• UJJWAL MANNA

A UAV can continue receiving sensor measurements even when one of its sensors has started producing misleading information. In that situation, the problem is not simply that data are missing. The estimator has to decide whether a measurement is still useful, whether its influence should be reduced, or whether it should be ignored altogether. In this work, we develop a sensor-confidence framework that uses recent GNSS, inertial, visual, and barometric observations to estimate the likelihood that each sensor remains within a predefined error tolerance. The confidence estimate is then used to adjust the measurement covariance of a state estimator, while persistent low confidence can lead to measurement rejection. Inertial degradation is handled separately because the IMU also contributes to the prediction step. We compare three approaches: fixed sensor fusion, a conventional innovation-based adaptive filter, and a lightweight learned confidence model. Five conditions are considered: healthy motion, progressive GNSS degradation, abrupt GNSS faults, GNSS outage, and concurrent GNSS–vision faults. A software-based pilot experiment was carried out on 30 held-out trajectories, with six test trajectories for each scenario. The learned method produced an overall 3-D position RMSE of 1.247 m, compared with 1.223 m for fixed fusion and 1.618 m for the conventional adaptive method. It performed slightly better than fixed fusion under progressive and abrupt GNSS faults, but the overall result did not improve, and performance was weaker when GNSS and vision failed together. These results suggest that temporal sensor confidence can be useful, but the current implementation does not justify claiming that it is generally better than simpler fusion methods. Expanded verification was then run as a separate reproducibility study on 240 unseen trajectories spanning eight conditions. The training distribution contained the five original scenario families, while intermittent GNSS spikes, visual degradation with dropout, and stale GNSS measurements were held out as new fault forms. Across the 240-trajectory suite, the learned-confidence estimator reduced mean 3-D RMSE from 0.750 m with fixed fusion to 0.659 m, a 12.2% reduction; on the three unseen fault forms the reduction was 22.5%. The conventional innovation-adaptive baseline remained stronger in several scenarios, so the contribution is not presented as a universal replacement for adaptive filtering. The stronger result is sensor-health evidence: the calibrated temporal model produced a Brier score of 0.095 and an expected calibration error of 0.098, with no sustained false-alarm trajectory among 30 healthy expanded tests. For 61 cases in which the fixed-fusion reference crossed a sustained 2 m error limit, the GNSS confidence alarm occurred first in all 61, with a median simulated lead of 1.5 s.

UAV navigation; GNSS degradation; sensor confidence; adaptive estimation; temporal learning; fault tolerance; uncertainty calibration

"Adaptive Sensor Confidence Learning for Resilient UAV Navigation Under GNSS Degradation and Multiple Sensor Failures", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.13, Issue 9, page no.b268-b278, September-2026, Available :http://www.jetir.org/papers/JETIR2609132.pdf

"Adaptive Sensor Confidence Learning for Resilient UAV Navigation Under GNSS Degradation and Multiple Sensor Failures", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.13, Issue 9, page no. ppb268-b278, September-2026, Available at : http://www.jetir.org/papers/JETIR2609132.pdf