Multi-branch Domain Adversarial Neural Network with dynamic weight allocation for multi-source EEG classification

To address challenges such as the strong non-stationarity and inter-subject distribution shifts of EEG data, as well as the limitations of conventional DANN-based methods in feature representation and multi-source domain adaptation, a Multi-branch Domain Adversarial Neural Network with Multi-scale Channel Attention (MBCA-DANN) is proposed. To enhance feature richness, a multi-scale channel attention Module (MSCA) is designed, which provides multi-scale features and adaptively adjusts the feature channel weights, improving the feature capture ability of the network. A multi-branch architecture is constructed by combining auxiliary Maximum Mean Discrepancy (MMD), domain discriminators, and label discriminators, ensuring optimal matching between the source and target domains. Furthermore, a multi-source domain method with dynamic weight allocation is introduced, enhancing classification performance and robustness. Experimental results demonstrate that the classification accuracy for single-source domain transfer on the MII and MIII datasets is 71.89% and 71.82%, respectively, while the multi-source domain transfer classification accuracy improves to 79.83% and 82.87%. The model achieves a classification accuracy of 98.69% on the fatigue detection dataset, outperforming all currently known state-of-the-art algorithms, validating its strong generalization ability and providing an effective solution for multi-source cross-subject EEG classification.