Simulation reveals limits of GNSS fault detection in multirotor UAVs

Summary

A simulation study evaluates five causal monitoring methods (GNSS–acceleration inconsistency, Kalman NIS, barometer consistency, and fixed and covariance-normalized physics scores) for detecting slow and physically plausible GNSS trajectory faults in multi-rotor UAVs. The study reveals that detection rates collapse on out-of-sample data and that adaptation-based slow lateral drifts are nearly undetectable by any of the tested methods.

Why it matters

As multi-rotor platforms become increasingly reliant on GNSS for autonomy, understanding the hard limits of physics-based fault detection under realistic sensor constraints is critical for safety. The article quantitatively identifies a fundamental trade-off: tightening calibration to suppress false alarms also eliminates the few true detections. This leaves operators with no free lunch in monitoring design.

Highlights

• In a panel of 1,800 missions, the covariance physics method captured 82.9% of faults; however, this rate dropped to 41.5% in a 600-mission out-of-sample set, indicating a significant generalization gap.
• In 400 adaptation-based slow lateral missions, Kalman NIS flagged only 9 faults, while the other four monitors detected none, highlighting a near-blind spot.
• Nominal physics calibration produced a 40–41.6% false alarm rate under clean model mismatch. Broader calibration eliminated these alarms but also wiped out the observed detections.

Critical take

The study is entirely simulation-based; RF-level attacks, receiver manipulation, or flight hardware loops are not included. Therefore, conclusions about 'limits' are limited to the fidelity of the chosen models and the specific slow lateral fault class. While the first author's high-school-level institutional affiliation is not a disqualifier on its own, the lack of real-world validation or comparison with commercial integrity monitors (e.g., RAIM derivatives) limits the practical weight of the headline 82.9% figure.