摘要
针对现有和差波束弹载雷达运动目标检测方法所需训练距离单元数量较大、实际检测性能较低的问题,该文提出了一种基于智能多分类和网络参数迁移学习的运动目标检测新方法,其基本思路为利用少量训练距离单元数据构建数据集对深度卷积神经网络进行训练,将待测距离单元数据分类为杂波类(即无目标类)和对应不同多普勒频率的目标类。考虑到利用实测数据进行在线训练所需的计算资源较多、时间较长,该文首先构建了和差波束弹载雷达运动目标检测的回波信号模型,并基于实测数据进行验证,用于产生仿真数据进行网络离线训练。针对现有典型卷积神经网络参数较多、复杂度较高、训练效率较低等问题,该文基于特征融合模块(FFM)和空间注意力模块(SAM)对DenseNet网络进行改进,构建了FFM-SAM-DenseNet智能多分类器。由于基于智能多分类的检测方法在对不同待测距离单元数据进行处理时需重新训练网络,其整体收敛时间较长、速度较低。为解决该问题,该文引入迁移学习策略,将不同待测距离单元所对应多分类器的网络参数进行共享,以加快所提方法的整体收敛速度。仿真和实测数据处理结果表明,该文所提方法可以基于少量训练距离单元数据,获得相比现有方法更优的运动目标检测性能。
this paper proposes a new detection method based on intelligent multiclassification and network parameter transfer learning.The proposed method uses a small set of training range cell data to construct a dataset for training a deep Convolutional Neural Network(CNN),which classifies data from the Range Cell Under Test(RCUT)into clutter(target-free)or target classes with different Doppler frequencies.To avoid the high computational cost and time associated with online training on measured data,an echo signal model is first established for moving target detection in the missile-borne sum-difference beam radar.This model is validated using measured data and subsequently used to generate simulated data for offline network training.In addition,to overcome common limitations of typical CNNs,such as large parameter sets,high computing complexity,and low training efficiency,this paper enhances the DenseNet architecture by incorporating a Feature Fusion Module(FFM)and a Spatial Attention Module(SAM),resulting in an improved FFM-SAM-DenseNet multiclassifier.Furthermore,conventional detection methods based on intelligent multiclassification require retraining the network when processing data from different RCUTs,leading to long convergence time and reduced efficiency.To solve this problem,transfer learning is introduced to share network parameters across multiclassifiers for different RCUTs,accelerating the overall convergence speed.Simulation and measured data show that,even with limited training range cell data,the proposed method achieves better moving target detection performance than existing typical methods.
作者
王赵隆
张小宽
冯为可
孙华龙
王童
WANG Zhaolong;ZHANG Xiaokuan;FENG Weike;SUN Hualong;WANG Tong(Air Defense and Antimissile School,Air Force Engineering University,Xi’an 710051,China)
出处
《雷达学报(中英文)》
北大核心
2025年第6期1430-1450,共21页
Journal of Radars
基金
陕西省创新能力支撑计划项目(2025ZC-KJXX-81)
陕西省教育厅青年创新团队科研计划项目(24JP221)
陕西省高校青年创新团队,国家自然科学基金(62201611,62301597,62301598)。
关键词
弹载雷达
和差波束
运动目标检测
卷积神经网络
迁移学习
Missile-borne radar
Sum and difference beams
Moving Target Detection(MTD)
Convolutional Neural Network(CNN)
Transfer learning
