Machine learning LTE path loss estimation from end-device measurements

Abstract

Predicting accurate path loss is essential in the design and planning phase of mobile networks since the network performance directly benefits from accurate path loss estimation. This dissertation is concerned with predicting path loss values for LTE mobile networks using end-user device measurements and machine learning. The study's objectives include developing a collaborative data-capturing platform for LTE networks using user equipment (UE), developing a localisation method to estimate the location of the LTE eNodeB (base station), and using machine learning to estimate path loss using the captured data. The LTE eNodeB localisation is based on using a random forest regressor to estimate the distance to the eNodeB and combine it with multilateration to yield the estimated coordinates of the eNodeB. To model the path loss, a random forest regressor combined with different path loss attributes identified in the literature was used. The tower localisation results obtained in the study clearly show that the timing advance parameter is essential for LTE eNodeB localisation purposes and, when combined with a random forest regressor and multilateration, yields sub-100 metre localisation errors. The path loss modelling results indicate that the developed machine learning regressor model can estimate path loss for unseen data equivalent to a fitted log-distance model when given groundtruth tower locations. It is also evident from the study that it is possible to create machine learning path loss models using localised eNodeB data, paving the way for creating truly diverse path loss models. In conclusion, it is possible to use UE to capture LTE parameter data collaboratively, localise eNodeBs accurately from the data, and predict accurate path loss using machine learning models from the captured data.