信号增强网络驱动的调制识别

Modulation recognition driven by signal enhancement

现有基于深度学习的调制识别在训练阶段需要大量IQ信号样本,而复杂电磁环境中很难获取大量样本,导致基于深度学习的调制识别算法泛化性能下降。针对网络泛化能力差的问题,提出了一种基于信号增强的调制识别(signal enhancement based modulation recognition,SEBMR)算法。首先,设计了捕获IQ信号全局特征的特征提取及重构模块;其次,提出了基于辅助分类生成对抗网络(auxiliary classifier generative adversarial network,ACGAN)的IQ信号增强网络,实现了样本数量及质量的双重增强;最后,利用支持向量机算法实现了调制方式识别分类。为了实现对复杂信道下调试信号的识别,训练时采用表征全局特征的重构信号,测试时采用经历无线衰落的IQ基带信号。实验结果表明,提出的方法相比现有的基于长短期记忆(LSTM)网络、卷积神经网络(CNN)、注意力机制等识别方法,在小样本训练集、衰落信道环境下可获得更优的识别准确度。

The existing modulation recognition algorithms based on deep learning theory require a large number of IQ signal samples during the training phase.It is difficult to obtain a large number of samples in complex electromagnetic environments,resulting in a decrease in the generalization performance of modulation recognition algorithms based on deep learning.A signal enhancement based modulation recognition(SEBMR)algorithm was proposed to address the issue of poor network generalization ability.Firstly,a feature extraction and reconstruction module was designed to capture the global features of IQ signals.Secondly,an IQ signal enhancement network based on auxiliary classifier generative adversarial network(ACGAN)was proposed to achieve dual enhancement of sample quantity and quality.Finally,the support vector machine algorithm was employed to achieve modulation recognition and classification.To achieve recognition of debugging signals in complex channels,reconstructed signals representing global features were for training,and IQ baseband signals which experienced wireless fading were used for testing.The experimental results show that the proposed method can achieve better recognition accuracy performance in small sample training sets and fading channel environments compared to existing recognition methods based on long shortterm memory(LSTM),convolutional neural network(CNN),attention mechanism,etc.