Research On Deep Learning Based Downlink CSI Acquisition In FDD Massive MIMO Systems

Massive multiple-input multiple-output(mMIMO)wireless communication technology which equips hundreds of antennas at the base station,has become one of the key enabling technologies of the future wireless communication system.mMMO fully utilizes wireless resources in the space domain,thus improving the spectral efficiency and energy efficiency of systems.However,with the significant increase of antennas,a series of transmission technology problems have been triggered,making it more challenging to explore the wireless transmission technology applied in the mMIMO communication systems.The base station(BS)needs to obtain the downlink channel state information(CSI)for downlink precoding and resource allocation to improve the frequency and energy efficiency of the system.In frequency division duplexing(FDD)mode,there is no channel reciprocity between the uplink and downlink,so the BS needs to send pilot signals to the user,and the user then estimates the downlink CSI from the pilot and compresses it back to the BS.Since the downlink pilot overhead and uplink feedback overhead increase in proportion to the number of antennas,the acquisition of downlink CSI has become the bottleneck restricting the development of FDD-mMIMO technology.In recent years,the artificial intelligence(AI)based downlink CSI compression and feedback scheme has shown excellent results.However,this vibrant research field mainly focuses on the design of AI algorithms for downlink CSI compression and feedback at a specific compression rate(CR),which limits the application.The change in the actual communication environment will result in the need to adjust the CR.Therefore,it is of significance to design a flexible compression rate adjustment mechanism.This thesis focuses on the application limitations of the downlink CSI compression and feedback technology in the FDD-mMIMO system.The specific research work is as follows:First of all,facing the complex and changeable actual communication scenario,this thesis proposes a downlink CSI feedback algorithm multiple CRs based on module sharing.Specifically,a compression module is designed to compress codeword sequences with multiple different CRs.And a filling module is added to the decoder to solve the problem of dimensional misalignment of codeword sequences.By sharing partial modules,a universal encoder and decoder are realized,which can compress and decompress the downlink CSI at different rates.The simulation results prove the effectiveness of the algorithm from the perspective of recovery performance and complexity.In addition,facing the dynamic compression scenario,this thesis proposes the concept of dynamic compression of downlink CSI and designs a downlink CSI dynamic compression feedback algorithm using deep transfer learning(DTL).Specifically,DTL is used to mine the common features across CRs in the pre-trained compression and feedback network.And by transferring these common features,DTL fine-tunes the compression and feedback network with different CRs at a small training cost.To reduce the training cost,the impact of various factors on the transfer effect is further explored to optimize the transfer strategy.The simulation results show that the algorithm can achieve a flexible downlink CSI compression and feedback with a small training cost while maintaining a small performance loss.