Research On Receiving Algorithm Of OFDM Underwater Acoustic Communications Based On Deep Learning

Underwater acoustic communication(UAC)is one of the important means of underwater information acquisition.Orthogonal frequency division multiplexing(OFDM)has become one of the important modulation technologies in UAC because of its resistance to channel frequency fading and high spectral efficiency.Traditional receiving algorithms rely on channel estimation,channel statistics and mathematical model,and their performance is limited in UAC environment with complex and changeable channels and difficult modeling.With the increasing computing power of hardware system,deep learning algorithm shows great potential in the field of communication.Therefore,this paper studies the UAC receiving algorithm based on deep learning,overcomes the shortcomings of the traditional receiving algorithm in OFDM UAC system,and improves the quality of UAC.The main contents of this paper are as follows:1.This thesis describes the signal model and channel model of OFDM UAC system,and summarizes the classical OFDM UAC system model based on deep learning.In addition,two underwater acoustic channel data sets are constructed based on Rayleigh simulation channel and bellhop simulation channel,and combined with the real sea area channel data set to jointly drive the deep learning algorithm.2.The receiving algorithm of encoded decoded convolutional neural network(EDCNN)in OFDM UAC system is proposed.The algorithm uses the convolution characteristics of convolution neural network and image filling to reduce the serious inter carrier interference in OFDM UAC system;By sliding convolution and mean pooling,the network parameters are reduced to avoid the dimension explosion of fully connected neural network in a large number of subcarrier OFDM UAC systems;Extracting signal features based on multi-layer parallel feature map to process complex received signals in communication system;According to the idea of auto-encoder,the algorithm structure is optimized to realize the compression and reconstruction of received signal to remove noise.Simulation results show that the bit error rate(BER)of EDCNN algorithm is lower than that of linear minimum mean square error(LMMSE)algorithm;And it is lower than LMMSE algorithm in the UAC scene with a small number of pilots and no cyclic prefix.Simulation experiments are carried out based on the number of different subcarriers,and the computational complexity is analyzed.Simulation experiments are carried out based on the number of different subcarriers,and the computational complexity is analyzed.In the OFDM UAC system with 512 subcarriers at 20d B signal-to-noise ratio,EDCNN algorithm reduces the BER by 29.93%and time consumption by 97.75%compared with full information LMMSE.3.This thesis designs unsupervised domain adaptation neural network receiving(DANR)algorithm to avoid dependence on a large number of real underwater acoustic channel data,so as to reduce the acquisition cost and improve the generalization ability of OFDM UAC receiving algorithm based on deep learning.The algorithm uses principal component analysis to evaluate the channel complexity,generates a large number of received signal samples by simulating the channel,extracts the channel common features and detection signals by using domain adaptive neural network,selects the migration channel environment according to the channel complexity,and constrains the learning direction of neural network based on generation countermeasure normalization.Simulation results show that the BER of DANR is not only significantly lower than that of EDCNN algorithm when the channel environment does not match,but also lower than that of LMMSE and least square(LS)algorithm.4.The experimental results based on the pool and the Penghu and Zhoushan wave energy power generation platforms off the coast of Zhuhai City,Guangdong Province show that the BER performance of the proposed EDCNN and DANR algorithms are better than LMMSE algorithms after channel decoding.Compared with LMMSE algorithms,for more than 9000pieces of data transmitted in a 3×0.85×0.25 m~3 pool,EDCNN algorithm reduces the BER from0.0441 to 0.DANR algorithm can achieve zero BER when using 10%of the data of the original received signal data set;In the sea trial with a water depth of about 16 meters and a distance of749 meters,EDCNN algorithm reduces the BER from 0.2023 to 0.0405 compared with LMMSE algorithm.Further,DANR algorithm achieves a BER performance of 0.0441 when using 5%of the data volume of the original received signal data set.