Limits of Physics-Based GNSS Fault Detection in Multirotor UAV Simulation

Abstract

A false Global Navigation Satellite System (GNSS) trajectory does not always look physically impossible. A slow horizontal offset can preserve smooth position, velocity, acceleration, and jerk, making it difficult to reject with a limited onboard sensor set. This simulation study compares five causal monitors for a multirotor UAV: GNSS–acceleration disagreement, Kalman normalized innovation squared (NIS), GNSS–barometer altitude consistency, a fixed physics score, and a covariance-normalized physics score. Thresholds were fitted on 300 nominal and 300 uncertainty-sampled clean missions, then frozen before 7,480 evaluation missions. On a predefined 1,800-mission panel, covariance physics detected 82.9% and fixed physics 78.3%; on a separately generated 600-mission holdout, their rates fell to 41.5% and 34.3%, while Kalman NIS reached 42.0%. In 400 adaptive slow-horizontal missions, Kalman NIS detected 9 and the other four monitors detected none. Nominal physics calibration also produced 40.0–41.6% false-alarm rates under clean model mismatch. Broader calibration removed those false alarms but also removed the observed physics detections. The results show that detectability depends on the independent measurements available and on model uncertainty. No radio-frequency attack, receiver manipulation, or flight hardware is modeled.