摘要
针对当前真实场景下远距离射频指纹识别难以准确提取特征且实时性较差的问题,提出一种基于奇异谱分析重构信号和改进残差神经网络的射频指纹识别的方法。首先,将采集到的信号进行奇异谱分析,根据贡献率大小对原始信号进行重构,随后通过STFT获得时频谱图作为神经网络的输入:其次,构建轻量级残差神经网络,加快模型收敛速度;然后,在轻量级网络的下采样过程中引入混合维度注意力机制,对网络中间的特征图进行重构,强调重要特征,抑制一般特征;最后,使用激活函数Leaky ReLU替换原有的ReLU,避免在负值区域的梯度永远为0,进而导致模型训练无法反向传播。使用公开数据集POWDER-4BS-Iqsample验证实验后的结果表明,所提方法仅需要训练10个epoch识别精度就能达到87%,在保证识别精度的前提下缩减了时间损耗。与多种经典模型和算法相比,所提方法更加兼具识别精度与实时性。
In view of the difficulty in accurately extracting features and poor real⁃time performance in remote RF fingerprint identification in current real scenarios,a method for RF fingerprint identification based on singular spectrum analysis(SSA)and improved deep residual network(ResNet)is proposed.The collected signal is subjected to SSA,and the original signal is reconstructed based on its contribution rate.Then,the time⁃frequency spectrum is obtained with STFT and is taken as the input of the neural network.A lightweight ResNet is constructed to accelerate the convergence speed of the model.Then,in the downsampling process of lightweight networks,a mixed dimension attention mechanism is introduced to reconstruct the feature map in the middle of the network,which emphasizes important features and suppresses general features.The activation function Leaky ReLU is used to replace the ReLU,so as to avoid the constant zero gradient in the negative value region,which will further result in impossible backpropagation of model training.The results of the validation experiment on the public dataset POWDER⁃4BS⁃Iqsample show that the proposed method only requires training 10 epochs to achieve recognition accuracy of 87%,which reduces time loss while ensuring recognition accuracy.In comparison with various classic models and algorithms,the proposed method has both recognition accuracy and real⁃time performance.
作者
凌浩然
朱丰超
姚敏立
LING Haoran;ZHU Fengchao;YAO Minli(Rocket Force University of Engineering,Xi’an 710025,China)
出处
《现代电子技术》
北大核心
2024年第5期102-107,共6页
Modern Electronics Technique
