Research On Beamforming Algorithm Based On Unsupervised Learning In Massive MIMO Scenario

In order to cope with the rapidly growing traffic demand,massive multi-input multi-output(massive MIMO)technology has received great attention in the industry.The difficulty of balancing real-time and performance with traditional mathematical algorithms has led academics to look at deep learning(DL),which can model non-linear functions with relatively low complexity and guarantee good performance.However,the scene environment in massive MIMO networks is complex,the scale of input data is large,and the resource scheduling is complicated.The existing deep learning-based beamforming model has poor scalability due to its simple structure,and the performance of this single-structured model degrades severely as the scene scale increases.Therefore,this thesis proposes a new model structure for massive MIMO beamforming to mitigate the model performance degradation;furthermore,beamforming in massive MIMO user movement scenarios is analyzed and a multi-branch deep learning model is designed to improve the system performance.The main work of this thesis is divided into two parts as follows.To address the problem of single model structure of the existing massive MIMO beamforming model,this thesis proposes an unsupervised convolutional neural network(CNN)based on the Inception structure to effectively avoid the problem of gradient disappearance and model performance degradation that tends to occur in deep networks in complex scenarios,unsupervised learning is used to further explore the potential of the model and improve its performance.The simulation results verify the robustness and scalability of the model architecture proposed in this thesis.To address the problem that the channel state information(CSI)reported by users in high-speed mobile scenarios is not ideal,which leads to frequent beam training between base stations and users,this thesis proposes a beamforming model based on multi-branch Inception neural network,which directly gives the beamforming matrix of the mobile user by combining the user’s channel state information and the user’s location information,avoiding the system performance degradation caused by the large number of beamforming training and the performance degradation caused by the use of codebooks.In addition,considering the fluctuation of the system environment due to user movement,this thesis uses transfer learning and meta-learning algorithms to improve the adaptability of the model in different environments.Finally,the simulation verifies the feasibility and advantages of the multi-branch structure,and analyses the performance gain from multiple source inputs.