计算机科学
干扰(通信)
校准
鉴定(生物学)
算法
人工智能
计算机视觉
感知
算法设计
门控
时频分析
噪音(视频)
钥匙(锁)
雷达跟踪器
频域
频率响应
信噪比(成像)
信号处理
噪声测量
作者
Haoran Shi,Dan Song,Bowen Ding,Bin Rao,Wei Wang
标识
DOI:10.1109/jiot.2026.3718421
摘要
In complex urban electromagnetic environments, uncrewed aerial vehicle (UAV) signal features are easily overwhelmed by co-channel interference and background noise, severely limiting identification system performance. Traditional explicit feature analysis methods struggle to extract robust features under low signal-to-interference-plus-noise ratio (SINR) conditions. Although existing deep learning methods are capable of automatic feature extraction, signal features are severely obscured by interference and noise under low SINR conditions, making it difficult for networks to effectively capture weak signal features and leading to the loss of critical discriminative information. To address the aforementioned challenges, this article proposes a UAV individual identification framework tailored for low SINR environments and designs an interference suppression network based on dynamic frequency perception and global gating calibration (DFP-GGCNet). DFP-GGCNet comprises two core modules: the dynamic frequency perception module (DFPM) and the global gating calibration module (GGCM). Specifically, DFPM leverages frequency dynamic convolution (FDConv) to generate overcomplete features, which are subsequently calibrated by GGCM using the global time–frequency structure of the signal. Furthermore, a hybrid loss function incorporating pixel-level reconstruction loss and structural similarity (SSIM) loss is constructed to guide the network to focus on the recovery of signal time–frequency features from the perspectives of numerical accuracy and the global time–frequency structure. Experiments on the DroneRFb-DIR dataset demonstrate that, when the SINR ranges from −20 to −5 dB, the proposed framework improves the identification accuracy by an average of 13.83% compared with existing mainstream methods and achieves an average identification accuracy of 99.16% in the non-line-of-sight (NLOS) scenario.
科研通智能强力驱动
Strongly Powered by AbleSci AI