现代防御技术 ›› 2026, Vol. 54 ›› Issue (2): 128-136.DOI: 10.3969/j.issn.1009-086x.2026.02.012

• 论文 • 上一篇    

样本对齐与半全局注意力机制的个体识别算法

张志强, 胡进, 许金鑫, 吴云松   

  1. 中国船舶集团有限公司 第七二四研究所,江苏 南京 211153
  • 收稿日期:2025-04-14 修回日期:2025-06-05 出版日期:2026-04-28 发布日期:2026-04-30
  • 通讯作者: 胡进
  • 作者简介:张志强(2002-),男,湖北黄冈人。硕士生,研究方向为信息与信号处理。

Specific Emitter Identification Algorithm Based on Sample Alignment and Semi-global Attention Mechanism

Zhiqiang ZHANG, Jin HU, Jinxin XU, Yunsong WU   

  1. The 724th Research Institute of China State Shipbuilding Corporation Limited,Nanjing 211153,China
  • Received:2025-04-14 Revised:2025-06-05 Online:2026-04-28 Published:2026-04-30
  • Contact: Jin HU

摘要:

针对雷达辐射源数据清洗机制不完备、特征自适应表征能力不足导致个体识别率不高的问题,提出一种样本对齐和半全局注意力机制的个体识别算法。数据预处理阶段,通过双重条件阈值提取雷达辐射源中频信号的主体部分,实现每个样本的到达时间同步,并进行脉宽对齐与幅度归一化等多维特征对齐操作,使得清洗后样本具备对同频点个体的泛化能力。引入半全局注意力机制,并与双通道卷积神经网络结合,增强卷积神经网络对个体特征表征能力的同时保证训练的高效性。该机制结合了半全局的几何相似度与可学习的相似度,通过并行计算注意力得分并保留长距离特征关联,得到更加准确的特征细节以及更高的计算效率。实验结果表明,对于实测雷达辐射源数据,该方法可以有效解决该信号数据分布不齐以及特征自适应表征能力不足的影响,能够有效提高雷达辐射源个体的识别准确率。

关键词: 辐射源个体识别, 半全局注意力机制, 样本对齐, 几何相似度, 双通道卷积

Abstract:

To address the issues of incomplete data cleaning mechanisms for specific emitter signals and insufficient adaptive feature representation capability leading to low individual recognition rates, this paper proposes an individual identification algorithm incorporating sample alignment and a semi-global attention mechanism. During the data preprocessing stage, the main components of specific emitter intermediate frequency (IF) signals are extracted through multi-condition dual-threshold processing, enabling arrival time synchronization across samples. Subsequently, multi-dimensional feature alignment operations, including pulse width alignment and amplitude normalization, are performed, ensuring the cleaned samples gain generalization capability for individuals with identical frequency points. A semi-global attention mechanism is introduced and integrated with a dual-channel convolutional neural network (CNN), strengthening the network's feature representation capacity while maintaining training efficiency. This mechanism combines semi-global geometric similarity with learnable similarity measures, employing parallel processing of sparse attention scores to preserve long-range feature dependencies, thereby achieving more accurate feature details and higher computational efficiency. Experimental results on measured specific emitter signal datasets indicate that the proposed method effectively address the challenges of uneven data distribution and inadequate adaptive feature representation, significantly improving model recognition accuracy.

Key words: specific emitter identification, semi-global attention mechanism, sample alignment, geometric similarity, dual-channel convolution

中图分类号: