Physics-Informed Probabilistic Learning of Low-Altitude Urban Wind from Onboard Motion Data

Minghao Chen

Keywords: Physics-informed machine learning; automation for smart cities;

TL;DR: We propose an effective physics-informed probabilistic learning approach to estimate complex urban low-altitude winds with uncertainty quantification using motion data from unmanned aerial vehicles.

Abstract: Urban wind fields at low altitude are highly variable and safety-critical for urban air mobility and other operations. This paper presents a physics-informed probabilistic machine learning approach to estimate 3D local winds: a quadrotor dynamics prior coupled with a learned discrepancy, inferred from onboard IMU data only. The latent wind vector and its uncertainty can be estimated in near-real time, enabling risk-aware planning in cities. Unlike existing methods that rely on dedicated sensors or complex models with high computational demands, our approach uses only motion data from standard onboard IMU sensors. Our results demonstrate that the method can estimate wind vectors with a root mean square error (RMSE) of less than 12\% (of speed range) under various simulated realistic wind conditions in both hovering and cruising cases, while providing uncertainty quantification of the estimates. The proposed framework offers an easy-to-implement low-altitude onboard wind estimator to support drone operations in urban environments.

Submission Number: 55