Deep Learning-assisted Efficient Receiving Method Of Wireless Communication

In the next decade,as the information society becomes highly digitized and intelligent,wireless communication networks need to provide comprehensive coverage to realize the internet of everything of the ultra-massive machine-type communication(um MTC)scenario.For this reason,the 5th generation mobile communication system(5G)will evolve to the 6th generation mobile communication system(6G).In order to realize the visions of 6G,researchers use technologies such as massive multiple input multiple output(MIMO)and grant-free transmission to tap time-frequency and space resources.However,in the application of these technologies,traditional wireless communication signal receiving methods have problems such as low accuracy and high complexity.Deep learning has powerful data-driven characteristics and data processing capabilities,which make it can be used as an optimization tool in wireless communications.Therefore,this thesis focuses on massive MIMO and grant-free transmission scenarios,and studies deep learning-assisted active user detection and channel estimation algorithms and massive MIMO signal detection algorithms.First,for the active user detection and channel estimation problems in the grant-free transmission system,a learning iterative soft thresholding network(LISTnet)based on deep unfolding is proposed.In the uplink grant-free transmission system,the multiple measurement vectors problem of active user detection can be converted into a single measurement vector(SMV)problem by the Khatri-Rao product,which can support more active users.Based on this,the iterative soft thresholding(IST)algorithm is used to solve the SMV problem,and the LISTnet is constructed through the deep unfolding of the IST algorithm to train the optimal parameters of the IST algorithm.Then,in order to further reduce the complexity of the network,a LISTnet-based residual learning iterative soft thresholding network(Res LISTnet)is proposed,which dynamically adjusts the number of effective layers of the network by deleting the redundant layers of LISTnet and further reduces the complexity of the network.The simulation results show that the active user detection performance of the proposed LISTnet is better than IST,traditional compressed sensing algorithms and detection algorithms based on deep neural networks,and Res LISTnet further reduces the complexity of the network while maintaining the performance of active user detection.Second,for the um MTC scenario,a modified hybrid message passing(MHMP)algorithm under the generalized spatio temporal traffic model(GSTTM)is proposed to solve active user detection and channel estimation problems.It is assumed that the devices evenly distributed within the coverage of the base station transmit data to the base station through the alarm event-driven mode.By defining the input of the alarm trigger probability function as the distance between the device and the alarm event,the device traffic is spatially correlated;by modeling the state of devices as a two-state markov modulated poisson process,the device traffic is temporally correlated,thus establishing GSTTM.Then,when performing active user detection and channel estimation under GSTTM,in order to make full use of the spatial-temporal correlation of the model,the historical active state information of the devices is input into a multi-layer long short-term memory(LSTM)network to obtain the predicted prior active probability of the devices at the current moment,and pilot scheduling is performed by calculating the orthogonality of the channel covariance matrices of different devices.The predicted active probability of the device and the scheduled pilot matrix are used as the initial input of the hybrid message passing(Hybrid Message Passing,HMP)algorithm to obtain the MHMP algorithm.The simulation results show that the MHMP algorithm has better active user detection and channel estimation performance than the traditional compressed sensing algorithm and HMP algorithm.Finally,aiming at the problem of massive MIMO signal detection,a soft output MIMO signal detection method based on graph neural network(GNN)is proposed.Starting from the maximum likelihood(ML)detection,an approximate iterative message passing method is derived,in which messages are directly passed between variables,and the messages is accumulated at the destination variable.In this way,a fully connected undirected graph is constructed,where the received information at the receiving end is used as the node feature and edge feature of the graph,and the message passing algorithm runs on this graph.Then,the corresponding GNN detector is designed based on the undirected graph to learn the message passing rules,where the nodes of the network finally output the estimated log-likelihood ratio of the corresponding transmission bits,and the detection result of ML is used as the network training label.The simulation results show that in the two scenarios of Rayleigh fading channel and bursty noise channel,the GNN detector can be directly applied and achieved achieves detection performance close to or even better than that of ML.In addition,compared with the traditional detection algorithm,the GNN detector has good robustness to the signal-to-noise ratio.