Research On Deep Learning-Based Physical Layer Algorithms For Underwater Acoustic Communications

With the increase in demand for marine resources,our nation’s maritime strategy is gradually moving from the offshore to the deep ocean and from the region to the world.There is an urgent need to develop underwater information acquisition technologies vigorously.Underwater acoustic(UWA)communications are the most effective way to achieve long-distance wireless transmission,which has significant research value.However,UWA channel delay,Doppler,and multipath significantly challenge various UWA communication systems.In particular,the time-space-frequency three-dimensional variation characteristics of UWA channels make it difficult to model accurately.Correspondingly,wireless communications have been well developed in recent years,from information theory to channel modeling.However,the discrepancy between theoretical models and experimental data motivates us to develop intelligent communication systems.Artificial Intelligence(AI)based communications enhanced by Deep Learning(DL)have achieved preliminary results in wireless communication.However,DL is sensitive to data,and it is impossible to copy the wireless terminal DL method directly due to the particularity of UWA channels.To this end,DL-based UWA communications are studied in this thesis.Orthogonal Frequency Division Multiplexing(OFDM)and Chirp spread spectrum communications are typical high spectral efficiency and robust communication examples,respectively.Using DL-based methods to solve mathematical model mismatch problems is focused on.Various data-assisted DL receivers are proposed,which support effective and reliable wireless underwater information transmission.The main contributions of this thesis are as follows:(1)An UWA channel impulse response(CIR)dataset,dubbed UWA-CIRset,is established.There is yet to be a public and unified data set to support DL-based UWA communications.To this end,a UWA channel dataset is established based on multiple simulation models and multiscenario measured data.(2)A channel estimation and equalization integrated receiver based on multi-task learning(MTL)is proposed,called TaskNet,to improve the DL-receiver’s generalization.The mismatch between the training data set and the test environment degrades the receiver performance for UWA communications.TaskNet considers that the classical receiver contains mutually dependent tasks,including channel estimation,equalization,and demodulation.To demonstrate the advantages of TaskNet,we make a comparison with the classical single-task-based DLreceivers,i.e.,FC-DNN and Com Net.Simulation results show that the proposed MTL-based DL receiver has a better generalization performance than the single-task-based ones,whether the signal suffers nonlinear damage or the training set does not match the real channel.(3)A lightweight DL-receiver based on a pruning method and convolutional neural network(CNN)framework is proposed,dubbed LightNet.Concretely,the pruning and CNN architecture are adopted to lighten the receiver model.Moreover,the pruning algorithm can control the tradeoff between the bit error rate(BER)and the number of network nonzero parameters.Simulation results show that the BER of the proposed LightNet is better than that of MMSE,FC-DNN over measured at-sea channels.(4)A model-driven DL equalizer based on the MIMO-OFDM framework is designed,because the data-driven approach is sensitive to training data and has poor generalizations.The deep unfolding approach is employed and minimum Kullback-Leibler(KL)is adopted with the one-hot coding instead of the MMSE criterion for MIMO-OFDM equalizer.Furthermore,we apply the measured at-sea UWA channel to evaluate the proposed equalizer.Experimental results show that the proposed algorithm performs better with low computational complexity.Concretely,the BER of the proposed method is nearly an order of magnitude lower than that of MMSE.(5)A strategy for selecting a proper receiver under different communication scenarios is proposed.UWA channel is bandwidth-constrained and experiences doubly selective fading.It is challenging to acquire perfect channel knowledge for OFDM communications using a finite number of pilots.On the other hand,DL approaches have great success in wireless OFDM communications.However,whether they will work underwater is still a mystery.In this thesis,we,for the first time,compare two categories of DL-based UWA OFDM receivers,i.e.,the data-driven(DD)method,which performs as an end-to-end black box,and the model-driven(MD)method,also known as the model-based data-driven method,which combines DL and expert OFDM receiver knowledge.We employ LightNet to establish the DD receiver.On the other hand,UDNet is adopted for the MD receiver.We analyze the characteristics of different receivers by Monte Carlo simulations under diverse communications conditions and propose a strategy for selecting a proper receiver under different communication scenarios.Field trials in the pool and sea are also conducted to compare the feasibility and advantages of the DL receivers.It is observed that DL receivers perform better than conventional MMSE receiver in terms of BER.(6)A data-driven Chirp binary orthogonal keying(Chirp-BOK)receiver,dubbed C-DNN,is proposed.Classical matched-filter(MF)receivers are susceptible to synchronization and Doppler.It is difficult to achieve accurate symbol synchronization and Doppler compensation in tough underwater scenarios,which results in a mismatch of mathematical models and a decrease in the BER performance.In response to the above problems,a DL-based Chirp-BOK signal detector with a fully connected network architecture is designed.Meanwhile,data augmentation to enrich the training dataset is employed to overcome the data mismatch problem.The simulations show that C-DNN has better BER performance than the MF receiver.Moreover,a pool experiment is conducted to demonstrate that the C-DNN receiver performs better than the MF receiver under lower SNR conditions.(7)An acoustic radio cooperative(ARC)training scheme for data-driven receivers based on federated meta-learning(FML)is proposed.Under the random scheduling wireless network,distributed training C-DNN receiver improves the receiver BER performance by exploiting the model parameters from multiple buoys.Furthermore,the convergence of the FML framework is theoretically analyzed,and a closed-form solution is derived.Finally,numerical simulation results verify the FML scheme’s superiority.The simulation results show that the proposed FML has better generalization than federated learning-based systems and outperforms the MF-based method after several communication rounds.