基于网络度量的三分支孪生网络调制识别算法

Triplet Siamese Network Modulation Recognition Algorithm Based on Network Measurement

针对基于孪生网络的小样本调制识别算法存在相似识别类别混淆的问题,提出一种基于网络度量的三分支孪生网络调制识别算法。通过三分支孪生神经网络(Triplet Siamese Neural Network,TSN)将输入三元样本组的原始特征映射至特征子空间中。将正负样本与参考样本特征并联输入至两个参数共享的关系网络(Relation Network,RN),学习一个非线性的度量函数。通过各个类别的特征向量生成各类别的类原型,作为测试过程中的类特征输入。为降低信道噪声和信号接收误差对均值类原型表达的影响,采用局部异常因子算法(Local Outlier Factor,LOF)剔除类别中偏差数据。在公开的调制数据集Deep Sig中进行验证,仿真结果表明,TSN-RN-LOF算法模型可以充分利用相似类别之间的差异信息,提取更具辨识度的特征,取得更优的识别性能。

Aiming at the problem of similar recognition category confusion in small sample modulation recognition algorithms based on Siamese Neural Network,and the Triplet Siamese Network modulation recognition algorithm based on network measurement is proposed.Firstly,the original features of the input triplet set are mapped into the feature subspace through the Triplet Siamese Neural Network(TSN).Then,the features of the positive and negative samples and the reference samples are input in parallel to the Relation Network(RN)shared by two parameters,and a nonlinear measure function is learned.Finally,the prototype of each category is generated through the feature vector of each category,which is used as class feature input in the testing process.At the time,in order to reduce the influence of channel noise and signal reception error on the expression of mean-type prototypes,the Local Outlier Factor(LOF)algorithm is used to eliminate the deviation data in the category.The results conducted in the public modulation data set DeepSig indicate that the algo-rithm model of TSN-RN-LOF makes full use of the difference information between similar categories,extracts more recognizable features and achieves better recognition performance.