A Study on Data-driven Methods for Selection and Evaluation of Beam Subsets in 5G NR

Abstract: 5G New Radio is the next generation of mobile networks and it comes with promises of ultra-high speeds, ultra-high reliability and ultra-low latency. This has posed a challenge for the engineers entrusted with the task of finding solutions which could fulfil the specification, and as a result, some promising areas have received increased attention in recent years. Signal processing techniques like directional transmission and reception, also referred to as beamforming, is believed to be a key area of advancement to meet the requirements. The general benefit of applying beamforming is to concentrate the power of a signal in a given direction, which also translates to a smaller area of coverage on the ground surface. As a user traverses the area surrounding a base station, it will thus need to keep an updated image of the radio environment including available beams for transmission and reception. This is done through a frequent and periodic synchronization signal procedure which, partially due to carrier aggregation, intensifies the workload put on the often battery powered user equipment as it continuously updates and stores state variables related to available physical channels. In order to reduce required processing power and cache memory usage in the user equipment, telecommunications engineers have introduced beam subsets which each user may operate on rather than the full set of available beams. This report investigates purely data-driven and machine learning based alternatives to the current, static, implementation responsible for selecting and evaluating beam subsets with the goal of mitigating the downsides posed by not considering all available channels, as is the case with any such subset strategy. The results show that the current static implementation of the subset selector can be improved in terms of reducing the selected subsets which do not include the widebeam of highest signal strength at a cost of more frequent subset updates, which are of relevance for distributed systems where such configurations take place over the air interface. A probabilistic approach involving a Markov Chain yielded the greatest benefit at the highest subset update cost, while a Machine Learning approach involving a Random Forest Regressor offers a smaller improvement at a much lower cost relative to the Markov method. A subset evaluation technique involving a Multi-Layer Perceptron classifier yielded promising results, being able to detect 76.8\% of subsets not including the strongest widebeam. Another evaluation technique based on a convolutional neural network resulted in accuracies of up to 95\% based on various different image inputs. Furthermore, it was shown via experiments performed on the Markov method that the frequency of selected subsets which did not include the best widebeam decrease exponentially with an increasing subset size.