Research On Adaptive Reconstruction Method Of 5G 3D Channel Sample Space

In the development of 5G technology,Massive MIMO has been widely used due to its advantages of greatly improving spectral efficiency and system capacity.During the 5G network planning phase,it is necessary to evaluate the performance characteristics of Massive MIMO.However,due to the large number of antennas in the network and its high complexity,the channel matrix obtained by the traditional systemlevel simulation method will take huge time and computational cost,which cannot meet the requirements of the actual network planning.Therefore,this paper combines with deep learning and uses BP neural network to predict the channel amplitude characteristics of Massive MIMO systems,which can quickly predict the real network,accurately evaluate the performance and reduce the planning time.The main research contents of this paper are as follows:1.Based on the ray-tracing data of the real environment and the 3D Massive MIMO channel model,this paper generates channel matrix by using system-level simulation platform,and further obtains channel amplitude to provide real values of training samples and prediction samples for BP neural network.2.Based on the analysis of the prediction results of the BP neural network,it is found that there are many high-prediction-error users.The fundamental reason is that the ray-tracing data sample space is incomplete and the training samples are missing.Therefore,this paper proposes an adaptive conditional variational auto-encoder with multiple learning distributions.The new samples generated by the model are used to supplement the original training set.Then the BP neural network is trained with the improved training set to reduce the number of high-prediction-error users and the prediction-error.3.In order to generate reliable new samples,this paper deeply studies the characteristics of sample space and classifies samples based on the sparsity of sample space.Based on the reconstruction ability of the adaptive conditional variational autoencoder,the most suitable distribution of each sparseness sample set is obtained.Then,the most suitable conditional variational auto-encoder for each sample set is determined to make the new sample distribution more consistent with their characteristics.The simulation results show that the BP neural network obtained by the proposed method is not only very close to the accuracy of system-level simulation,but also reduces the time consumption and computational cost.The simulation results are verified by real 5G environment data.Therefore,the effectiveness of the method in this paper is confirmed,which provides important guidance for the planning of the actual network and the application of deep learning tools to the base station side.