Cell Tower Localization using crowdsourced measurments

Abstract: This thesis explores the application of a neural network approach to cell tower localization using crowdsourced measurements. The deployment of cell tower infrastructure has been increasing exponentially in recent times as it is a crucial element of mobile communications. Location information is key to the quality of 4G LTE and 5G wireless service, establishing accurate coverage maps and different connectivity studies. Mobile carriers do not usually disclose the location of their cell towers due to security concerns, regulatory requirements, or market competition. In addition, open-source datasets on cell tower localization available online are often incomplete, inaccurate, or non-existent. Crowdsourcing enables the collection of large amounts of signal measurements from several mobile devices. By labeling these measurements with ground truth locations of base stations, we can address this challenge, employing a machine learning framework to predict the geographical locations of cell towers. The methodology followed in this project involves data preprocessing and feature engineering of a crowdsourced dataset along with the implementation and tuning of a multi-layer perceptron (MLP) neural network model. The cell tower approximations obtained with this method excelled other state-of-the-art localization algorithms and provide a better estimation of telecommunication infrastructure deployments than open-source datasets. Overall, this thesis discusses the feasibility of employing a neural network model for predicting cell tower locations, while addressing some limitations and possible improvements for the localization problem.