User Equipment Characterization using Machine Learning

Abstract: With the ever increasing demand for higher data rates and reliability, efficient management of cellular networks remains a challenge. Among other technologies, fifth generation systems are expected to tackle this challenge using large electronically controllable antenna arrays operating in the time division duplex mode. In such systems, spatial beamforming is implemented with accurate channel estimates at the cellular base station (BS). As a consequence of the high channel dimensions, large amounts of data are being collected at the BS which can be further utilized in order to optimize resource allocation, and to observe trends to facilitate more efficient beamforming to user equipments (UEs). To this end, machine learning methods play a vital role to identify useful patterns in data. In this thesis, four machine learning models have been built to categorize whether an UE is moving at a given velocity, or remaining stationary. In particular, binary neural network, multiclass neural network, support vector machine and logistic regression techniques are implemented and analyzed. The uplink sounding reference signal (SRS) channel estimates values are used as input data to the machine learning techniques. The SRS data are generated from a lab simulator at Ericsson AB, Lund. A binary neural network is first built to classify if the UE is moving or remaining stationary. Furthermore, the multiclass neural network is extended to classify movement of the UE at different speeds of 30 km/h or 100 km/h. Further to this, a support vector machine and logistic regression are implemented to compare performance and computational complexity of such approaches, relative to a binary neural network. The obtained results show that the binary neural network has the highest classification accuracy (98%) compared to the support vector machine (95%) and logistic regression (93.8%). For binary classification in this thesis, a larger amount of samples are input into the neural network therefore achieving the highest accuracy. Additionally, the multiclass neural network showed an accuracy of 89.2%. The accuracy of the machine learning algorithms depend on the problem scenario and the size of the dataset.