Scholarlink

<p>Automatic Modulation Recognition (AMR) constitutes a critical enabling technology for satellite cognitive radio systems, enabling receivers to identify modulation schemes without prior channel state information (CSI). However, satellite-to-ground communication links present unique challenges, including dynamic Doppler frequency shifts arising from Low Earth Orbit (LEO) satellite motion and nonlinear amplitude-phase distortions from Traveling Wave Tube Amplifiers (TWTA) operating near saturation regions. Existing deep learning approaches predominantly focus on time-domain signal analysis, thereby underutilizing the interpretable spectral fingerprints that these channel impairments leave in the frequency domain. To address these limitations, this paper proposes a Frequency-aware Cross-scale Attention Network (FCAN) that employs a dual-branch parallel architecture for joint time–frequency feature extraction. The time-domain branch utilizes Cross-Scale Dilated Convolution (CSDC) modules with adaptive Multi-scale Feature Selection (MFS) to capture temporal patterns across varying symbol rates and modulation orders. The frequency-domain branch integrates learned deep spectral representations with four physics-driven statistical descriptors—spectral centroid, spectral flatness, spectral bandwidth, and low-frequency energy ratio—that encode interpretable domain knowledge about satellite channel characteristics. A bounded gating fusion mechanism with constrained modulation coefficients dynamically recalibrates features from both domains while preventing feature extinction under adverse low signal-to-noise ratio (SNR) conditions. Comprehensive experiments on the RML24 satellite benchmark dataset demonstrate that FCAN achieves state-of-the-art performance with 72.78% overall accuracy across 22 modulation classes using only 422&#xa0;K parameters, outperforming the best baseline method by 1.64% while maintaining computational efficiency suitable for resource-constrained satellite terminals. Systematic ablation studies validate the synergistic contributions of each architectural component, confirming the effectiveness of integrating physical priors with data-driven representations for robust satellite AMR.</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

FCAN: A frequency-aware cross-scale attention network for automatic modulation recognition in satellite communications

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月23日

查看全文
  • Yong Zhang,
  • Yong Wang,
  • Qingsong Zhao,
  • Shihua Pan,
  • Fei Xie,
  • Sheng Luo,
  • Yadi Zhai,
  • Jiaxiong Yang

摘要

Automatic Modulation Recognition (AMR) constitutes a critical enabling technology for satellite cognitive radio systems, enabling receivers to identify modulation schemes without prior channel state information (CSI). However, satellite-to-ground communication links present unique challenges, including dynamic Doppler frequency shifts arising from Low Earth Orbit (LEO) satellite motion and nonlinear amplitude-phase distortions from Traveling Wave Tube Amplifiers (TWTA) operating near saturation regions. Existing deep learning approaches predominantly focus on time-domain signal analysis, thereby underutilizing the interpretable spectral fingerprints that these channel impairments leave in the frequency domain. To address these limitations, this paper proposes a Frequency-aware Cross-scale Attention Network (FCAN) that employs a dual-branch parallel architecture for joint time–frequency feature extraction. The time-domain branch utilizes Cross-Scale Dilated Convolution (CSDC) modules with adaptive Multi-scale Feature Selection (MFS) to capture temporal patterns across varying symbol rates and modulation orders. The frequency-domain branch integrates learned deep spectral representations with four physics-driven statistical descriptors—spectral centroid, spectral flatness, spectral bandwidth, and low-frequency energy ratio—that encode interpretable domain knowledge about satellite channel characteristics. A bounded gating fusion mechanism with constrained modulation coefficients dynamically recalibrates features from both domains while preventing feature extinction under adverse low signal-to-noise ratio (SNR) conditions. Comprehensive experiments on the RML24 satellite benchmark dataset demonstrate that FCAN achieves state-of-the-art performance with 72.78% overall accuracy across 22 modulation classes using only 422 K parameters, outperforming the best baseline method by 1.64% while maintaining computational efficiency suitable for resource-constrained satellite terminals. Systematic ablation studies validate the synergistic contributions of each architectural component, confirming the effectiveness of integrating physical priors with data-driven representations for robust satellite AMR.