| 摘要: |
| 针对辐射源个体识别（Specific Emitter Identification，SEI）在信道干扰下识别正确率显著下降的问题，利用子域适应能够对齐不同子领域的特征分布，提出了一种有监督子域适应的SEI方法。首先采用傅里叶分析网络（Fourier Analysis Networks，FAN）替代传统卷积神经网络（Convolutional Neural Network，CNN）中的多层感知机，设计了CNN-FAN模型，能够直接从RAW I/Q信号中提取信号特征，然后将提取到的特征进行分类并计算局部最大均值差异（Local Maximum Mean Discrepancy，LMMD），通过不断优化网络参数降低分类损失以及最小化LMMD，最终模型能够对齐不同信道下同一个体的特征分布，提高在信道干扰下的SEI性能。实验结果表明，所设计的CNN-FAN在信噪比为10 dB 的加性高斯白噪声信道干扰下识别准确率为97.08%；所提出的基于有监督子域适应的SEI方法在3种实际信道干扰下识别准确率分别达到99.17%、96.83%和93.83%。 |
| 关键词: 辐射源个体识别(SEI) 深度学习 傅里叶分析网络(FAN) 子域适应 |
| DOI：10.20079/j.issn.1001-893x.250708005 |
|
| 基金项目:国家自然科学基金资助项目(62461015)；广西自然科学基金项目(2023GXNSFAA026060) |
|
| A Supervised Subdomain Adaptation Method for Specific Emitter Identification |
| HE Jialong,LIU Xiangguo,XIE Yuelei |
| (School of Information and Communication，Guilin University of Electronic Technology，Guilin 541004，China) |
| Abstract: |
| For the significant decline in identification accuracy of specific emitter identification (SEI) under channel interference，a supervised subdomain adaptation method is proposed.This method leverages subdomain adaptation to align feature distributions across different subdomains.Initially，a Fourier analysis network (FAN) is employed to replace the multilayer perceptron in a traditional convolutional neural network (CNN)，resulting in the CNN-FAN model.This model directly extracts signal features from RAW I/Q signals.Subsequently，the extracted features are classified，and the local maximum mean discrepancy (LMMD) is computed.By iteratively optimizing network parameters to minimize classification loss and the LMMD，the model aligns the feature distributions of the same emitter under different channel conditions，thus enhancing SEI performance under channel interference.Experimental results demonstrate that the CNN-FAN achieves an identification accuracy of 97.08% under additive white Gaussian noise channel interference with a signal-to-noise ratio of 10 dB.Furthermore，the proposed supervised subdomain adaptation-based SEI method achieves identification accuracies of 99.17%，96.83%，and 93.83%，respectively，under three types of actual channel interference. |
| Key words: specific emitter identification(SEI) deep learning Fourier analysis networks(FAN) subdomain adaptation |
《电讯技术》编辑部 