多尺度时空卷积结合Transformer的脑电情绪分类算法研究  

Research on EEG Emotion Classification Algorithm Combining Multi-Scale Temporal-Spatial Convolution and Transformer

作　　者：钟献彪 王俊青 任飞廉 童澄达 侯振中 赵兴群 ZHONG Xianbiao;WANG Junqing;REN Feilian;TONG Chengda;HOU Zhenzhong;ZHAO Xingqun(School of Internet of Things Engineering,Wuxi Taihu University,Wuxi Jiangsu 214063,China;School of Biological Science&Medical Engineering,Southeast University,Nanjing Jiangsu 210096,China)

机构地区：[1]无锡太湖学院物联网工程学院,江苏无锡214063 [2]东南大学生物科学与医学工程学院,江苏南京210096

出　　处：《传感技术学报》2025年第1期104-111,共8页Chinese Journal of Sensors and Actuators

摘　　要：基于脑电的情绪识别分类算法通常需要大量电极通道以获得良好的分类性能,然而这限制了其在实际应用中的灵活性。基于多尺度时间卷积及空间卷积,结合通道注意力和Transformer编码器模块,提出了一种混合模型(MTSACT),能够基于少数通道的脑电信号进行情绪分类。提出的算法在DREAMER公开数据集和使用自研设备采集的数据集上开展了受试内、混合被试2种条件下的实验,在两个情绪维度上均取得95%以上的平均分类准确度。最后探索了EEG各频带的贡献度以及对模型进行可解释化工作,这些结果表明,利用额叶和颞叶的6个脑电通道足以实现高分类准确度。EEG-based emotion recognition classification algorithms typically necessitate a substantial number of electrode channels to attain satisfactory classification efficacy,yet this limits their flexibility in practical applications.A hybrid model(MTSACT)that employs multi-scale temporal and spatial convolutions,channel attention,and Transformer encoding is presented to classify emotions from EEG sig-nals with fewer channels.The algorithm proposed herein is subjected to experimental validation on the publicly available DREAMER data-set,as well as on a dataset amassed using bespoke devices,under two distinct experimental paradigms,i.e.,within-subject and cross-sub-ject.The model demonstrates an average classification accuracy exceeding 95%across two emotional dimensions.Furthermore,this study delves into the contributory significance of various EEG frequency bands and the interpretability of the model.The findings corroborate that the utilization of six EEG channels positioned in the frontal and temporal lobes is sufficient to yield high classification accuracy.

关 键 词：情绪识别 脑电 深度学习 少数通道 可视化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]