Research On Transceiver Design Of Massive MIMO Communication System Baesd On Machine Learning

The fast development of the intelligent device and the Internet-of-things has stimulated various new applications,which pose new challenges on both transmission rate and flexibility of wireless networks.The traditional method depends heavily on the simplified model assumptions and can hardly meet the customization requirements of future applications.To overcome the challenges mentioned above,this thesis considers applying data-driven machine learning techniques to massive Multiple-Input Multiple-Output(MIMO)systems to improve transmission efficiency and flexibility.Specifically,this thesis focuses on designing domain-specified machine learning systems to improve the wireless systems by harnessing properties of the massive MIMO systems.Inspired by the cluster phenomenon of received signals in MIMO systems,a clustering learning based signal detection framework is proposed in this thesis for time-division duplex(TDD)massive MIMO systems,which is a unified framework for flexibly designing both linear and nonlinear signal detection algorithms with different performance requirements.The proposed clustering receiver requires no explicit channel estimation procedure.Furthermore,this thesis proposes a modulation constrained clustering framework based on properties of the MIMO channel and digital modulation for computational efficiency in high order modulation systems.The model complexity of the modulation constrained clustering algorithm is independent of the modulation order.The thesis demonstrates how to approximate the detection performance with perfect channel information under the proposed clustering framework with small pilot overhead.Finally,theoretical analysis and experimental simulations verify the efficiency of the proposed framework.For downlink channel estimation problems in frequency-division duplex(FDD)massive MIMO systems,the dictionary learning based channel estimation system is studied in this thesis.The proposed paradigm aims at designing the cell-specific dictionary to represent the geometry characteristics of wireless channel for better sparse channel estimation.The physical implications of the learned dictionary are theoretical studied to explain the performance improvement incurred by dictionary learning.Building on that,the block-structured dictionary learning algorithm is proposed for better geometric channel representation.Finally,the performance of the proposed algorithms is tested in simulated channel environment according to 3GPP standards.This thesis successfully incorporates machine learning techniques into wireless communications to provide a new data-driven perspective for designing MIMO systems.Based on the specific structure of MIMO channels,efficient machine learning algorithms and systems are proposed for signal detection problem in TDD MIMO and downlink channel estimation problem in FDD massive MIMO,respectively.The performance of the proposed data-driven algorithm designs demonstrates superior performance compared to traditional designs in both theoretical analysis and experimental simulations,which provides a new aspect for research of massive MIMO communication systems.