Vehicular Outdoor Localization Using CellularNetwork Signals and Machine Learning

Accurate vehicular localization is an essential parameter for enabling intelligent fleet management and predictive maintenance, especially in connected Electric Vehicles (EVs). Global Navigation Satellite Systems, such as the Global Positioning System (GPS), often suffer from signal blockage, high energy use, and hardware constraints in dense urban areas. This paper proposed an advanced cellular network-based localization approach to support EV health monitoring by addressing two key challenges. The work addressed two linked challenges. The first challenge concerned incomplete Long Term Evolution (LTE) signal measurements along the vehicle trajectories. These data gaps occurred due to handovers between cell towers, signal obstructions, and vehicle motion. The second challenge involved the regression of these cellular features to geographic coordinates with meter-level error. This study evaluates imputation methods for missing values and introduces two methods tailored to time series of cellular signals, named Blockwise Endpoint-Averaging (BEA) and Blockwise Endpoint-Propagation (BEP). It then applied feature selection methods. These methods reduces the feature space from nineteen to seven cellular and geometric features. Several regression models then mapped the selected features to latitude and longitude, with a stacking ensemble that combines the base models with the highest validation scores. Experiments on an urban drive dataset showed that BEA imputation, feature selection, and stacking reduced the mean localization error from about 46.87 metres in the baseline to about 2.69 metres.