| As one of the key technologies in the new generation of mobile communication systems,massive multiple-input multiple-output(MIMO)offers advantages such as high system capacity,reliability,and strong robustness.However,in order to fully utilize these advantages,it is essential to accurately obtain the channel state information(CSI)of the uplink and downlink channels.In a Frequency Division Duplex(FDD)system,as the uplink and downlink channels use different transmission frequencies,the base station cannot directly infer the downlink CSI from the uplink CSI based on channel reciprocity.In recent years,utilizing deep learning techniques to establish neural network models at the base station for predicting downlink CSI has become a research hotspot.However,some small base stations have insufficient data to train effective prediction models.To address this,this thesis introduces the concept of distributed CSI prediction and proposes a downlink CSI prediction scheme based on federated learning and deep learning.The specific research work is as follows:Firstly,this thesis establishes a deep learning model for predicting downlink CSI.In FDD massive MIMO systems,where there are a large number of antennas and channel reciprocity is not available,obtaining CSI is extremely challenging.This thesis proposes a deep learning-based downlink CSI prediction scheme that utilizes the mapping relationship between the uplink and downlink channels and uses a data-driven approach to achieve the goal of predicting downlink CSI based on uplink CSI through an offline training and online deployment process.To make the neural network more suitable for processing high-dimensional CSI data,this thesis designs the 3D-Csi Net network model,which uses 3D convolution instead of 2D convolution and improves the feature extraction and residual network modules.Simulation results show that the proposed 3D-Csi Net model has higher prediction accuracy and generalization ability than the traditional Csi Net model.Next,this thesis proposes a distributed downlink CSI prediction scheme based on federated learning.In this scheme,edge small base stations train local models on local datasets and send the trained model parameters to the central macro base station.The macro base station aggregates and updates these model parameters and sends the latest global model parameters back to the small base stations.The entire training process is iterated through the federated learning framework until convergence is reached.Experimental results show that compared to the centralized learning method that requires uploading local data,the federated learning method can significantly reduce the amount of transmitted data during model training and lower communication costs.Finally,this thesis proposes an efficient model aggregation algorithm for federated learning.The model aggregation process at the macro base station is a crucial part of the entire federated learning framework,and the aggregation algorithm also determines the final model performance.Traditional federated averaging(Fed-Avg)algorithms perform poorly in both model performance and convergence speed,and it is difficult to train a qualified global model when the local data is imbalanced.To address this,this thesis proposes a new model aggregation algorithm,Fed-WG,in which the central macro base station receives local model weight and gradient information from edge small base stations and performs two rounds of continuous parameter updates on the global model.Additionally,the algorithm adds a feedback mechanism during the global model training process to improve the model’s generalization ability.Simulation results show that compared with the traditional Fed-Avg algorithm,the proposed Fed-WG algorithm can effectively accelerate the convergence speed of the global model and improve its prediction accuracy. |
