A Fault-Tolerant Multisensor Fusion Architecture for Robust Attitude Estimation in UAVs

Abstract

Accurate and robust attitude estimation is essential for the control and navigation of unmanned aerial vehicles (UAVs), especially under inertial measurement unit (IMU) degradations. This work presents a fault-tolerant multisensor fusion architecture for attitude and angular-rate estimation using three redundant IMU-driven estimator branches and an analytical UAV dynamic reference model. The framework combines model-based analytical referencing, weighted voting, Kalman refinement, fault diagnosis, online model calibration, and gated residual-covariance adaptation. Fusion weights are computed componentwise from pairwise Mahalanobis distances between sensor-based estimation residuals relative to the analytical model, mapped into continuous statistical consistency scores and modulated by Allan-variance-based reliability factors. The voter output and calibrated model prediction are then fused through a discrete-time Kalman filter to obtain a dynamically consistent and noise-reduced 6-D estimate. Residual consistency analysis enables fault-tolerant weighting, fault detection, branch isolation, and recovery actions. The architecture is evaluated under representative hard and soft IMU fault scenarios and benchmarked against a weighted-voting fusion architecture, an adaptive federated Kalman filter, and an M-estimator-based robust Kalman filter. Results demonstrate improved estimation accuracy, noise attenuation, fault-detection capability, and computational feasibility on the target flight controller.