基于自注意力多尺度空洞卷积网络的雷达工作模式识别

Radar working mode recognition based on self-attention multi-kernel dilated convolution network

关键词:  雷达工作模式识别  深度学习  自注意力机制  多尺度卷积  空洞卷积

Key words:  radar working mode recognition  deep learning  self-attention mechanism  multi-scale convolution  dilated convolution

摘要:

在复杂电磁环境下,雷达对抗侦察信号常因大量脉冲丢失与虚假脉冲干扰,造成训练数据与实战场景间存在显著分布差异,严重影响有源相控阵雷达空空工作模式的识别准确率。针对这一问题,本文提出一种基于自注意力多尺度空洞卷积网络（self-attention multi-kernel dilated convolution network,SAMKDCN）的识别模型。该模型以空洞卷积、多卷积核选择与残差结构为核心,构建特征图提取模块,实现在时间维度上对多尺度特征图的提取。进一步通过自注意力机制,实现对特征图权重的自适应调整,以突出关键特征,增强特征表示能力,从而提升AESA雷达工作模式识别的准确率。仿真实验表明:SAMKDCN能够有效学习AESA雷达空空工作模式的核心特征；在理想环境下,其准确率最高可达99.14%；在0%~50%的脉冲丢失与虚假脉冲条件下,平均识别率达到95.11%；在50%丢失率与50%虚假率的极端条件下,仍保持88.23%的识别准确率,显示出良好的泛化能力与鲁棒性。

Abstract:

In complex electromagnetic environments, radar countermeasure reconnaissance signals often suffer from significant distribution differences between training data and actual combat scenarios due to substantial pulse loss and false pulse interference, which seriously degrades the recognition accuracy of the air-to-air working mode of active phased array radar. To address this issue, this paper proposes a recognition model based on self-attention multi-kernel dilated convolution network (SAMKDCN). Centered on dilated convolution, multi-kernel selection, and residual structures, this model constructs a feature-map extraction module for multi-scale feature learning across the temporal dimension. Moreover, a self-attention mechanism is incorporated to adaptively adjust feature-map weights, thereby highlighting critical features and strengthening feature representation, which ultimately enhances the accuracy of AESA radar working-mode identification. Simulation experiments show that SAMKDCN can effectively learn the core features of the air-to-air working mode of AESA radar. Under ideal conditions, it achieves a peak accuracy of 99.14%. With pulse-loss and false-pulse ratios ranging from 0% to 50%, the average recognition rate attains 95.11%; Even under the extreme scenario of 50% loss rate and 50% false-pulse rate, this model retains a recognition accuracy of 88.23%, demonstrating favorable generalization ability and robustness.