一种采用预分类式神经网络的混叠信号识别方法

An Overlapped Signal Recognition Method Used Pre-classification Neural Network

在非合作通信系统中,对未知信号调制方式的识别是对此信号进行准确处理的前提。随着通信环境的日益复杂,会不可避免地接收到时频混叠信号。针对混叠信号的信号分量{2FSK, BPSK, QPSK, 8PSK, 16QAM, 64QAM}的时频、高阶特征,提出了一种预分类式分类网络,实现了对分量信号两两混合的21种混叠信号调制方式的识别。改进了传统的稠密连接网络(DenseNet),简化其结构,引入自编码(Auto-encoder)和注意力机制(Attention),提出了AttEn-DenseNet。结合双向长短时记忆网络(Bi-LSTM),组成预分类式神经网络。先对混叠信号进行预分类,而后再循环式筛选网络极易混淆的高阶调制信号。在此方法中,预分类结构和注意力/自编码机制,能够帮助网络有效提取特征。通过实验对比,验证了预分类结构对识别信号的促进作用和AttEn-DenseNet对特征的高敏性。

The indispensable prerequisite for accurately signal processing in non-cooperative communication is the accurate modulation recognition of unknown signals.With the increasing complexity of communication environment,it is inevitable to receive time-frequency overlapped signals.The time-frequency and high order characteristic of{2FSK,BPSK,QPSK,8PSK,16QAM,64QAM}signal component of overlapped signals are analyzed and further a new type of pre-classification network to realize modulation recognition of 21 kinds of overlapped signals with component signals are mixed in pairs is proposed.First,Dense Connection Network(DenseNet)is improved and simplified.The attention and auto-encoder mechanism is introduced and AttEn-DenseNet is proposed.Then,the Bidirectional Short and Long Term Memory Network(Bi-LSTM)is combined with AttEn-DenseNet to compose the entire pre-classification neural network.The overlapped signals are first pre-classified,and then the high-order modulated signals that are easily confused by the network are recursively filtered.The pre-classification structure and attention/auto-encoder mechanism can assist the network to effectively extract features.The experimental results illustrate the validity of the pre-classification structure on modulation recognition and the high sensitivity of AttEn-DenseNet to features.