Research On Underwater Acoustic Multi-carrier Signal Joint Detection Technology Based On Deep Learning

At present,underwater acoustic communication has become a necessary means for underwater information transmission.However,unlike land-based wireless channels,the complexity and variability of the marine environment led to underwater channels with physical characteristics like narrow bandwidth,long time delay,and strong multipath,which easily affect the quality of information transmission.Therefore,the exploration of high speed and stable underwater acoustic communication technology has become the focus of research.Orthogonal Frequency Division Multiplexing(OFDM)has become the standard waveform in underwater acoustic multicarrier communication systems,however,it has high out-of-band leakage,and the system performance will be greatly reduced when the frequency bias is serious.Filter Bank Multi-Carrier(FBMC)uses a prototype filter with excellent timefrequency focusing characteristics to reduce band leakage and relax strict orthogonality between sub-carriers,which can further resist inter-code crosstalk and has great potential for application in hydroacoustic communication.However,due to the presence of imaginary part interference in FBMC,the existing OFDM related channel estimation methods cannot be applied directly,and the signal interference needs to be estimated or cancelled using a complex frequency guide design,which also poses a challenge for efficient detection at the receiver side.At present,the use of artificial intelligence methods to improve communication transmission performance is gradually becoming a research hot topic,and deep learning theory is tried to be applied to the physical layer design of communication systems,and the construction of intelligent transceiver models becomes a future development trend.To address the above problems,this thesis investigates the channel estimation and equalization methods for hydroacoustic FBMC systems,and the main research contents are as follows:Firstly,the transmission characteristics of underwater acoustic channel and the related principles of deep learning are introduced.The effects of Doppler effect,multipath effect,etc.on the underwater acoustic communication system are elaborated,and the channel model is also given.Then,starting from the most basic neuron structure,we introduce the composition and hyperparameter settings of the classical fully connected neural network,and analyze the training process such as forward transmission and backward propagation of the neural network.Next,we analyze the basic principles of the FBMC system,including the construction of the model and the derivation process of the fast implementation method.For the problem that the subcarriers remain orthogonal only in the real domain,which makes it difficult to estimate the channel information accurately because the symbols at each time-frequency grid point are affected by the interference of the imaginary part,we proposed a channel estimation method based on interference approximation and analyzed the system performance under three block-guided designs.In addition,we investigate a channel equalization algorithm based on virtual time-reversal mirror to eliminate the effect of inter-code crosstalk on the signal,and verify the system performance by simulation experiments.Then,to address the problems of traditional FBMC system with large guide frequency overhead at the receiver side,a convolutional neural network-based signal joint detection method for the underwater acoustic FBMC(CNN-FBMC)system is proposed.Using neural networks to replace the traditional receiver-side channel estimation and equalization based on the frequency guide of the communication module,and we also compare the BER performance of CNN-FBMC system with Deep Neural Network based FBMC(DNN-FBMC)and conventional hydroacoustic FBMC system by simulation experiments.Finally,a Autoencoder network with a self-supervised mechanism is introduced into the system transceiver,and proposing an end-to-end underwater acoustic FBMC communication system based on convolutional autoencoder.By using the compressed reconfiguration process of the compiled code network to fit the constellation mapping and symbol demodulation process of the transceiver of the communication system respectively,more degrees of freedom are given to the system.At the same time,the dense connection mode is used in the decoding network to realize the multifeature multiplexing of the signal.Simulation results show that the end-to-end system BER performance based on convolutional Autoencoder outperforms the least-squares estimation and virtual time-reversal mirror equalization methods.