基于迁移学习的水声通信信号调制识别方法

Modulation Recognition Method of Underwater Acoustic Communication Signal Based on Transfer Learning

支持多种载波调制方式的多制式自适应水声通信,可高效利用时变、频变、空变水声信道的通信带宽,是水声通信技术的发展趋势之一。针对多制式自适应水声通信的信号载波调制方式自动识别需求,本文提出了一种基于迁移学习的水声通信信号载波调制方式识别方法(UACCMR/TL)。通过多制式水声通信系统建模、水声信道建模,生成了单载波时间域扩频(SCTDSS)、单载波频域均衡(SC-FDE)、正交频分复用(OFDM)和多载波频域扩频(MC-FDSS)4种典型水声通信载波调制信号的时间序列数据集,采用时频分析方法将时间序列信号数据集转换成时频图像数据集,利用迁移学习方法将图像识别领域的深度学习模型VGG16应用于水声通信信号的载波调制方式识别。仿真结果表明:所提出的UACCMR/TL方法对SC-TDSS、SC-FDE、OFDM和MC-FDSS信号具有良好的识别性能,信噪比为5 dB以上时,分类准确率均达到90%以上,信噪比为15 dB时,分类准确率均接近于100%;同时,UACCMR/TL方法可以在不同信道较少数据样本模型微调下达到良好的识别性能,具有良好的信道泛化性能。

Multi-system adaptive underwater acoustic communication supporting multiple carrier modulation modes can make efficient use of the communication bandwidth of time-varying,frequency varying and space-varying underwater acoustic channels.It is one of the development trends of underwater acoustic communication technology.Aiming at the requirement of automatic recognition of signal carrier modulation mode in multi-system adaptive underwater acoustic communication,an underwater acoustic communication signal carrier modulation recognition based on transfer learning(UACCMR/TL)is proposed.Through multi-system underwater acoustic communication system modeling and underwater acoustic channel modeling,the time series data sets of four typical underwater acoustic communication carrier modulation signals are generated,including single carrier time domain spread spectrum(SC-TDSS),single carrier frequency domain equalization(SC-FDE),orthogonal frequency division multiplexing(OFDM)and multicarrier frequency domain spread spectrum(MC-FDSS).The time series signal data sets are converted into time-frequency image data sets by time-frequency analysis method,and the deep learning model VGG16 in the field of image recognition is applied to the recognition of carrier modulation mode of underwater acoustic communication signal by transfer learning.Simulation results show that the proposed UACCMR/TL method has good classification effect on SC-TDSS,SC-FDE,OFDM and MC-FDSS signals.When the SNR is more than 5dB,the classification accuracy of the four signals remains above 90%.When the SNR is 15dB,the classification accuracy of the four signals is close to 100%.At the same time,the proposed UACCMR/TL method can achieve good recognition performance and good channel generalization performance in small data sample model fine tuning of different channels.