Downlink CSI Limited Feedback Method In FDD Massive MIMO Systems Based On Deep Transfer Learning

In a wireless communication system,the acquisition of downlink channel state information(CSI)is very importment for the base station(BS)to ensure the quality of wireless communication.Compared with the traditional methods,the use of deep neural network(DNN)can effectively compress the downlink CSI,so as to greatly reduce the feedback overhead.However,the generalization of DNN is poor,therefore,it is necessary to retrain a DNN for the new wireless channel environment.Nevertheless,in massive multiple input multiple output(MIMO)systems,retraining a DNN faced huge time and data overhead.This paper mainly discusses CSI limited feedback algorithm in FDD massive MIMO systems.The main contents are as follows:First of all,a downlink CSI feedback method based on deep transfer learning(DTL)is proposed for different wireless channel environments in ubiquitous communication scenarios.In this paper,the pre-trained model is transfered to other wireless channel environments,and a small number of data is used to fine-tune the pre-trained model,and a new model with excellent performance can be obtained soon.The experimental results show that only a small number of training cost is required in the whole process,which proves the effectiveness and superiority of DTL.Secondly,considering the high real-time requirement in communication systems,a downlink CSI feedback method considering the balance of training cost and model performance is proposed.This paper explored the effects of different numbers of samples,iteration times and network layers on the model performance in the process of DTL.The experimental results show that on the premise of a certain decline in model performance,this method further reduces the training cost of the DNN,and can quickly get a model with good performance.Finally,a downlink CSI feedback method based on model-agnostic meta-learning(MAML)is proposed to solve the problem that the pre-trained model needs a large number of samples.Based on a good weight initialization learned by using the small datasets of different wireless channel environments,a small number of samples of different wireless channel environments are used to train the DNN to obtain the model of different wireless channel environments.Compared with the model obtained by DTL,the experimental results show that the model obtained by MAML algorithm has slightly reduced in terms of NMSE,but still shows excellent performance.