Scholarlink

<p>To address the critical limitations of conventional complex-valued residual networks in radio frequency (RF) signal analysis, including inadequate recognition accuracy under low signal-to-noise ratio (SNR) conditions, and computational inefficiency in processing I/Q modulated RF signals, we propose a complex adaptive denoising residual network based on feature aggregation network (CADRN_FAN) for RF fingerprinting identification. Our main contributions are as follows: First, we design a complex residual neural network with an adaptive denoising module. We achieve effective elimination of noise and redundant data by introducing the proposed complex channel attention mechanism and the improved complex feature aggregation network structure. Second, a customized complex depthwise separable convolution (CDSC) lightweight technique for RF signals is proposed, which simplifies the network structure while ensuring model performance. Third, an end-to-end framework of “fast Fourier transform (FFT) feature extraction—adaptive denoising—complex lightweighting” is constructed to achieve the synergistic optimization of recognition performance and computational efficiency. Experimental results show that in the SNR range of 13–18&#xa0;dB, the proposed method achieves an average recognition accuracy of 86.9% on 100 classes of ADS-B data, which is 2.1 percentage points higher than that of the traditional complex residual network. Compared with the traditional complex residual network, the lightweight network achieves a 30.5% reduction in parameters with an accuracy loss of no more than 1%. Compared with other methods, the proposed method has significant advantages: Compared with real-valued networks that discard phase information, this method completely retains the amplitude and phase features of I/Q signals, resulting in higher recognition accuracy; compared with traditional complex residual networks, this method exhibits stronger adaptive denoising capability through the complex channel attention mechanism specifically designed in this study and the improved complex feature aggregation network structure, thus achieving superior generalization performance under low SNR conditions. Furthermore, relying on the CDSC adapted to RF signals, the proposed method effectively solves the accuracy loss problem caused by lightweight design, ultimately achieving a balance between efficient RF signal recognition and model compression.</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

CADRN_FAN: A Robust Radio Frequency Fingerprinting Identification Method for Low Signal-to-Noise Ratio Environments

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Qi Cheng,
  • Huojun Zhao,
  • Xiaoli Gao,
  • Jie Li,
  • Jicong Zhou

摘要

To address the critical limitations of conventional complex-valued residual networks in radio frequency (RF) signal analysis, including inadequate recognition accuracy under low signal-to-noise ratio (SNR) conditions, and computational inefficiency in processing I/Q modulated RF signals, we propose a complex adaptive denoising residual network based on feature aggregation network (CADRN_FAN) for RF fingerprinting identification. Our main contributions are as follows: First, we design a complex residual neural network with an adaptive denoising module. We achieve effective elimination of noise and redundant data by introducing the proposed complex channel attention mechanism and the improved complex feature aggregation network structure. Second, a customized complex depthwise separable convolution (CDSC) lightweight technique for RF signals is proposed, which simplifies the network structure while ensuring model performance. Third, an end-to-end framework of “fast Fourier transform (FFT) feature extraction—adaptive denoising—complex lightweighting” is constructed to achieve the synergistic optimization of recognition performance and computational efficiency. Experimental results show that in the SNR range of 13–18 dB, the proposed method achieves an average recognition accuracy of 86.9% on 100 classes of ADS-B data, which is 2.1 percentage points higher than that of the traditional complex residual network. Compared with the traditional complex residual network, the lightweight network achieves a 30.5% reduction in parameters with an accuracy loss of no more than 1%. Compared with other methods, the proposed method has significant advantages: Compared with real-valued networks that discard phase information, this method completely retains the amplitude and phase features of I/Q signals, resulting in higher recognition accuracy; compared with traditional complex residual networks, this method exhibits stronger adaptive denoising capability through the complex channel attention mechanism specifically designed in this study and the improved complex feature aggregation network structure, thus achieving superior generalization performance under low SNR conditions. Furthermore, relying on the CDSC adapted to RF signals, the proposed method effectively solves the accuracy loss problem caused by lightweight design, ultimately achieving a balance between efficient RF signal recognition and model compression.