Abstract
The advent of 5G technology has facilitated the adoption of private cellular networks in industrial settings. Ensuring reliable coverage while maintaining certain requirements at its boundaries is crucial for successful deployment yet challenging without extensive measurements. In this article, we propose the leveraging of unmanned aerial vehicles (UAVs) and Gaussian processes (GPs) to reduce the complexity of this task. Physics-informed mean functions, including a detailed ray-tracing simulation, are integrated into the GP models to enhance the extrapolation performance of the GP prediction. As a central element of the GP prediction, a quantitative evaluation of different mean functions is conducted. The most promising candidates are then integrated into an informative path-planning algorithm tasked with performing an efficient UAV-based cellular network mapping. The algorithm combines the physics-informed GP models with Bayesian optimization and is developed and tested in a hardware-in-the-loop simulation. The quantitative evaluation of the mean functions and the informative path-planning simulation are based on real-world measurements of the 5G reference signal received power (RSRP) in a cellular 5G-SA campus network at the Port of Lübeck, Germany. These measurements serve as ground truth for both evaluations. The evaluation results demonstrate that using an appropriate mean function can result in an enhanced prediction accuracy of the GP model and provide a suitable basis for informative path planning. The subsequent informative path-planning simulation experiments highlight these findings. For a fixed maximum travel distance, a path is iteratively computed, reducing the flight distance by up to 98% while maintaining an average root-mean-square error of less than 6 dBm when compared to the measurement trials.