基于生成对抗网络的雷达脉内信号去噪与识别

Generative adversarial model for radar intra-pulse signal denoising and recognition

近年来,深度神经网络在计算机视觉等领域取得了突破性进展,然而在射频信号处理领域,如电子支援侦察系统中的雷达辐射源识别任务,相关技术的发展仍处于起步阶段。在实际军事应用场景中,噪声的存在是影响深度神经网络性能发挥的关键因素。例如,在高信噪比环境下训练至收敛的深度模型分类器在处理低信噪比数据时往往性能下降严重。为了解决上述问题,提出了一种生成对抗式的去噪网络,实现了端到端的雷达信号去噪和脉内调制类型识别。该模型由生成器、鉴别器和分类器三部分组成,其中,生成器为编解码器结构,通过对称的上采样和下采样操作提取输入雷达信号中高阶特征向量,从噪声中恢复出干净信号;鉴别器则用来判断生成器输出去噪结果的真伪;在此基础上,将分类器与上述两者级联,使得去噪结果符合分类所需的语义信息。实验结果表明,所提算法在密集噪声环境下具备高质量的信号去噪效果和较高的分类准确度;与已有算法相比,算法在低信噪比环境数据上的迁移能力具有一定的优越性。

While deep neural networks have achieved an impressive success in computer vision,the related research remains embryonic in radio frequency signal processing,i.e.,a vital task in modern wireless systems,for example,the electronic reconnaissance system.Noise corruption is a harmful but unavoidable factor causing severe performance degradation in the signal processing procedure,and thus has persistently been an intractable problem in the radio frequency domain.For example,a classifier trained on the high signal-to-noise ratio(SNR)data might experience a severe performance degradation when dealing with low SNR data.To address this problem,in this paper we leverage the powerful data representation capacity of deep learning and propose a Generative Adversarial Denoising and classification Network(GADNet)for radar signal restoration and a classification task.The proposed GADNet consists of a generator,a discriminator and a classifier fulfilling an end-to-end workflow.The encoder-decoder structure generator is trained to extract the high-level features and recover signals.Meanwhile,it fools the discriminator’s judges by bewildering the denoising results coming from the clean data.The classification loss from the classifier is adopted jointly to the training procedure.Extensive experiments demonstrate the benefit of the proposed technique in terms of high-quality restoration and accurate classification for radar signals with intense noise.Moreover,it also exhibits superior transferability in low SNR environments compared to the state-of-the-art methods.