基于注意力残差网络的重叠尖锋分类算法  

Overlapping Peak Classification Algorithm Based on Attention Residual Network

作　　者：余海翔 汪曼青 王录涛[1] YU Haixiang;WANG Manqing;WANG Lutao(School of Computer Science,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610225

出　　处：《软件导刊》2025年第4期62-68,共7页Software Guide

基　　金：四川省科技厅重点研发计划项目(2023YFG0018)。

摘　　要：在神经解码研究中,多个神经元同时放电会产生尖锋重叠,进而影响解码准确性,因而分类重叠尖锋是解码神经元集群活动的关键步骤。提出一种基于格拉姆角和场与注意力残差网络(GASF-CBAM-ResNet)的重叠尖锋分类算法,首先使用独立尖锋数据制作重叠尖锋模板训练集,接着采用格拉姆角和场方法将一维尖锋序列转换成二维图像,用于训练注意力残差网络模型,最后对独立尖锋和重叠尖锋进行分类识别。该方法在“Wave_clus”数据集上的平均准确率达92.737%,优于传统方法。结果表明,GASF-CBAM-ResNet不仅可以有效地区分重叠尖锋,还节省了大量时间和人力成本,能够为神经解码研究提供可靠支持。In the study of neural decoding,simultaneous firing of multiple neurons creates spike overlap,which affects the accuracy of decod‐ing.Therefore,the classification of overlapping spikes is critical in decoding the activity of neuronal clusters.This paper proposes an algorithm for overlapping spike classification based on Gramian Angular Summation Field and Attention Residual Network(GASF-CBAM-ResNet).First,the training set of overlapping spike templates is created using independent spikes.The Gramian angular summation field is then applied to convert the 1D spike signal into a 2D image,which is used as input to the attention residual network.Finally,both independent and overlap‐ping spikes are classified and identified.The average accuracy of the proposed network reaches 92.737%on the"Wave_clus"data sets,and outperforms the traditional methods.The results show that GASF-CBAM-ResNet not only discriminates overlapping spikes efficiently,but al‐so provides reliable support for neural decoding with significant time and labor savings.

关 键 词：重叠尖锋 格拉姆角和场 卷积块注意力 残差网络 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]