG training data.Experimental results show that the accuracy of the enhanced DE feature binary classification is 0.1343 and 0.1318 higher than that of the original DE feature in the dimensions of arousal and valence of DEAP dataset,respectively,which proves that the features generated by this method are robust.(2)To address the problem of data scarcity in emotion recognition of EEG signals and the resulting low accuracy of emotion classification,a conditional Wasserstein generative adversarial network with self-attention mechanism is designed,which is represented by SA-cWGAN.The self-attention module is used to learn long-term context-dependent global features from the training data.The Lipschitz constraints of Wasserstein distance and gradient penalty are used to optimize the loss function of the network,and then high-quality EEG data is generated to enhance the original training set.Two and three classification comparison experiments are carried out on DEAP and SEED datasets,respectively,and the DE and power spectral density(PSD)features close to the EEG training data distribution were generated.SVM classifier is used to classify the emotion of the enhanced EEG features.The experimental results show that in the arousal and valence dimensions of the DEAP dataset,the classification accuracy of the enhanced DE and PSD features is 0.1663,0.0648 and 0.1755,0.0834 higher than that of the original DE and PSD features,respectively.In SEED dataset,the accuracy of the three classifications increased by 0.0464 and 0.0518 respectively,indicating that the accuracy and stability of EEG emotion recognition can be effectively improved by introducing the features generated by the self-attention mechanism into the generative adversarial network to enhance the original training dataset.(3)A conditional boundary equilibrium generative adversarial network is designed,which is represented by SA-Res-cBEGAN.The self-attention module is used to learn the long-term context-dependent global features information from EEG training data,and the convolutional network in generative adversarial network is replaced by the residual neural networks.In addition to increasing the network depth of effective training,the problem of gradient disappearing will not be generated,so that high-quality and stable truth-like data can be generated to enhance the original training set.Two and three classification comparison experiments are carried out on DEAP and SEED datasets,respectively,to generate DE and PSD features that are close to the distribution of EEG training data.SVM classifier is used to classify the enhanced EEG features.The experimental results show that the binary classification accuracy of the enhanced DE and PSD features is increased by 0.1829,0.0744 and 0.2003,0.0938,respectively,compared with the original DE and PSD features in the dimensions of arousal and valence of DEAP dataset.In SEED dataset,the accuracy of classification was increased by 0.0634 and 0.0704,respectively,which were better than the benchmark method and the existing augmentation model for EEG emotion recognition data.