Fast and Noise-Resilient Magnetic Field Mapping on a Low-Cost UAV Using Gaussian Process Regression.

Abstract

This study presents a comprehensive approach to mapping local magnetic field anomalies with robustness to magnetic noise from an unmanned aerial vehicle (UAV). The UAV collects magnetic field measurements, which are used to generate a local magnetic field map through Gaussian process regression (GPR). The research identifies two categories of magnetic noise originating from the UAV's electronics, adversely affecting map precision. First, this paper delineates a zero-mean noise arising from high-frequency motor commands issued by the UAV's flight controller. To mitigate this noise, the study proposes adjusting a specific gain in the vehicle's PID controller. Next, our research reveals that the UAV generates a time-varying magnetic bias that fluctuates throughout experimental trials. To address this issue, a novel compromise mapping technique is introduced, enabling the map to learn these time-varying biases with data collected from multiple flights. The compromise map circumvents excessive computational demands without sacrificing mapping accuracy by constraining the number of prediction points used for regression. A comparative analysis of the magnetic field maps' accuracy and the spatial density of observations employed in map construction is then conducted. This examination serves as a guideline for best practices when designing trajectories for local magnetic field mapping. Furthermore, the study presents a novel consistency metric intended to determine whether predictions from a GPR magnetic field map should be retained or discarded during state estimation. Empirical evidence from over 120 flight tests substantiates the efficacy of the proposed methodologies. The data are made publicly accessible to facilitate future research endeavors.