Scholarlink

Abstract <p>Automatic modulation classification (AMC) plays a critical role in identifying radar and communication waveforms under diverse and noisy signal conditions. In congested environments, reliable classification helps reduce interference and optimize communication resource allocation. This study proposes two deep learning architectures designed for radar modulation classification under varying Signal-to-Noise Ratios (SNRs), a baseline ConvGRU model and an enhanced Fused ConvGRU model. The models utilize Convolutional Neural Networks (CNNs) to extract spatial features, Gated Recurrent Units (GRUs) to capture temporal dependencies, and Multi-Head Attention to improve feature interaction. The Fused ConvGRU model introduces a trainable fusion gate to integrate spatial and temporal information more effectively. The proposed Fused ConvGRU model achieves a maximum accuracy of 99.72% at 10 dB SNR and 95.59% at 0 dB SNR on the DeepRadar dataset, utilizing only 162,391 trainable parameters. In comparison, the baseline ConvGRU model attains 99.24% accuracy at 10 dB and 93.23% at 0 dB SNR, with 154,135 trainable parameters. The exceptional performance of the proposed models highlights their utility for blind classification and adaptation to dynamic EW scenarios, making them a versatile tool for future EW applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A CNN and GRU based deep learning architecture for radar signal classification

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Asim Saleem,
  • Guoyun Lv,
  • Fan YangYu,
  • Muhammad Saad Ayub

摘要

Abstract

Automatic modulation classification (AMC) plays a critical role in identifying radar and communication waveforms under diverse and noisy signal conditions. In congested environments, reliable classification helps reduce interference and optimize communication resource allocation. This study proposes two deep learning architectures designed for radar modulation classification under varying Signal-to-Noise Ratios (SNRs), a baseline ConvGRU model and an enhanced Fused ConvGRU model. The models utilize Convolutional Neural Networks (CNNs) to extract spatial features, Gated Recurrent Units (GRUs) to capture temporal dependencies, and Multi-Head Attention to improve feature interaction. The Fused ConvGRU model introduces a trainable fusion gate to integrate spatial and temporal information more effectively. The proposed Fused ConvGRU model achieves a maximum accuracy of 99.72% at 10 dB SNR and 95.59% at 0 dB SNR on the DeepRadar dataset, utilizing only 162,391 trainable parameters. In comparison, the baseline ConvGRU model attains 99.24% accuracy at 10 dB and 93.23% at 0 dB SNR, with 154,135 trainable parameters. The exceptional performance of the proposed models highlights their utility for blind classification and adaptation to dynamic EW scenarios, making them a versatile tool for future EW applications.