Theory And Key Technologies Of Transceiver Design For Intelligent Communications

The fifth-generation mobile communication(5G)systems commercialized in 2019 can achieve air interface enhancement,spectrum expansion,and network intensification by largescale multiple-input multiple-output(MIMO),millimeter-wave communications,and ultradense networking.However,with the initial deployment of 5G commercial systems,the advantages and disadvantages of the main core enabling technologies have become increasingly significant.Mobile communications are facing challenges in development models,architecture,and security,making it difficult to meet the explosive growth of mobile data traffic and business diversification Demand.Therefore,it is urgent to start research on the 6th Generation Mobile Communication Systems(6G).Among the candidate technologies for next-generation mobile communication systems,the intelligent communication with the help of artificial intelligence to enhance the wireless transmission efficiency has become the international frontier.At present,related theoretical methods and key technologies in physical layer communications are still in their infancy,requiring a lot of research work.This thesis takes model-driven deep learning as the core,and conducts research on advanced transceiver design theory and key technologies for intelligent communication.First,we study the model-driven-deep-learning-based advanced transceiver design for the physical layer of intelligent communication.After introducing the concept and principle of model-driven deep learning in detail,and aiming at the problem that data-driven deep learning network does not introduce expert knowledge in the communication field,the modeldriven-deep-learning-based transceiver design is proposed.We explain the model-driven-deeplearning-based transceiver design theory and key technologies from MIMO detector,OFDM receiver design,massive MIMO channel state Information(CSI)feedback and precoding design,and point out the direction for subsequent research on advanced transceiver design for intelligent communications.Secondly,we study the advanced receiver design for millimeter-wave MIMO orthogonal frequency division multiplexing(OFDM)system with low-resolution analog to digital converter(ADC).Aiming at signal distortion caused by low-resolution ADC,a generalized expectation consistent signal recovery(GEC-SR)algorithm is proposed for signal detection in millimeterwave MIMO-OFDM systems.By using the replica method in statistical physics,the state evolution equation of the GEC-SR algorithm is derived,and the performance limit of the algorithm is analyzed.Based on the analytical framework,we obtain the symbol error rate(SER)performance loss caused by the 3 bit quantization does not exceed 1.02 d B,and it reveals how the high-resolution ADC and low-resolution RF chain impact the system SER performance.The simulation results prove the effectiveness of the proposed algorithm and the accuracy of the analytical framework.Subsequently,we investigate the model-driven-deep-learning-based channel estimation methods for the millimeter-wave beam-domain massive MIMO system.Aiming at the narrowband three-dimensional lens antenna millimeter wave massive MIMO system,a channel estimation network based on the learning denoising approximate message passing algorithm is proposed to achieve accurate CSI acquisition,and further considering the challenges brought by the beam squint phenomenon in the broadband millimeter wave system.A channel estimation network based on the learning and denoising expectation consistency signal recovery algorithm is proposed.The Steins unbiased estimator is used to achieve unsupervised training of the network,and accurate CSI can still be obtained under a small number of RF links and lowprecision ADCs.The simulation results show that the performance of the two model-drivendeep-learning-based channel estimation networks are significantly better than the traditional compressed sensing algorithm.Then,we study the model-driven-deep-learning-based MIMO detection algorithm for point-to-point MIMO systems.By unfolding the traditional orthogonal-approximate-messagepassing-based MIMO detector into the deep neural network and introducing the trainable parameters reasonably,we can obtain the network structure.We further consider imperfect CSI,introduce channel estimation error in the model-driven deep learning MIMO detector,and improve the channel estimation performance through the data-aided channel estimation method to achieve the joint channel estimation and signal detection for the MIMO system.Simulation results confirm that model-driven deep learning can accelerate the convergence of iterative MIMO detectors,improve data detection performance,and show strong robustness to system parameters mismatches.Finally,we study the model-driven-deep-learning-based finite-alphabet and constant envelope precoding for multi-user massive MIMO systems.Aiming at the problems of traditional iterative precoding algorithms that require a large number of iterations and computational overhead,the model-driven-deep-learning-based multi-user massive MIMO precoding method is proposed.The traditional iterative discrete signal estimation and Riemannian manifold optimization precoder are unfolded into the deep neural network,and trainable parameters are added to adjust the search step size,gradient direction and damping factor,reducing the multi-user interference,computational overhead and the number of iterations.Simulation results demonstrate that the two model-driven-deep-learning-based precoders can improve the bit error rate performance and exhibit robustness to channel estimation errors and channel model mismatches.