面向特定辐射源识别的强关联师生模型  

Strong Correlation Teacher-student Model for Specific Emitter Identification

作　　者：王艺卉 张志亮 于轩懿 徐芹 郭团奇 WANG Yihui;ZHANG Zhiliang;YU Xuanyi;XU Qin;GUO Tuanqi(Naval Aviation University,Yantai 264001;No.31401 Troops of PLA,Yantai 264001;No.32151 Troops of PLA,Xingtai 054000;No.92485 Troops of PLA,Laiyang 265200)

机构地区：[1]海军航空大学,烟台264001 [2]31401部队,烟台264001 [3]32151部队,邢台054000 [4]92485部队,莱阳265200

出　　处：《舰船电子工程》2025年第2期124-128,146,共6页Ship Electronic Engineering

摘　　要：针对特定辐射源识别网络模型复杂、高质量数据不足问题,提出了一种面向特定辐射源识别的强关联师生模型,旨在实现快速精准识别的同时压缩网络模型、节约计算资源,以实现平衡识别准确率和识别效率。首先对数据集进行重采样构建K折交叉验证数据集以最大程度上提供数据利用率,防止过拟合;然后,依据分支结构在训练过程与预测过程中的作用程度不同改进基于ResNet网络的识别网络模型,引入分组卷积的思想适当添加网络的多维特征提取能力,利用结构解耦重新参数化的方式去除分支结构,简化预测网络的卷积核构成;最后,利用原始网络和简化网络构建基于知识蒸馏的强关联师生模型架构,实现了在特定辐射源保证快速精准识别的同时网络模型的轻量化。To solve the problems of complex and insufficient high-quality data of specific radiation source recognition network model,a strong correlation teacher-student model for specific radiation source recognition is proposed,which aims to achieve fast and accurate recognition,compress the network model and save computing resources,so as to achieve a balance between recognition accuracy and recognition efficiency.Firstly,the data set is resampled to construct K-fold cross-validation data set to provide maximum data utilization and prevent overfitting.Then,the recognition network model based on ResNet network is improved according to the role of branch structure in the training process and the prediction process.The idea of grouping convolution is introduced to appropriately add the multi-dimensional feature extraction capability of the network,and the branch structure is removed by structural decoupling and re-parameterization to simplify the convolution kernel composition of the prediction network.Finally,a strong correlation teacher-student model based on knowledge distillation is constructed by using the original network and the simplified network,which realizes the lightweight of the network model while ensuring the rapid and accurate identification of specific radiation sources.

关 键 词：特定辐射源识别 交叉验证 知识蒸馏 分组卷积 参数重构 

分 类 号：TN911.7[电子电信—通信与信息系统]