A machine learning approach for localization in cellular environments

Abstract

A machine learning approach is developed for localization based on received signal strength (RSS) from cellular towers. The proposed approach only assumes knowledge of RSS fingerprints of the environment, and does not require knowledge of the cellular base transceiver station (BTS) locations, nor uses any RSS mathematical model. The proposed localization scheme integrates a weighted K-nearest neighbor (WKNN) and a multilayer neural network. The integration takes advantage of the robust clustering ability of WKNN and implements a neural network that could estimate the position within each cluster. Experimental results are presented to demonstrate the proposed approach in two urban environments and one rural environment, achieving a mean distance localization error of 5.9 m and 5.1 m in the urban environments and 8.7 m in the rural environment. This constitutes an improvement of 41%, 45%, and 16%, respectively, over the WKNN-only algorithm.