Geometric Information for Wireless Communications and Localization

Abstract

This thesis presents two distinct yet complementary approaches to advance wireless communication systems. First, leveraging geometric information, I address the challenge of predicting the outdoor signal strength of cellular networks, which is critical for network planning and predicting the coverage maps of cellular networks. Traditional methods relying on ray tracing or stochastic models often encounter limitations in accuracy and data availability. To overcome these challenges, I propose a practical received signal strength prediction system that uses trajectory information and satellite maps surrounding base stations. The map-based deep neural network architecture, which uses the split learning framework and is enhanced with the deep vision transformer, achieves high prediction accuracy while protecting user privacy. Evaluation results using real-world datasets demonstrate significant performance of the proposed approach over conventional models in both centralized and distributed scenarios, with reduced computational overhead. In the second work, I address indoor localization using a zero-shot learning framework, reducing reliance on real-world measurements in novel communication environments. The approach uses a scalable graph neural network for coarse localization, supplemented by floor-plan-aided deep neural networks for improved accuracy. A synthetic data generator enhances generalization ability in scenarios lacking real-world samples. Experimental results show a significantly improved accuracy compared to existing literature, demonstrating the effectiveness of the approach in real-time location estimation of mobile devices.