基于双向时序和图卷积特征融合的脑电信号分类

打开文本图片集

EEGSignalClassificationBasedonBi-directional TimeSeries andGraph Convolutional Feature Fusion

Xu Yanjun， Li Xianhua，Du Pengfei， Zhang Kang （AnhuiUniversityofScienceand Technology，Huainan232oo1，China）

【Abstract】To enhance the synergy between time-domain and frequency-domain features in motor imagery EEG signal classification，the paper proposesatime-frequency feature fusionalgorithm basedon bidirectional temporal and graph convolutional networks.Themodel adopts adual-branch structure：the bidirectional temporal fusionmoduleextracts bidirectionalcontextual temporal features through multi-scaleconvolutional groupsandabidirectional long shortterm memory network （BiLSTM），combined with across-attention mechanism to reinforce key information;the graph convolutional moduleconstructs frequency-domain features generated from time-domain signals viaFourier transform intoan electrode topology graph，capturing local and global frequency-domain pattrns through node neighborhood informationaggregation.The time-domainand frequency-domain features output bythetwo modules are deeplyfused and then fed into a Softmax classifier.This method achieves a classification accuracy of 81.4% on the BCI Competition IVDataset 2a， outperforming benchmark methods such as FBCSP and FBCNet by 5.2% to 13.6% .Ablation experiments validate the effectiveness ofeach module.Thecollaborative designoftheattentionmechanismandBiLSTMsignificantly improves temporal modeling capabilities，providing new insights for EEG signal analysis.

【Key words】motion imagery; electroencephalogram（EEG）; graphical convolutional network; bidirectional long and short term memory networks; attention mechanism

[中图分类号]G64;06 [文献标识码]A

[文章编号]1674-3229（2025）03-0055-08

0 引言

运动想象脑电信号（MI-EEG）研究致力于构建通用的脑信号基础模型，以精确捕捉和分析脑电信号，从而深入揭示大脑工作机制。（剩余9942字）