Deep learning-based energy efficient LSFD weights prediction for user centric cell free massive MIMO system

Abstract

Cell free massive Multiple Input Multiple Output (mMIMO) is expected to be utilized in Sixth Generation (6G) mobile generation as it provides high macro diversity gain and uniform coverage. Access Point (AP)–User Equipment (UE) association is one of the main problems in cell free mMIMO. In this paper, a joint Large Scale Fading Decoding (LSFD) and AP-UE association is studied to reduce the computational time while achieving high energy and spectral efficiencies. Two deep neural network models are proposed called Per User Equipment Deep Neural Network (PUEDNN) and Per Access Point Deep Neural Network (PAPDNN). PUEDNN model predicts the LSFD weights between each UE and all APs, while PAPDNN model has the advantage of predicting the LSFD weights between each AP and all UEs. Accordingly, this model could be implemented in a more distributed fashion at each AP. These models are trained using dataset generated from heuristic sparse LSFD optimization algorithm, this allows the models to learn the sparsity nature of the system and apply AP-UE association based on the values of the predicted LSFD weights at the receiver side while using the large scale fading coefficients as the models’ input. Simulation results show that the computational time of both PUEDNN and PAPDNN models is reduced by 74 % and 92 % compared to optimum LSFD and sparse LSFD designs, respectively. Furthermore, PUEDNN enhanced the EE significantly compared to optimum LSFD and sparse LSFD designs, respectively, while PAPDNN outperforms the EE of optimum LSFD. Moreover, both models achieve comparable SE compared to previous heuristic designs. Finally, the proposed models are simulated using different parameter settings to validate their robustness, and complexity analysis is conducted for the models’ inference.