Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (5): 160-168.DOI: 10.3778/j.issn.1002-8331.2207-0372

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Self-Attention GAN for EEG Data Augmentation and Emotion Recognition

CHEN Jingxia, TANG Zhezhe, LIN Wentao, HU Kailei, XIE Jia   

  1. College of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xi’an 710021, China
  • Online:2023-03-01 Published:2023-03-01



  1. 陕西科技大学 电子信息与人工智能学院,西安 710021

Abstract: To solve the problem of data scarcity in emotion recognition based on electroencephalogram(EEG) and the resulting low accuracy of emotion classification, a conditional Wasserstein generative adversarial network(SA-cWGAN) with self attention mechanism is proposed. The self-attention module is used to learn long-term context related global features from training data through, and Wasserstein distance and Lipschitz constraint of gradient penalty are used to optimize the loss function of the network, so as to generate high-quality EEG data to enhance the original training set. A large number of comparative emotion classification are carried out on DEAP and SEED datasets, in which the proposed SA-cWGAN method is used to generate differential entropy(DE) and power spectral density(PSD) features that are close to the distribution of the input EEG training data, to augment the original EEG training dataset. SVM classifier is then used to classify the augmented features into different emotion categories. The experimental results show that in arousal and valence dimensions of DEAP dataset, the accuracy of the augmented DE and PSD features is 16.63, 17.55?percentage points and 6.48, 8.34?percentage points higher than that of the original DE and PSD features, respectively. On the SEED dataset, the accuracy of the three classifications has been improved by 4.64 and 5.18?percentage points respectively. It proves that the features generated by the proposed method have good robustness. The accuracy and stability of EEG emotion recognition can be effectively improved by augmenting original training dataset with the features generated by introducing self-attention mechanism to GAN network.

Key words: electroencephalogram(EEG), emotion recognition, data augmentation, generative adversarial network(GAN), self-attention, conditional Wasserstein

摘要: 针对脑电信号(electroencephalogram,EEG)情绪识别中数据稀缺及由此导致的情感分类精度不高的问题,提出了一个引入自注意力机制的条件Wasserstein生成对抗网络(SA-cWGAN),通过自注意力模块从训练数据学习长时上下文相关的全局特征,采用Wasserstein距离和梯度惩罚的Lipschitz约束对网络的损失函数进行优化,进而生成高质量的EEG数据对原有训练集进行增强。所提方法分别在DEAP和SEED数据集上进行了大量的二分类和三分类对比实验,生成了与EEG训练数据分布接近的微分熵(DE)和功率谱密度(PSD)特征,以此来增强EEG训练数据集,采用SVM分类器对增强后的EEG特征进行情绪分类。实验结果表明,在DEAP数据集上的唤醒度和效价维度下,增强后的DE、PSD特征较原有DE、PSD特征二分类准确率分别提高了16.63、17.55个百分点和6.48、8.34个百分点;在SEED数据集下,三分类准确率分别提高了4.64、5.18个百分点,证明所提方法生成的特征具有良好的鲁棒性,也表明通过对GAN网络引入自注意力机制生成的特征增强原有训练数据集能够有效提高EEG情绪识别的准确率和稳定性。

关键词: 脑电信号(EEG), 情绪识别, 数据增强, 生成对抗网络(GAN), 自注意力, 条件Wasserstein