基于残差模块和自注意力机制GAN的脑电信号增广方法  

EEG signal augmentation method based on generative adversarial network with ResBlock and self-attention machenism

作　　者：李明爱 彭伟民[1] LI Ming’ai;PENG Weimin(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Computational Intelligence and Intelligent System(Beijing University of Technology),Beijing 100124,China;Engineering Research Center of Digital Community,Ministry of Education(Beijing University of Technology),Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]北京市计算智能和智能系统重点实验室(北京工业大学),北京100124 [3]教育部数字社区工程研究中心(北京工业大学),北京100124

出　　处：《计算机应用》2022年第S01期80-86,共7页journal of Computer Applications

基　　金：国家自然科学基金资助项目(11832003,81471770)。

摘　　要：针对脑电信号(EEG)数据量过少的问题,提出一种基于残差模块(ResBlock)和自注意力(Self-Attention)机制的生成对抗网络(GAN),记为RBSAGAN。该模型首先对ResBlock进行改进,设计了Up ResBlock和Down ResBlock网络用于提取信号中不同尺度感受野的特征并对数据维度进行扩大和缩小;然后根据Self-Attention机制设计1D Self-Attention网络挖掘EEG中各离散时刻之间的时间相关性;最后通过生成器和判别器的对抗训练生成逼真的信号。该模型在公开的BCI Competition IV dataset 2a数据集进行了大量实验,结果表明,RBSAGAN具有生成接近于真实脑电信号样本的能力,并且将分类器1D卷积网络(CNN)的平均识别率提升至96.04%,可以为EEG数据增广任务提供参考。Concerning the problem that the amount of ElectroEncephaloGram(EEG)data is too small,a Generative Adversarial Network(GAN)based on ResBlock and Self-Attention mechanism is proposed,namely RBSAGAN.The model first improved the ResBlock by designing the Up ResBlock and Down ResBlock networks to extract the features of different receptive fields in the signal and to expand and reduce the data dimension;then,the 1D Self-Attention network was designed according to the Self-Attention mechanism to mine the temporal correlation between discrete moments in the EEG;finally,realistic signals were generated by adversarial training of generator and discriminator.The model was extensively experimented on the publicly available BCI Competition IV dataset 2a.Experimental results show that the proposed RBSAGAN has the ability to generate samples close to real EEG signals and improve the average recognition rate of the classifier 1D convolutional network(CNN)to 96.04%.The proposed method can provide a reference for EEG data augmentation tasks.

关 键 词：脑电信号 生成对抗网络 卷积网络 残差网络 自注意力机制 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]