Beam selection using Machine Learning in Massive MIMO systems

Abstract: In mobile communications the demand for increased capacity and speeds is a constant. 5G NR is the latest generation which is now seeing widespread adoption. Crucial technologies like Massive MIMO and beamforming are enabled by the use of large antenna arrays. While these arrays allow a 5G network to handle larger amounts of data at higher speeds they do have drawbacks. Specifically, the power requirements associated with using many large arrays are sizable. Antenna selection is a potential solution to this problem, allowing the antenna array to use only a subset of elements at off-peak times. Selecting the most favorable beams for transmission and reception is an important step towards achieving this goal. The aim of this thesis is therefore to apply machine learning techniques to this problem in order to predict the locations of optimal beams. Throughout the testing process several machine learning models and approaches to solving the problem were explored. In each step of the process the prediction accuracy was evaluated and improvements were made. Several of the models displayed the ability to make accurate predictions, which will aid in solving the antenna selection problem. The highest accuracy was achieved when predicting one of two predefined beam regions based on Precoding Matrix Indicator-values using a Deep Neural Network model. This test yielded a prediction accuracy of 87.4%.