基于半监督矩形网络的通信电台个体识别

通信电台信号的小样本条件造成了电台个体识别准确性欠佳的问题，本文首次提出基于半监督矩形网络进行通信电台个体识别，克服小样本条件对电台个体识别效果的影响.首先提取电台信号的矩形积分双谱特征，人为注入噪声构成污染样本，在半监督矩形网络编码器中有监督训练，其训练结果通过网络径向连接传给解码器，解码器再无监督学习，重构未污染的原始样本，从网络顶层提取电台个体特征，输入softmax分类器实现分类识别.在实际采集的电台数据集上的实验结果说明，本算法在小样本条件下相比现有算法能更准确识别同型号的电台个体.

Communication Radio Individual Recognition Based on Semi-Supervised Rectangular Network

Small sample condition of communication radio signal caused poor individual recognition on radios.To solve this problem,a method about communication radio individual recognition based on semi-supervised rectangular network was proposed innovatively.Firstly,the square integral bispectrum feature was extracted from radio signal and then was corrupted by Gaussian noise.The corrupted sample was passed to the encoder of semi-supervised rectangular network for supervised training.The trained parameterization was then mirrored to decoder through the lateral connection across the model.And the output was forced by decoder through unsupervised learning to be close to the clean input.Then the essential feature extracted was referred as the individual feature of radio signals.Individual recognition was finally accomplished by a softmax classifier.The experiment results on several radio datasets collected in actual environment indicated that the method had superior performance on identifying radio individuals with the same types under small sample condition.