机场终端区RF信号深度调制识别方法

Deep Modulation Identification Approach of RF Signal in Airport Terminal Area

论文构建了24种不同信号调制类型的数据集,并提出一款端到端的信号调制识别神经网络。研究了网络卷积层数、卷积核以及训练数据集大小对信号调制识别性能的影响。所提方法避免了基于特征提取的信号调制识别方法中所需的特征选择、信号同步、载波跟踪、信噪比估计等繁杂的处理流程。最后,引入迁移学习技术解决因信道环境变化导致网络识别性能下降的问题,提升了所提网络的环境适应能力。实验结果表明当信道环境发生变化时,通过迁移学习,仅利用不到总样本数的40%作为训练样本即可实现与全部数据集相近的识别性能,并且网络训练时长同比降至总训练样本所需训练时长的1/3。

In the paper,a data set containing 24 kinds of signal modulation waveforms is constructed and an end-to-end signal modulation recognition neural network is developed. The modulation identification accuracy of the proposed neural network versus the convolution layer number,convolution kernel size and training data set size is analyzed,respectively. With the proposed approach,the tedious signal processing procedure including feature selection,signal synchronization,carrier tracking,and SNR estimation which must be done in the baseline approaches are no longer needed. Besides,in order to improve the performance of the neural network when the wireless channel response changed,the transfer learning trick is adopted. Even though the channel response is changed,by virtue of transfer learning in the proposed approach,just using about 40% of data set samples to train the neural network is enough,but the network training time is 1/3 of that when using the whole training samples.