Abstract: To address the limitations of existing deep learning models in EEG signal decoding, this paper proposes an RCA-Conformer network architecture for motor imagery EEG classification. The architecture extracts local features through multi-scale temporal and spatial convolutions, incorporates a residual channel attention mechanism to further enhance convolutional feature representations, and utilizes a self-attention mechanism to capture global dependencies. In addition, by introducing residual connectivity, the features extracted by the convolution module and the Transformer encoder module are fused and directly inputted into the classifier for classification, thus significantly improving the classification performance of the model. In the subject-dependent evaluation, the architecture achieved 80.29% classification accuracy (Kappa = 0.7371) on the BCI Competition IV-2a dataset, and 85.74% classification accuracy (Kappa = 0.7148) on the BCI Competition IV-2b dataset. By showing excellent performance compared with existing methods, RCA-Conformer validates the effectiveness of the method and provides a new solution for the field of EEG decoding.