Deep Learning-Based CSI Estimation Using Synthetic Dataset

Abstract

Multiple-input Multiple-output (MIMO) plays a vital role in 5G technology. The MIMO transmission is beneficial only when Channel State Information (CSI) is known. However, gathering CSI provides challenges like high dynamic channel and feedback overheads. In this paper, a CSI estimation technique using deep learning techniques is proposed for highly dynamic vehicular networks. The propagation environments like scatters and reflectors are almost the same. This allows the designed Deep Neural Network (DNN) architecture to experience with negligible overhead with the non-linear CSI relations. The proposed method focuses on synthetic data generation using MATLAB and training the neural network which will be used for the prediction of the CSI of the eMBB channel in the same cell. Hence resulting in the overhead reduction and increases in the threshold violations. The network used for the training results in proper convergence of the loss function. This indicates a reduction of CSI overhead of eMBB vehicles in order realize non-linear CSI with the DNN model. MATLAB is used to generate synthetic data. The trained neural network is implemented using TensorFlow 2.0 framework in python 3.6.0. using the Nvidia RTX2070ti graphic Card with CUDA and CUDNN support. From the simulation study, results show that the angular domain transformation is more logical with the real-time data as the data is already in complex number form so, to make some similarities, to improve in the estimation of CSI.