基于多阶段融合模型的运动想象脑电信号分类方法OA

To address the challenges of insufficient temporal dependency modeling and limited channel selection capability in motor imagery electroencephalography(MI-EEG)classification,this paper designs a deep neural network architecture based on multi-stage feature fusion.The proposed model integrates time-frequency features,channel attention information,and temporal dependencies across multiple stages:a one-dimensional convolutional module inspired by EEGNet extracts local time-frequency representations;a multi-layer Transformer encoder captures global temporal dependencies;an SE(Squeeze-and-Excitation)block enhances responses from key brain regions;followed by 1×1 convolutions and residual connections for spatial feature mapping and fusion;and finally,a temporal convolutional network(TCN)performs deep-level temporal modeling and decoding output.In within-subject experiments on the BCI Competition IV-2a and BCI Competition IV-2b datasets,the model achieved average classification accuracies of 82.75%and 87.82%,respectively.In comparison with established methods such as ShallowConvNet,DeepConvNet,EEGNet,MI-CAT and CTNet,the proposed model demonstrates superior accuracy in within-subject MI-EEG classification tasks,validating the effectiveness and generalization ability of multi-stage fusion mechanism in time-frequency feature modeling.