Method Of Underwater Acoustic Communication Signal Recognition Based On Data Augmentation And Transfer Learning

As an important part of underwater communication countermeasures,underwater acoustic communication reconnaissance plays a key role in underwater network-centric warfare.The underwater acoustic communication signal recognition technology is the basis for subsequent coherent interference and information deciphering.In recent years,deep learning methods represented by Convolutional Neural Networks(CNN)have achieved remarkable results in the field of signal modulation pattern recognition and underwater acoustics.However,due to the lack of data,the underwater acoustic communication signal recognition technology based on CNN cannot be effectively applied to underwater acoustic communication reconnaissance.Therefore,in this paper,aiming at the measured underwater acoustic communication signals,the data augmentation is achieved through the generative adversarial network,and the transfer learning method is used to improve the recognition accuracy of the underwater acoustic communication signal of the convolutional neural network.The main research content of this paper is divided into the following three parts:First of all,this paper designs a recognition scheme that combines inter-class recognition and intra-class recognition to realize the inter-class recognition of MPSK,OFDM,DSSS and MFSK signals and the intra-class recognition of MPSK signals and MFSK signals.The simulation experiment results show that the CNN model designed in this paper can realize underwater acoustic communication signal recognition and has good noise immunity.The Songhua Lake test results show that under the condition of sufficient data,the scheme proposed in this paper can realize the recognition of the underwater acoustic communication signal,but under the condition of insufficient data,only the 2DCNN-S model based on cyclic spectrum characteristics can realize the recognition of 2FSK and 4FSK signals.Afterward,in response to the insufficient data of underwater acoustic communication signals,this paper proposes an underwater acoustic communication signal data augmentation scheme based on Generative Adversarial Network(GAN),designs Underwater Communication Signal GAN(UWACS-GAN),and carries out an experimental study of underwater acoustic communication signal recognition technology based on data augmentation.The experimental results show that compared with the use of simulated signals for data set expansion,the use of generated signals based on UWACS-GAN for data augmentation shows better recognition results.For the inter-class recognition of four modulation patterns of signals,the recognition accuracy of the 1DCNN-DA model based on UWACS-GAN generated signal data augmentation reaches 77.13%,and the recognition accuracy of the 2DCNN-DA model reaches73.17%;for the intra-class recognition of MPSK signals,the recognition accuracy of the2DCNN-DA model based on UWACS-GAN generated signal data augmentation reaches84.67%.In order to further improve the recognition accuracy of underwater acoustic communication signals under the condition of insufficient data,this paper carries out an experimental study of underwater acoustic communication signal recognition technology based on transfer learning.The experimental results show that the underwater acoustic communication signal recognition technology based on transfer learning ccan improvethe recognition effect more effectively,and close to the recognition effect under the condition of sufficient data.For the inter-class recognition of four modulation patterns of signals,the recognition accuracy of the 1DCNN-GTL model based on the generated signal recognition model transfer reaches 83.92%,and the recognition accuracy of the 2DCNN-GTL model reaches 91.42%;for the intra-class recognition of MPSK signals,the recognition accuracy of2DCNN-GTL model based on the generated signal recognition model transfer reaches 86.00%.Based on all the research results,the underwater acoustic communication signal recognition scheme designed in this paper finally uses the 2DCNN-GTL model based on the UWACS-GAN generated signal recognition model transfer to realize the inter-class recognition and the intraclass recognition of MPSK signals,and the 2DCNN-S model based on spectral characteristics to realize the intra-class recognition of MFSK signals.