基于GRU和ResNet的短时水声通信信号调制识别  

Automatic Modulation Recognition of Short-time Underwater Acoustic Communication Signal Based on GRU and ResNet

作　　者：瞿逢重[1] 朱江[1] 涂星滨 杨劭坚 魏艳[1] 方浩[1] QU Fengzhong;ZHU Jiang;TU Xingbin;YANG Shaojian;WEI Yan;FANG Hao(Ocean College of Zhejiang University,Zhoushan,Zhejiang 316021,China)

机构地区：[1]浙江大学海洋学院,浙江舟山316021

出　　处：《信号处理》2023年第10期1793-1804,共12页Journal of Signal Processing

基　　金：国家杰出青年科学基金(62225114);国家自然科学基金(62101489,62171405);中国科学院战略性先导科技专项(NO.XDA22040402)。

摘　　要：针对水声通信信号识别中的准确率低和要求输入信号时间长的问题,本文提出一种基于门控循环单元(gate recurrent unit,GRU)和残差网络(residual network,ResNet)的短时水声通信信号识别方法,设计并实现该识别网络。首先,设计GRU层的输出结构,从短时通信数据中提取出特征图像,将其用于后续的网络进行进一步的特征提取。其次,对原始的神经残差单元进行改进,解除单元中卷积核通道数必须与输入信号通道数相同的限制,保留残差单元能避免梯度消失的功能。随后,使用改进的残差单元组建ResNet网络层,该网络层使用GRU层的特征图像作为输入,并进行高维度和深层次的挖掘。之后,使用重新设计的GRU与ResNet组成融合神经网络。最后,在仿真中生成单径信道、深海信道和浅海信道下的数据作为数据集来验证该网络的调制识别性能,利用实际海试中采集真实的水声通信数据来验证该网络的性能,并将几何视觉组(visual geometry group network,VGG)、长短期记忆网络(long short-term memory,LSTM)、门控神经单元与卷积长短期记忆网络(CNN-LSTM)等传统神经网络与本文提出的网络进行比较。仿真结果表明,该网络结构在时不变的水声信道下信噪比为8 dB时识别率能达到100%,在时变的水声信道下调制识别性能优于上述传统网络;海试结果表明,该网络在真实的水声信道下识别准确率对比上述传统神经网络有较大优势,正确率达到65.17%,比准确率第二的GRU高11.06%,比LSTM高17.04%,比CNNLSTM高21.52%,比VGG高41.7%。Underwater acoustic communication signal modulation recognition has long been notorious for its low accuracy and long input signal time.To tackle this challenge,a short-term underwater acoustic communication signal modulation recognition method based on gate recurrent unit(GRU)and residual network(ResNet)was proposed of in.First,we de-signed a new output structure of the GRU layer,extracted the feature image from the short-term communication data,and used it in the subsequent network for further feature extraction.Secondly,the original neural residual unit was improved.This step not only removed the restriction that the number of convolution kernel channels in the unit must be the same as the number of input channels but also retained the function of the residual unit to avoid gradient disappearance.Subse-quently,the improved residual unit was used to form the ResNet network layer.This layer used the feature image of the GRU layer as input for high-dimensional and deep-level mining.After that,we used the redesigned GRU and ResNet to form a fusion neural network.Finally,in the simulation,the data of single-path channel,deep sea channel,and shallow sea channel were generated,respectively,as collected data sets to verify the modulation recognition performance of the net-work,and the real underwater acoustic communication data were collected in the actual sea test as well to ascertain the per-formance of the network.The proposed network in this paper were compared with the traditional neural networks such as vi-sual geometry group network(VGG),long short-term memory(LSTM),gated neural unit,convolutional long short-term memory network(CNN-LSTM).The simulation results demonstrate that the recognition rate of the proposed network struc-ture can reach 100%when the signal-to-noise ratio is 8 dB or higher under the time-invariant underwater acoustic channel,and the modulation recognition under the time-varying underwater acoustic channel outperforms the above-mentioned tradi-tional network.The sea test shows that the modul

关 键 词：水声通信 调制识别 深度学习 循环网络 残差网络 

分 类 号：TN911.3[电子电信—通信与信息系统]