| 摘要: |
| 针对通信信号辐射源个体样本稀缺导致的个体识别性能不足的问题，提出了一种基于Noise-WGAN-GP框架的信号辐射源个体样本增强与识别方法。首先对通信信号数据随机加噪，生成高斯分布的加噪数据，将加噪数据作为输入WGAN-GP（Wasserstein Generative Adversarial Network with Gradient Penalty）的生成器生成样本，判别器引入真实数据和生成样本分布的Wasserstein距离作为损失项，添加梯度惩罚。在ACARS（Aircraft Communications Addressing and Reporting System）与BPSK（Binary Phase Shift Keying）辐射源个体数据集进行仿真验证，以余弦相似度和误差向量幅度作为相似度度量，生成样本与真实样本具有较高相似度。基于生成后的ACARS和BPSK数据进行辐射源个体识别，准确率分别达86.38%和86.54%，相比原小样本数据识别准确率分别提升40.38%与36.54%，且优于频率偏移、相位偏移、纯噪声过采样增强等增强方式。 |
| 关键词: 辐射源个体识别 小样本学习 Noise-WGAN-GP Wasserstein距离 |
| DOI：10.20079/j.issn.1001-893x.251012005 |
|
| 基金项目: |
|
| Signal Individual Sample Augmentation and Specific Emitter Identification Based on Noise-WGAN-GP Framework |
| LU Xuemin,ZHOU Xi |
| (1.Southwest China Institute of Electronic Technology，Chengdu 610036，China;2.Unit 92728 of PLA, Shanghai 200436，China) |
| Abstract: |
| To address the insufficient individual recognition performance caused by the scarcity of individual samples from communication signal radiation sources，a Noise-WGAN-GP framework-based method is proposed for individual sample augmentation and recognition of signal radiation sources.First，communication signal data is randomly added with noise to generate Gaussian-distributed noisy data，which is then input into the generator of Wasserstein generative adversarial network with gradient penalty(WGAN-GP) to produce samples.The discriminator incorporates the Wasserstein distance between real data and generated sample distributions as a loss term，with gradient penalty added.Experimental validation on aircraft communications addressing and reporting system(ACARS) and binary phase shift keying(BPSK) radiation source individual datasets shows that when using cosine similarity and error vector magnitude (EVM) as similarity metrics，the generated samples exhibit high similarity to real samples.Individual radiation source recognition based on the generated ACARS and BPSK data achieves accuracies of 86.38% and 86.54% respectively，representing improvements of 40.38% and 36.54% compared with recognition using the original small-sample data.This method also outperforms other augmentation approaches such as frequency offset，phase offset，and pure noise oversampling augmentation. |
| Key words: specific emitter identification few-shot learning Noise-WGAN-GP Wasserstein distance |
《电讯技术》编辑部 