基于EfficientNet网络的水声通信信号调制识别

Modulation recognition of underwater acoustic communication signal based on EfficientNet network

提出一种基于EfficientNet网络的复杂水声信道下非合作水声通信信号调制识别方法。与传统的深度学习网络相比,EfficientNet网络具有更高的效率和更小的模型大小。该方法通过对2FSK、4FSK、BPSK、QPSK、DSSS-BPSK和OFDM水声通信信号的时频特征进行分析,将二维时频图作为EfficientNet网络的训练集和测试集,使用训练集训练EfficientNet网络模型,并使用测试集确定模型的准确性和效率,完成对水声通信信号的调制识别。实验结果表明:在内场实验中,信噪比大于8 dB时,所研究6种信号的调制识别率均在85%以上;通过外场数据测试,信号调制识别率均在80%以上,验证了该方法在减少模型的大小和计算成本的同时,可保证较高的水声通信信号的调制识别准确率。

To address the difficulty in recognizing hydroacoustic communication signals that are non-cooperative and transmitted under complex hydroacoustic channels,an efficient modulation recognition method based on EfficientNet is proposed.EfficientNet is a deep learning network structure that uses the basic idea of ResNet(residual network)to solve the problem of gradient vanishing in the training of deep network and improve the model’s performance.Compared with traditional deep learning networks,EfficientNet uses a technique called compound coefficient to scale up models in a simple but effective manner.In this method,the time-frequency characteristics of six hydroacoustic communication signals(2FSK,4FSK,BPSK,QPSK,DSSS-BPSK,and OFDM)are analyzed to derive two-dimensional time-frequency maps that serve as the training set and test set of EfficientNet.The training set is used to train the EfficientNet model,and the test set is used to determine the accuracy and efficiency of the model,thus realizing the modulation recognition of hydroacoustic communication signals.The experimental results show that this method guarantees a high modulation recognition accuracy of hydroacoustic communication signals while reducing the size and computational cost of the model.In indoor field experiments,the modulation recognition rate of the six signals under study is above 85%when the signal-to-noise ratio is greater than 8 dB.In outdoor field experiments,signal modulation recognition rates are all above 80%.