基于迁移学习的跨被试脑电疲劳驾驶检测  

Transfer learning-based cross-subject EEG fatigue driving detection

作　　者：邱轶辉 江琼 魏玲玲[1] 张卫平 邱桃荣[1] QIU Yihui;JIANG Qiong;WEI Lingling;ZHANG Weiping;QIU Taorong(School of Mathematics and Computer Sciences,Nanchang University,Nanchang 330031,China;School of Computer and Big Data Science,Jiujiang University,Jiujiang,Jiangxi 332005,China)

机构地区：[1]南昌大学数学与计算机学院,江西南昌330031 [2]九江学院计算机与大数据科学学院,江西九江332005

出　　处：《南昌大学学报（理科版）》2023年第4期397-402,共6页Journal of Nanchang University(Natural Science)

基　　金：国家自然科学基金资助项目(61762045,61841201)。

摘　　要：在利用脑电信号进行疲劳驾驶跨被试检测中,克服脑电的个体差异是一项重大挑战,欧氏空间对齐是一种解决方法,然而该方法要求目标域上有大量数据。为降低跨被试检测中对目标域数据的依赖,我们提出一种基于模型迁移学习和改进欧式空间对齐的方法,以提高模型在有少量目标域数据时的分类能力。所提出的方法首先对源域数据进行欧式空间对齐以降低个体间差异,接着对目标域数据进行参考矩阵相似度加权平均对齐,使用深度卷积神经网络用于特征提取和分类,在源域上预训练后在目标域上微调。测试结果显示所提出的对齐方法能有效提高少量目标域数据可用时的跨被试分类准确率,最好的准确率达到96.12%。In cross-subject detection of fatigue driving with EEG signals,overcoming individual differences in EEG is a major challenge,and Euclidean spatial alignment is a solution,However,the method requires a large amount of data on the target domain,To reduce the reliance on target domain data in cross-subject detection,we propose a method based on model migration learning and improved Euclidean spatial alignment was proposed to improve detection when a small amount of target domain data was available capability.The proposed method first performed Euclidean spatial alignment on the source domain data to reduce inter-individual differences,followed by a reference matrix similarity-weighted average alignment on the target domain data,and used a deep convolutional neural network for feature extraction and classification,which is pre-trained on the source domain and then fine-tuned on the target domain.The results showed that the proposed alignment method can effectively improve the cross-subject classification accuracy with a small amount of target domain data,with the best recognition rate of 96.12%.

关 键 词：迁移学习 数据对齐 脑电信号 疲劳驾驶 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]