APPLICATION OF VARIATIONAL ENCODER IN DEEP LEARNING BASED CHANNEL ESTIMATION IN MASSIVE MIMO SYSTEMS

Abstract

Advancements in physical layer technology in 5G and beyond has necessitated researchers to contribute novel techniques to further the aim of effective implementation of Massive MIMO technology. To achieve high data rates by harnessing multiplexing gains and provide low mean square error rates by exploiting spatial diversity of Massive MIMO has paved way towards dimensionality reduction and to estimate channel out of sparse channel. Due to scalability and model-based paradigm in Massive MIMO, its computational cost and signaling overhead has increased. In time division duplexing mode, the CSI obtained in uplink channel estimation is utilized by the base station to send the feedback by means of reciprocity. Classical CSI estimation methods like LS and LMMSE where problem of getting accurate statistical model, interference management, converge management and certain strategies have limitation and it is difficult to get system parameters and reduce signaling overhead due limitation of scalability and its applicability. In this dissertation, we have studied existing classical models and compared with the Deep Learning method using Variational Autoencoder where, the latent space having smaller dimension which is derived as a probability distribution generated stochastically as normal gaussian distribution after encoding the CSI and decoded by taking the latent attribute from each sample to generate a sample vector. Further, any minor change in the input will result in perfect reconstruction and estimation of CSI, also by reducing the divergence between them and by measuring the KL divergence. Which leads to generation of new data identical to the training data. The above model has led to reduction in root mean squared error and cost function as compared to classical methods.