摘要
目的运动想象脑电信号(motor imagery electroencephalography,MI-EEG)是脑机接口(brain-computer interfaces,BCI)中的一种经典范式,而数据量不足问题制约着它的研究与应用。随着深度学习的发展,生成对抗网络(generative adversarial networks,GAN)成为一种有潜力的数据增强方法,但其却面临着生成数据特征难以自动控制和数据多样性不佳的困难。为此,本文提出了一种基于信息最大化生成对抗网络(information maximization GAN,InfoGAN)的MI-EEG数据增强方法。方法首先,针对MI-EEG的多导联时间序列特点,设计基于深度卷积的InfoGAN(deep convolution information maximizing GAN,DCIMGAN)。然后,基于BCI Competition IV 2a/2b数据集,使用3种经典深度学习模型(EEGNet、FBMSNet和MI-BMInet)在有无生成数据参与训练的条件下进行性能对比分析,以评估生成数据对分类模型训练的影响。最后,以EEGNet作为基础分类模型,进一步对比了本文提出的增强方法与现有的多种数据增强策略(包括随机噪声添加、时间偏移、幅值缩放和信号翻转)在分类性能上的差异,探讨不同数据增强方法的优劣性。结果DCIMGAN由一个生成器G及两个鉴别器D和Q组成。G的输入为随机噪声和潜在变量,输出生成数据;D和Q的输入为真实数据和生成数据,分别用于判别样本真假和估计潜在变量的值。通过在训练过程中引入生成数据与潜在变量之间的互信息损失,实现对生成器的有效控制。增加DCIMGAN生成数据可使3种模型对于BCI Competition IV 2a/2b数据集的分类准确率分别提升3.51%/7.00%、2.55%/2.10%和3.92%/9.26%。结论DCIMGAN能够自动控制生成数据特征,获得与真实数据相似度高的MI-EEG数据,相对常用数据增强方法更有效提升分类模型性能。
Objective Motor imagery electroencephalography(MI-EEG)is a classic paradigm in brain-computer interfaces,but the limited data available hinders its research and application.With the development of deep learning,generative adversarial networks(GANs)have emerged as a potential data augmentation method,though they face challenges in controlling the generated data features and ensuring data diversity.To address these issues,this paper proposes a motor imagery(MI)EEG data augmentation method based on an information maximization GAN(InfoGAN).Methods Firstly,based on the multi-lead time series characteristics of MI-EEG,an InfoGAN(deep convolution information maximizing GAN,DCIMGAN)was designed.Then,based on the BCI Competition IV 2a/2b datasets,three classic deep learning models(EEGNet,FBMSNet,and MI-BMInet)were used to conduct a comparative analysis of performance with and without generated data in training,in order to evaluate the impact of generated data on the training of classification models.Finally,using EEGNet as the basic classification model,a further comparison was made between the proposed enhancement method and various existing data augmentation strategies(including random noise addition,time shifting,amplitude scaling,and signal flipping)in terms of classification performance,to explore the advantages and disadvantages of different data augmentation methods.Results DCIMGAN consisted of a generator(G)and two discriminators(D and Q).G taked random noise and latent variables as inputs to generate data,while D and Q differentiate between real and generated samples and estimate the values of latent variables,respectively.Mutual information loss between the generated data and latent variables was introduced during training to effectively control the generator.The inclusion of DCIMGAN-generated data improved the average classification accuracy of three classic deep learning models(EEGNet,FBMSNet,and MI-BMInet)by 3.51%/7.00%,2.55%/2.10%,and 3.92%/9.26%,respectively.Conclusions DCIMGAN can automatically control the features of generated data,producing MI-EEG signals with high similarity to real data,and is more effective than commonly used data augmentation methods in improving classification model performance.
作者
肖楠
李明爱
XIAO Nan;LI Ming’ai(School of Information Science And Technology,Beijing University of Technology,Beijing 100124;Beijing Key Laboratory of Computational Intelligence and Intelligent System,Beijing 100124;Engineering Research Center of Digital Community,Ministyr of Education,Beijing 100124)
出处
《北京生物医学工程》
2025年第5期449-456,共8页
Beijing Biomedical Engineering
基金
国家自然科学基金(62173010)资助。
关键词
运动想象信号
生成对抗网络
深度卷积神经网络
数据增强
motor imagery signals
generative adversarial networks
deep convolutional neural networks
data augmentation
