Scholarlink

<p>Aiming to address the limitations of traditional unmanned aerial vehicle (UAV) detection methods that predominantly rely on radar echo signals, acoustic signals, and optoelectronic signals, this study proposes a UAV recognition method based on multi-dimensional signal characteristics. These conventional technologies are prone to environmental interference, exhibiting notable limitations such as low signal-to-noise ratios for radar echoes in cluttered environments and the ineffectiveness of optical signals under adverse visibility conditions (e.g., at night or during bad weather), which impede precise drone positioning and identification. The proposed method employs Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and wavelet packet threshold transform to denoise UAV signals and separate them from the received radio frequency (RF) signals. Wavelet Transform (WT) and Short-Time Fourier Transform (STFT) are utilized to process the signal after noise reduction, enabling effective feature extraction. Specifically, the wavelet transforms result scattergrams, amplitude spectrum, and phase spectrum obtained after STFT and WT are combined to form a three-channel input network, facilitating multi-modal analysis. Research findings indicate that when UAV signals processed by CEEMDAN combined with wavelet packet thresholding are classified using an enhanced EfficientNet-UAV model with improved three-channel features extracted after WT and STFT processing, the accuracy rate for detecting UAV presence reaches 100%, while the accuracy rate for 5 modes is 99.4%. The accuracy for 8 modes and 12 modes are 95.9% and 95.2%. Compared with previous methods, this approach effectively mitigates UAV signal noise, and the unique selection of three features forming a three-channel input significantly enhances recognition accuracy. Additionally, the selected network model exhibits advantages such as fewer parameters, high precision, and faster training speed compared to commonly used network models.</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

Detection and recognition of UAV radio frequency signals based on time–frequency processing and transfer learning with multi-channel input

  • 期刊论文

  • 发表日期：

    2025年7月11日

查看全文
  • Shuoyang Wang,
  • Yongqiang Luo,
  • Yifan Zheng,
  • Zhongsen Sun,
  • Yu Zheng,
  • Tian Liu,
  • Zhejun Jin

摘要

Aiming to address the limitations of traditional unmanned aerial vehicle (UAV) detection methods that predominantly rely on radar echo signals, acoustic signals, and optoelectronic signals, this study proposes a UAV recognition method based on multi-dimensional signal characteristics. These conventional technologies are prone to environmental interference, exhibiting notable limitations such as low signal-to-noise ratios for radar echoes in cluttered environments and the ineffectiveness of optical signals under adverse visibility conditions (e.g., at night or during bad weather), which impede precise drone positioning and identification. The proposed method employs Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and wavelet packet threshold transform to denoise UAV signals and separate them from the received radio frequency (RF) signals. Wavelet Transform (WT) and Short-Time Fourier Transform (STFT) are utilized to process the signal after noise reduction, enabling effective feature extraction. Specifically, the wavelet transforms result scattergrams, amplitude spectrum, and phase spectrum obtained after STFT and WT are combined to form a three-channel input network, facilitating multi-modal analysis. Research findings indicate that when UAV signals processed by CEEMDAN combined with wavelet packet thresholding are classified using an enhanced EfficientNet-UAV model with improved three-channel features extracted after WT and STFT processing, the accuracy rate for detecting UAV presence reaches 100%, while the accuracy rate for 5 modes is 99.4%. The accuracy for 8 modes and 12 modes are 95.9% and 95.2%. Compared with previous methods, this approach effectively mitigates UAV signal noise, and the unique selection of three features forming a three-channel input significantly enhances recognition accuracy. Additionally, the selected network model exhibits advantages such as fewer parameters, high precision, and faster training speed compared to commonly used network models.