基于残差原型网络的辐射源个体识别

Specific emitter identification based on residual prototype network

利用深度学习进行通信辐射源识别分类时,现有算法在较低信噪比下的识别能力还不足,且均着重关注各类辐射源个体的类间距离,忽视了类内紧密性。针对此问题,结合残差网络和原型学习基本思想,提出残差原型网络,对输入信号的差分星座轨迹图进行识别。此外,在基于距离的交叉熵损失函数基础上加入原型损失,通过提高类内紧密度的方式进一步扩增了类间距离。通过对5种ZigBee设备的实验,结果表明所提算法在相同信噪比条件下相较于其他算法具有更好的识别性能,在信噪比高于8 dB时,可达到99%以上的准确率。

When deep learning methods are used for specific emitter identification,the existing algorithms are insufficient under the low signal to noise ratio.Meanwhile,they all focus on the inter-class distance but ignore the intra-class compactness.To solve this problem,a residual prototype network is proposed to recognize the differential constellation trace figure of input signals by combining the residual network and prototype learning.In addition,prototype loss is combinedwith the distance-based cross entropy loss to further amplify the inter-class distanceby improving the intra-class compactness.The results show that the proposed algorithm has better recognition performance under the same signal to noise ratio condition through experiments on five ZigBee devices.And the accuracy can reach more than 99%when the signal to noise ratio is higher than 8 dB.