基于条件双重对齐注意力U-Net的跨被试脑电情绪识别方法

Cross-Subject EEG Emotion Recognition Method Based on Conditional Dual-Aligned Attention U-Net

摘要: 针对跨被试脑电情绪识别中的个体差异与特征分布偏移问题，本文提出一种条件双重对齐注意力U-Net (CDA-AUNet)模型。该模型由脑电特征编码器与双重域对齐模块构成。特征编码阶段构建了融合压缩–激励(SE)通道注意力机制的U-Net架构，以实现对多频带和空间通道特征的自适应加权；并在深层引入信息瓶颈机制以抑制生理噪声、缓解过拟合。在跨域对齐阶段，模型结合条件域对抗网络(CDAN)与相关对齐(CORAL)算法，通过特征与预测概率的张量外积对齐联合条件分布，并在二阶协方差维度约束特征偏移，实现源域与目标域特征的深度协同对齐。在SEED数据集上的留一被试交叉验证结果表明，CDA-AUNet的平均分类准确率达到85.18%，显著优于现有主流基线模型。消融实验进一步验证了各核心模块在提升跨被试泛化能力中的有效性。

Abstract: To address the challenges of individual differences and feature distribution shifts in cross-subject EEG-based emotion recognition, this paper proposes a Conditional Dual-Aligned Attention U-Net (CDA-AUNet) model. The proposed model consists of an EEG feature encoder and a dual-domain alignment module. During the feature encoding phase, a U-Net architecture integrated with a Squeeze-and-Excitation (SE) channel attention mechanism is constructed to perform adaptive weighting across multi-band and spatial channel features. Furthermore, an information bottleneck mechanism is introduced in the deep layers to suppress physiological noise and mitigate overfitting. During the cross-domain alignment phase, the model integrates a Conditional Domain Adversarial Network (CDAN) with the Correlation Alignment (CORAL) algorithm. By utilizing the tensor outer product of deep features and prediction probabilities to align the joint conditional distribution, and constraining feature shifts at the second-order covariance dimension, the model achieves deep collaborative alignment of feature distributions between the source and target domains. Leave-one-subject-out (LOSO) cross-validation results on the SEED dataset demonstrate that the CDA-AUNet achieves an average classification accuracy of 85.18%, significantly outperforming current mainstream baseline models. Ablation studies further validate the effectiveness of the core modules in enhancing cross-subject generalization capabilities.