LOS/Multipath/NLOS Classifiers using Machine learning and Raytracing. A preliminary study to identify and address the Mulitpath error

Abstract

This thesis explores the application of machine learning algorithms to address one of the challenges in localization in GNSS called the Multipath errors. The approach involves data collection via a receiver mounted on an excavator which is placed in front of the building, which acts as one of the sources for mulitpath error. In order to perform supervised machine learning, wireless communication tool box within Matlab is used for Raytracing simulation to label the data. Drawing inspiration from existing literature, we introduce a novel feature, the ’differ ence in range acceleration between code and carrier signals,’ which exhibits promis ing distribution and metrics when analyzed with respect to ray-tracing results. The support vector machine (SVM) achieves an average class-wise recall and preci sion of approximately 75% on the recorded measurments. Additionally, we explore the use of an Autoencoder with a 1D-CNN layer to extract new features aimed at enhancing classification performance. By conducting three different simulations with varying data sorting methods, we demonstrate how sorting the data and the machine learning algorithm can influence the learned features which in turn impacts classification performance. Lastly, in order to take advantage of the insights gained from different sorting meth ods, we transform the problem from a Multiclass to Multi-label-Multi-output clas sification problem, wherein we utilize a deep neural network architecture with GRU units to classify the signals. Although complicated in terms of data restructuring and handling, the network demonstrated robust performance, achieving more than 85% average class-wise recall and precision and exceeding 96% for signals labeled LOS.