一种基于深度学习残差网络的模糊函数赋型方法  

An Ambiguity Function Shaping Method Based on Deep Learning Residual Network

作　　者：肖相青 王元恺 胡进峰[1,2] 刘军 钟凯[1,2] 赵紫薇 李会勇 XIAO Xiangqing;WANG Yuankai;HU Jinfeng;LIU Jun;ZHONG Kai;ZHAO Ziwei;LI Huiyong(University of Electronic Science and Technology of China,Chengdu 611731,China;The Yangtze Delta Region Institute(Quzhou),University of Electronic Science and Technology of China,Quzhou 324000,China;The 41st Research Institute of China Electronics Technology Group Corporation,Qingdao 266555,China)

机构地区：[1]电子科技大学,四川成都611731 [2]电子科技大学长三角研究院(衢州),浙江衢州324000 [3]中国电子科技集团公司第四十一研究所,山东青岛266555

出　　处：《雷达科学与技术》2024年第6期613-619,627,共8页Radar Science and Technology

基　　金：国家自然科学基金(No.62231006);国家重点研发计划(No.2023YFF0717400);衢州市财政资助科研项目(No.2023D040,2023D009,2022D009,2022D013和2022D033);四川省科技计划项目(No:2023YFG0176)。

摘　　要：基于模糊函数(Ambiguity Function,AF)赋型的恒模波形设计是雷达系统中的一项关键技术。该问题可构造为一个非线性的复四次问题(NP-hard)。现有的方法可分为两类:第一类方法通过松弛方式来求解该问题,但不可避免地会引入近似误差;第二类方法直接求解该问题,但该类方法的参数选取较为困难。我们注意到深度神经网络是一个天然的非线性系统,与上述的非线性问题模型高度契合。因此,本文提出了一种基于深度学习残差网络的方法来对AF赋型,该方法不需要松弛操作以及复杂的参数选取。具体步骤为:1)将该问题转化为一个无约束的相位优化问题;2)将该无约束问题的非凸目标函数构造为网络的损失函数;3)使用残差网络直接优化波形的相位。仿真结果表明,所提方法的信干比(Signal-to-Interference Ratio, SIR)有显著提升并且有着更好的目标探测性能。Unimodular waveform design based on ambiguity function(AF)shaping is a crucial technique in radarsystems. This problem is formulated as a nonlinear complex quartic optimization problem(NP‑hard). The existing methodscan be classified into two categories: the first one solves the problem by relaxing the original problem, but inevitablyintroducing approximation errors;the second one solves the problem directly, but the selection of parameters in this categoryis difficult. We notice that deep neural network is a naturally nonlinear system that is highly compatible with thatnonlinear problem. Motivated by this, this paper proposes a method based on deep learning residual network for AFshaping without any relaxation or complex parameters selection. The specific steps are as follows: 1)The problem istransformed into an unconstrained phase optimization problem;2)The non‑convex objective function of the unconstrainedproblem is constructed as the loss function of the network;3)The residual network is used to directly optimizethe phase of the waveform. The simulation results show that the signal‑to‑interference ratio(SIR)of the proposed methodis improved significantly, and it has better target detection performance.

关 键 词：深度学习 模糊函数 残差网络 恒模约束 波形设计 

分 类 号：TN958.2[电子电信—信号与信息处理]