一种基于轻量化神经网络的电磁信号识别方法  

An Electromagnetic Signal Recognition Method Based on Lightweight Neural Network

作　　者：曾昕 钟立俊 杨玲 李京泽 ZENG Xin;ZHONG Lijun;YANG Ling;LI Jingze(Southwest China Research Institute of Electronic Equipment,Chengdu 610036,China)

机构地区：[1]中国电子科技集团公司第二十九研究所,成都610036

出　　处：《电子信息对抗技术》2024年第1期37-44,共8页Electronic Information Warfare Technology

摘　　要：针对电磁频谱管控领域中基于神经网络模型的电磁信号识别方法,占用内存、计算时间和传输资源消耗较大,导致难以边缘部署应用的问题,提出了一种基于轻量化神经网络的电磁信号识别方法。首先,对电磁I/Q路数据进行空白值去除、划窗切片、归一化、频域特征提取四个信号预处理步骤,接着训练残差神经网络模型对其进行分类识别,最后通过参数剪枝和聚类量化两个步骤完成网络轻量化。所提方法在实测信号判识准确率较之前变化较小的前提下,内存压缩率为4.65,相较于深度残差网络(Deep Residual Network,ResNet)计算时间加快61.15 s,表明该方法能够在达到高识别率的同时有效降低模型存储规模,在计算时间、计算量方面也具有优势。Aiming at the problem that the electromagnetic signal recognition method based on neural network model in the field of electromagnetic spectrum management and control has a large consumption of memory,calculation time and transmission resources,which makes it difficult to deploy applications at the edge,an electromagnetic signal recognition method based on lightweight neural network is proposed.Firstly,the electromagnetic I/Q data is subjected to four signal preprocessing steps:blank value removal,window slicing,normalization,and frequency domain feature extraction.Then the residual neural network model is trained for classification and recognition.Finally the network lightweight is completed through two steps:parameter pruning and cluster quantification.On the premise that the accuracy of the proposed method has no significant change compared with the previous,the memory compression rate is 4.65,which is 61.15 s faster than the deep residual network(ResNet).It shows that this method can effectively reduce the storage scale of the model while achieving high recognition rate,and has advantages in terms of calculation time and amount of calculation.

关 键 词：模型压缩 神经网络 电磁信号识别 边缘智能 电磁频谱管控 

分 类 号：TN974[电子电信—信号与信息处理]