Construction of Mobile Performance Maps using Clustered Crowdsourced Measurements

Abstract

The goal is to create forecasts of network performance maps based on collected measurements from different data sources. Even though we focus on Reference Signal Receive Power (RSRP) as the measurement metric, the underlying model must be extendable to other parameters, such as data-rate or service quality. Additionally,we aim to exploit the underlying spatial properties of the data.We employ a Gaussian Process Regression (GPR). The superiority of the GPR against other regressionmodels is its ability to provide a distribution of the prediction value rather than just a single value. This uncertainty can be exploited to determine the optimal location for the next measurement, such that the prediction error of the entire map is minimized. The task of collecting measurements can be outsourced to end-userdevices, thus saving the network operators valuable time. We employ the clustering of the data, which offers a possibility of computational time reduction for the GPR prediction and the measurement noise averaging effect. We design three different measurement distribution scenarios and analyze the MSE degradation over different measurement point densities when clustering is applied. We investigate different cluster identification methods, applicable in the real-data measurements, where the nature of clusters is unknown in advance. By comparing their influence on the MSE of the GPR prediction, we conclude that the K-Means method is best suited for our purposes. Finally, we apply the K-Means cluster identification method in two experiments, using measurements collected by the XY-positioning table and a drone.We show that we can considerably reduce the number of GPR training points by clustering measurements, without significant loss of the prediction quality. The experimental results also confirm our simulation results on the number of requiredclusters for an accurate GPR prediction.