结合双尺度动态卷积与时空注意力的运动想象分类OA

Motor imagery based on EEG signals,as a key link in brain computer interface,plays an important role in achieving human-computer interaction and auxiliary motor rehabilitation.However,due to the low signal-to-noise ratio and significant dif-ferences between individuals,the decoding performance is limited.Therefore,a decoding network based on two-scale convolu-tion selective dynamic fusion and channel-time separation gating attention,named DSCTANet,is proposed.First,initial feature extraction is performed on EEG signals by spatiotemporal deep convolution module.Then,local temporal features of different scales are extracted in parallel by dual-branch deep convolution,and a selective dynamic fusion mechanism is designed to real-ize weighted fusion of cross-scale features.Finally,time and channel separation attention mechanisms are constructed to opti-mize channel interaction and long-range time dependence respectively,strengthen global feature correlation,and adaptively fuse dual-path features by gating mechanism.Experiments on BCICompetition Ⅳ-2a and BCICompetition Ⅳ-2b data sets,the results show that the average accuracy of the proposed method reaches 80.59％in four-classification task and 85.10％in two-classification task,which is significantly improved compared with existing benchmark models.