统一全局空间表达的脑电信号跨被试情感识别

Unified Global Spatial Representation for EEG Subject-Independent Emotion Recognition

脑电信号(ElectroEncephaloGram, EEG)的跨被试情感识别,充分利用EEG信号库中大规模信息,并避免单被试模型训练对被试数据过度依赖产生的模型失效等问题,进一步推广了脑电识别的广泛应用.然而,不同被试生理与心理等客观差异加剧了模型建立难度.基于此,本文提出统一全局空间表达(Unified Global Spatial Representation, UGSR)的跨被试识别模型.本文构建自适应在线自编码网络,通过对时序数据增量学习,提取EEG信号潜在统一特征,实现生理偏差校正.进一步,本文利用格拉姆角场(Gramian Angular Fields, GAF)转换局部时序特征为全局连续空间表达,避免相同环境下因被试心理差异产生反应信号时序不一致等问题,并建立全局注意力机制的深度卷积神经网络,获得更具判别性的非线性样本表达,提升识别精度.本文模型被验证在流行的脑电信号数据集上,并获得了更好的跨被试识别精度与泛化性.

Electroencephalogram(EEG)subject-independent emotion recognition fully utilizes the built database of EEG,and avoids models of depending so heavily on training subjects.However,subject-independent emotion recognition suffers from the fairly low accuracy and generalization due to subjects born with individual difference in physical and psy-chological.To address above challenges,this paper proposes the unified global spatial representation model(UGSR).This paper presents self-adaption incremental auto-encoder network to obtain the latent unified features of all subjects without ground-truth to correct errors originating from physiological difference.Furthermore,this paper utilizes the gramian angu-lar fields(GAF)to transfer from local time-features to global spatial-features dealing with the semantic invalidation,more-over,exploits attention-CNN(Convolutional Neural Network)with the non-linear representation ability to extract the dis-criminate representation.The proposed model is verified in popular datasets,and achieves better performances than state-of-the-art methods.