基于MCSANet網(wǎng)絡的運動想象腦電分類

打開文本圖片集

中圖分類號：TN929.5-34；TP391 文獻標識碼：A 文章編號：1004-373X（2025）16-0067-08

MotorimageryEEGclassificationbasedonMCSANet

DU Jiang1，2， BI Feng3 （1.SchoolofComputerScienceand Technology，Shenyang UniversityofChemicalTechnology，Shenyang11o142，China;

2.LiaonngProvincialKeyLaboratoryofIntellgentTechnologyforChemicalProcessIndustryhenyang1O142China;

3.SchoolofInformationEngineering，LiaodongUniversity，Dandong118Oo3，China）

Abstract：Inordertosolve theproblemof insuficient feature miningand insufficient utilization indecoding electroencephalography（EG）signalsbymeansofthetraditionaldeeplearningmethod，aeeplearningmodel，MANet，is proposed，whichcombinestheparalelmulti-scaletemporalconvolutionwithsliding windowtechnologyandatentionmchanism. Theparalelmulti-scaletemporalconvolutionissedtofectivelycapturethetmporalcharacteristicsandspatialchracteistics ofEEGsignalsatdiferenttimescales.Theslidingwindowslicingtechnologyisusedtodividethefeaturesequencesand increase the numberof sequence samples.The weights ofeach partofthe sequence samples areasignedandfused by means of themulti-headself-atentionmechanism，whichcanfurther highlightmorekeyfeatures.Thefullyconnectedlayerandthe SoftMaxlayerareusedtowork together，soastoperformin-depthlearningandaccurateclasificationforthecapturedfeatures. Inorder tovalidatetheperformaneofthemodel，anexhaustiveexperimentalanalysiswasperformedontheBCICIV-2adataset. Theexperimentalresultsshowthattheaverageclassificationaccuracyofallsubjectsisashighas81.69%，whichverfisthe effectivenessoftheproposedmethodinminingthedeeppotential featuresofEEGand improvingtheclasificationperformance of motor imagery EEG.

Keywords：braincomputer interface;EEGsignal;parallelmulti -scale temporal convolution；sliding window slicing technology;multi-head self-attentionmechanism;ablation experiment

0 引言

解碼人或動物大腦產(chǎn)生的腦電信號，實現(xiàn)腦與外部設備之間的信息交換[。（剩余11652字）