FA-TSception: A Multi-Frequency Spatiotemporal Attention Network for Cross-Subject EEG-Based Emotion Recognition

针对跨学科EEG情绪识别中个体差异显著和时频特征泛化不足的挑战，本文提出了一种多频时空注意力网络FA-TSception。该模型创新性地整合了多频率自适应机制和高效的通道注意力，构建了一个基于TSeption多尺度时空架构的三级处理框架。多频动态时间层通过参数化比例因子生成自适应卷积核组，以精确匹配Alpha、Beta、Gamma等情绪相关频带的时频特征；非对称空间层结合半球卷积核提取前额叶和时间区域的空间激活模式；集成了高效的信道注意力模块(ECA)，实现了多频特征的自适应校准。DEAP数据集上的跨学科实验表明，FA-TSception在唤醒和效价维度上的平均分类准确率分别达到62.73%和60.12%。与TSception相比，它提高了1.16%，仅增加了5.6%的模型参数计数。FA-TSception不仅提高了跨个体EEG情绪识别的准确性，而且通过引入有效的注意力机制，同时保持相对稳定的模型参数数量，增强了模型识别情绪相关特征的能力。In response to the challenges of significant individual differences and insufficient generalization of time-frequency features in interdisciplinary EEG emotion recognition, this paper proposes a multi-frequency spatiotemporal attention network FA-TSception. This model innovatively integrates multi-frequency adaptive mechanisms and efficient channel attention, constructing a three-level processing framework based on TSeption multi-scale spatiotemporal architecture. Multi-frequency dynamic time layers generate adaptive convolution kernels by parameterizing scaling factors to accurately match the time-frequency features of emotion-related frequency bands such as Alpha, Beta, Gamma, etc.; use asymmetric spatial layers combined with hemispherical convolution kernels to extract spatial activation patterns in the frontal lobe and temporal regions; integrate with an efficient Channel Attention Module (ECA) to achieve adaptive calibration of multi-frequency features. Interdisciplinary experiments on the DEAP dataset showed that the average classification accuracy of FA-TSception in the arousal and valence dimensions reached 62.73% and 60.12%, respectively. Compared with TSeption, it improved by 1.16% and only increased the model parameter count by 5.6%. FA-TSception not only improves the accuracy of cross-individual EEG emotion recognition, but also enhances the model’s ability to recognize emotion-related features by introducing effective attention mechanisms while maintaining a relatively stable number of model parameters.