基于多尺度时频特征与通道注意力的无人机射频指纹识别方法研究

Research on UAV Radio-Frequency Fingerprint Recognition Method Based on Multi-Scale Time-Frequency Features and Channel Attention

摘要: 针对无人机射频信号识别过程中单尺度时频表示难以兼顾不同时间–频率分辨率，以及噪声扰动下识别性能易下降的问题，本文提出一种基于多尺度时频特征与通道注意力的无人机射频指纹识别方法。首先，对DroneRF公开数据集中的AR drone、Bepop drone和Phantom drone三类射频信号进行格式转换、文件级划分与固定窗口切片，并采用不同窗口长度的短时傅里叶变换(Short-Time Fourier Transform, STFT)构建多尺度时频表示。随后，将窗口长度为128、256和512获得的时频特征统一尺寸并进行通道拼接，以ResNet18为主干网络，引入压缩激励(Squeeze-and-Excitation, SE)通道注意力机制，并结合幅值缩放、时域平移和噪声扰动等数据增强策略，构建Multi-STFT ResNet18-SE-Aug识别模型。实验结果表明，本文模型在文件级验证集上的识别准确率达到0.7946，Macro F1达到0.7833，其中准确率较对应的单尺度模型提高1.65个百分点。进一步的特征域噪声扰动实验表明，多尺度模型在不同扰动强度下均表现出更好的识别稳定性，在10 dB和5 dB条件下准确率分别较单尺度模型提高8.42和9.81个百分点。结果表明，多尺度时频表示能够有效融合不同时间–频率分辨率下的互补特征，并提升无人机射频信号分类及中低信噪比特征扰动条件下的识别性能。

Abstract: To address the limitations of single-scale time-frequency representations in capturing complementary temporal and spectral characteristics of UAV radio-frequency (RF) signals, as well as the degradation of recognition performance under noise perturbations, this paper proposes a UAV RF fingerprint recognition method based on multi-scale time-frequency features and channel attention. First, RF signals from three UAV categories, namely AR drone, Bepop drone, and Phantom drone, are selected from the public DroneRF dataset and processed through format conversion, source-file-level partitioning, and fixed-length segmentation. Short-time Fourier transforms with window lengths of 128, 256, and 512 are then applied to construct multi-scale time-frequency representations. The resulting spectrograms are resized to a uniform resolution and concatenated along the channel dimension. ResNet18 is employed as the backbone network, while squeeze-and-excitation channel attention and data augmentation strategies, including amplitude scaling, temporal shifting, and noise perturbation, are incorporated to construct the Multi-STFT ResNet18-SE-Aug model. Experimental results show that the proposed model achieves an accuracy of 0.7946 and a Macro F1 score of 0.7833 on the file-level validation set, with an accuracy improvement of 1.65 percentage points over the corresponding single-scale model. Further feature-domain noise perturbation experiments demonstrate that the multi-scale model maintains better recognition stability under different perturbation levels, achieving accuracy improvements of 8.42 and 9.81 percentage points at 10 dB and 5 dB, respectively. These results indicate that multi-scale time-frequency representations can effectively integrate complementary RF characteristics at different time-frequency resolutions and improve UAV RF signal classification performance, particularly under medium- and low-SNR feature perturbations.