UC Davis

Radio coverage estimation plays an important role in network planning and resource management for modern Internet-of-Things (IoT) and cellular systems. To capture the distribution of radio spectrum, a vital tool is “radio map”, which describes the geometrical signal power spectrum density (PSD). Usually, a dense radio map is estimated from sparse observations collected by deployed sensors or user devices. Thus, a practical problem is to estimate fine-resolution radio maps from sparse radio strength measurements. However, most existing methods focus only on data distributions or propagation model, neglecting their correlations, which leads to poor performance in realistic applications. To leverage the radio propagation information in data-driven learning machines, this dissertation investigates the development of Physics-Inspired Machine Learning (PIML) for radio map estimation by integrating data statistics and physical principle.Specifically, this dissertation proposes two innovative physics-inspired, learning-based radio map estimation (RME) methods as solutions to the high cost of manual data collection. The first method introduces a novel feature, named by radio depth map (RDM), into the conditional generative adversarial network (cGAN) model, which enriches the learning process using radio propagation and shadowing models, demonstrating enhanced performance and robustness over the traditional learning-based and model-based methods. The second method incorporates physics rewards into the learning workflow, taking initial steps to integrate domain-specific knowledge into the learning process. A case study on the physics-rewarded RME indicates the validity of this approach and showcases its potential for future optimization. Our experimental results demonstrate the power of the proposed RDM in capturing the radio PSD distribution, as well as the efficacy of the proposed PIML methods for radio map reconstruction.