一种基于静息态欧氏空间对齐迁移学习的运动想象源域样本筛选方法  

A Source Domain Trial Selection Method for Motor Imagery Based on Resting-State Data Euclidean Space Alignment Transfer Learning

作　　者：楚超 祝磊 杨君婷 黄爱爱 张建海[2,3] CHU Chao;ZHU Lei;YANG Junting;HUANG Aiai;ZHANG Jianhai(School of Automation,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China;School of Computer Science and Technology,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China;Key Laboratory of Brain Machine Collaborative Intelligence of Zhejiang Province,Hangzhou Zhejiang 310018,China)

机构地区：[1]杭州电子科技大学自动化学院,浙江杭州310018 [2]杭州电子科技大学计算机学院,浙江杭州310018 [3]浙江省脑机协同智能重点实验室,浙江杭州310018

出　　处：《传感技术学报》2024年第6期997-1004,共8页Chinese Journal of Sensors and Actuators

基　　金：浙江省重点研发计划项目(2020C04009)。

摘　　要：针对脑电信号的非平稳性及校准过程复杂费时的问题,提出一种静息态欧氏空间对齐源域样本筛选方法。首先在欧氏空间中将受试者静息态数据与其运动想象信号进行对齐,以减少与运动想象任务无关的因素的干扰,然后依据欧氏距离度量准则剔除距离目标域样本较远的源域样本,进一步减少源域样本与目标域样本之间的差异,从而提升迁移学习的效果。在运动想象公开竞赛数据集BCI Competition 4的Dataset 1和Dataset 2a上分别获得了80.71%与74.46%的平均准确率,实验结果表明所提方法可以有效改善脑电信号的非平稳性,提高运动想象信号的分类正确率。To solve the problem of non-stationarity of EEG signal and complex and time-consuming calibration process,a source domain trial selection method based on Euclidean space rest data alignment is proposed.Firstly,the subjects’EEG trials are aligned with their rest-state data in Euclidean space to reduce the interference of factors unrelated to motor imagery task.Then,according to Euclidean distance measurement criteria,the source domain samples that are far away from the target domain samples are eliminated to further re-duce the difference between the source domain and the target domain samples,so as to improve the effect of transfer learning.The aver-age accuracy rates of 80.71%and 74.46%are obtained on two open motor imagery datasets of BCI Competition IV dataset1 and data-set2a,respectively.The experimental results show that the proposed method can effectively ameliorate the non-stationarity of EEG signal and improve the classification accuracy of motor imagery signals.

关 键 词：脑机接口 运动想象 迁移学习 域适应 个体差异 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]