基于空间变换网络的盲信号调制识别方法

Blind Signal Modulation Recognition Method Based on Spatial Transformer Network

通信信号的自动调制识别(Automatic Modulation Recognition,AMR)在各种领域都扮演着较为重要的角色。基于神经网络的调制识别方法,能够根据提取出已调信号中可以用于分类的高层次抽象特征,相比于传统的基于人为定义特征的识别方法,具有识别率更高的优势。在实际的调制识别应用场景中,由于盲接收参数估计不准确等原因,调制识别的输入信号样本存在较大的相位、频率偏移,或时间尺度改变。已有的深度学习调制识别方法未考虑上述原因带来的影响,导致调制识别率的降低。提出了一种基于空间变换网络(Spatial Transformer Network,STN)的调制识别方法,在网络模型中引入了通信中同步的一些先验模型,通过空间变换子网络实现先验模型,在一定程度上减轻相位、频率偏移以及不同时间尺度对调制识别的影响。通过仿真数据集的实验证明,提出的方法相比于传统的CNN网络、ResNet网络和CLDNN网络的识别率在SNR>0的条件下分别提高了8.3%、4.9%和5.2%,且更加容易训练,其收敛时所需的训练时间相比于CNN网络、ResNet网络和CLDNN网络分别下降了3.5%、27%和85%。

features.However,in real ARM application scenarios,the input for AMR has significant frequency and phase bias or time scale changes due to inaccurate blind receiving parameter estimation.Existing AMR method based on neural network can thus encounter a decrease in recognition rate due to the above mentioned reasons.An AMR method based on Spatial Transformer Network(STN)is proposed,in which some a priori modulation models are introduced in modulations to alleviate the effects of frequency bias,phase bias and different time scales.The proposed method is validated through a simulated dataset,which proves that the proposed method achieves an increase of 8.3%,4.9%and 5.2%in recognition rate over traditional CNN network,ResNet network and CLDNN network respectively when SNR is greater than 0.The proposed method can be trained more easily with a decrease of 3.5%,27%and 85%in training time as compared with traditional CNN network,ResNet network and CLDNN network respectively.