基于SSL-DDPM的脑电疲劳状态检测方法  

EEG fatigue state detection method based on SSL-DDPM

作　　者：张麟华 郭彩萍 许骁哲 富丽贞[3] 邢珍珍 ZHANG Linhua;GUO Caiping;XU Xiaozhe;FU Lizhen;XING Zhenzhen(Department of Computer Engineering,Taiyuan Institute of Technology,Taiyuan 030008,China;Shanxi Cultural Tourism Group Shantou Information Co.,Ltd.,Taiyuan 030008,China;School of Software,North University of China,Taiyuan 030008,China)

机构地区：[1]太原工业学院计算机工程系,山西太原030008 [2]山西文旅集团山投信息有限公司,山西太原030008 [3]中北大学软件学院,山西太原030008

出　　处：《现代电子技术》2024年第21期40-45,共6页Modern Electronics Technique

基　　金：国家自然科学基金青年科学基金资助项目(61602427);山西省科技厅重点研发计划(201903D121171);山西省教育厅科技创新项目(2023L352)。

摘　　要：疲劳检测对日常生活是至关重要的,尤其对于驾驶领域。基于脑电(EEG)信号的疲劳驾驶检测已吸引了众多学者的关注,但由于高质量带标签的EEG样本稀少问题严重阻碍了疲劳检测领域的发展。因此,文中首次将自监督学习(SSL)与扩散模型(DDPM)相结合应用于EEG的疲劳检测研究中,提出一种基于SSL-DDPM的脑电疲劳状态检测方法。该方法分为预训练和下游任务两部分,预训练阶段中首先对原始信号进行DDPM扩增,然后以ResNeXt代替ResNet为骨干网络对扩增前后的EEG信号进行特征提取,最后对提取的特征进行信号重构。下游任务的网络以共享预训练网络参数为主,对扩增前后的信号进行疲劳检测。通过SEED数据集和Multi-channel数据集进行实验验证,最终分类准确率分别达到88.23%和86.14%,验证了文中疲劳状态检测方法的有效性。Fatigue detection is vital to daily life,especially to the field of driving.Fatigue driving detection based on EEG signals has attracted much attention.However,the lack of high-quality labeled EEG samples has hindered the development of fatigue detection seriously.Therefore,a self-supervised learning(SSL)combined with diffusion model(abbreviated as DDPM)is applied to the fatigue detection of EEG for the first time,and an EEG fatigue detection method based on SSL-DDPM is proposed.The method is divided into two parts,named pre-training and downstream tasks.At the stage of pre-training,the original signal is amplified by DDPM,and then ResNeXt is used as the backbone network instead of ResNet to extract the features of EEG signals before and after amplification.Finally,the extracted features are subjected to signal reconstruction.The network for downstream tasks performs the signal fatigue detection before and after amplification by mainly sharing the pre-training network parameters.The SEED data set and Multi-channel data set are used for experimental verification,and the experimental results show that the final classification accuracy rates on the two data sets reach 88.23%and 86.14%,respectively,which verifies the validity of the fatigue state detection method in this paper.

关 键 词：脑电信号 疲劳检测 自监督学习 扩散模型 骨干网络 信号重构 

分 类 号：TN911.7-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]