西安电子科技大学学报 ›› 2023, Vol. 50 ›› Issue (6): 62-74.doi: 10.19665/j.issn1001-2400.20231005

• 电磁空间安全专栏 • 上一篇    下一篇

面向雷达行为识别的多尺度卷积注意力网络

熊敬伟(),潘继飞(),毕大平(),杜明洋()   

  1. 国防科技大学 电子对抗学院,安徽 合肥 230000
  • 收稿日期:2023-03-07 出版日期:2023-12-20 发布日期:2024-01-22
  • 通讯作者: 潘继飞(1978—),男,教授,E-mail:panjifei17@nudt.edu.cn
  • 作者简介:熊敬伟(1995—),男,国防科技大学硕士研究生,E-mail:xiongjingwei1995@nudt.edu.cn;|毕大平(1965—),男,教授,E-mail:bdpeei@163.com;|杜明洋(1994—),男,讲师,E-mail:dumingyang17@nudt.edu.cn
  • 基金资助:
    国家自然科学基金(62071476)

Multi-scale convolutional attention network for radar behavior recognition

XIONG Jingwei(),PAN Jifei(),BI Daping(),DU Mingyang()   

  1. College of Electronic Engineering,National University of Defense Technology,Hefei 230000,China
  • Received:2023-03-07 Online:2023-12-20 Published:2024-01-22

摘要:

针对低信噪比条件下雷达信号特征提取难、识别稳定性低的问题,提出了一种基于深度分组卷积、多尺度卷积和自注意力机制的雷达行为模式识别算法,在不增加训练难度的情况下提高复杂环境下的识别能力。该算法先采用深度分组卷积在浅层网络分离弱相关的通道,再通过多尺度卷积代替常规卷积实现多维特征的提取,最后利用自注意力机制调整优化不同特征图的权值,抑制低相关和负相关的通道与空间带来的影响。对比实验表明,所提MSCANet在0~50%丢失脉冲和虚假脉冲条件下平均识别率达到约92.25%,与基线网络AlexNet、ConvNet、ResNet、VGGNet相比,准确率提升了约5%~20%,不同雷达行为模式识别稳定,模型具有更好的泛化性和鲁棒性。同时,消融实验证明了深度分组卷积、多尺度卷积和自注意力机制对模式识别的有效性。

关键词: 深度学习, 机器学习, 模式识别, 深度分组卷积, 多尺度卷积, 自注意力机制

Abstract:

A radar behavior mode recognition framework is proposed aiming at the problems of difficult feature extraction and low recognition stability of the radar signal under a low signal-to-noise ratio,which is based on depth-wise convolution,multi-scale convolution and the self-attention mechanism.It improves the recognition ability in complex environment without increasing the difficulty of training.This algorithm employs depth-wise convolution to segregate weakly correlated channels in the shallow network.Subsequently,it utilizes multi-scale convolution to replace conventional convolution for multi-dimensional feature extraction.Finally,it employs a self-attention mechanism to adjust and optimize the weights of different feature maps,thus suppressing the influence of low and negative correlations in both channels and the spatial domains.Comparative experiments demonstrate that the proposed MSCANet achieves an average recognition rate of 92.25% under conditions of 0~50% missing pulses and false pulses.Compared to baseline networks such as AlexNet,ConvNet,ResNet,and VGGNet,the accuracy has been improved by 5% to 20%.The model exhibits stable recognition of various radar patterns and demonstrates enhanced generalization and robustness.Simultaneously,ablation experiments confirm the effectiveness of deep grouped convolution,multi-scale convolution,and the self-attention mechanism for radar behavior recognition.

Key words: deep learning, machine learning, mode recognition, depth-wise convolution, multiscale convolution, self-attention mechanism

中图分类号: 

  • TN971