快速自注意力结合SRU的癫痫信号识别模型  

Epilepsy signal recognition model with fast self attention combined with SRU

作　　者：沈彦平 龙善丽 尚宇哲 范雪梅 马士民 SHEN Yanping;LONG Shanli;SHANG Yuzhe;FAN Xuemei;MA Shimin(East China Institute of Photo-Electronic Integrated Device,Suzhou 215000,China)

机构地区：[1]华东光电集成器件研究所,江苏苏州215000

出　　处：《电子设计工程》2025年第7期16-21,共6页Electronic Design Engineering

基　　金：国家重点研发计划重点专项(2023YFB4704100);兵器工业应用创新项目(D03258F3)。

摘　　要：针对目前癫痫脑电信号研究中存在循环网络计算效率低,缺乏对不同通道间依赖关系的学习等问题,提出一种快速自注意力结合SRU的癫痫信号识别模型。对癫痫原始数据进行滑动窗口分段,通过一维卷积网络提出原始信号中高维局部信号特征,输入到双向快速自注意力结合SRU模块进行信号序列建模,并捕捉不同通道间的信号依赖,用全局注意力机制计算隐状态权重,由线性层输出癫痫类型。在癫痫信号数据集CHB-MIT上的实验结果表明,该模型取得了96.0%的平均F1值,高于实验对比的优秀深度学习模型,双向快速自注意力结合简单循环单元特征捕捉能力更强。At present,the computational efficiency of recurrent networks in the study of epileptic EEG signals is low,and there is a lack of learning on the dependency relationships between different channels.A fast self attention combined with SRU epilepsy signal recognition model is proposed,which segments the original epilepsy data into sliding windows.A one-dimensional convolutional network is responsible for proposing high-dimensional local signal features in the original signal,and then inputs them into a bidirectional fast self attention combined with SRU module for signal sequence modeling and capturing signal dependencies between different channels.The global attention mechanism calculates hidden state weights,and outputs epilepsy types through linear layers.The experimental results on the epilepsy signal dataset CHB-MIT show that the model achieved an average F1 score of 96.0%,which is higher than the excellent deep learning models compared in the experiment.The bidirectional fast self attention combined with simple recurrent unit feature capture ability is stronger.

关 键 词：癫痫信号识别 快速自注意力 一维卷积 简单循环单元 全局注意力 

分 类 号：TN91[电子电信—通信与信息系统]