Abstract
Federated Kalman Filter (FKF) is the most widely used distributed data fusion algorithm for UAV navigation. However, FKF typically requires local systems to share the same system model, a condition often difficult to meet in practice. This paper proposes a Random Vector Space (RVS) approach to balance estimation accuracy and computational load. By treating state predictions and estimations as RVS bases, the optimal state estimation is derived through a linear combination of these bases. Replacing the global time-updating module in FKF with the RVS approach yields higher accuracy without increasing calculation time. Simulation results demonstrate improvements in position and velocity estimation accuracy across three axes, with velocity accuracy increasing by up to 19.9%.
Introduction
Unmanned Aerial Vehicles (UAVs) rely on integrated navigation systems to maintain precise Guidance, Navigation, and Control (GNC). These systems typically fuse data from Inertial Navigation Systems (INS), Global Positioning Systems (GPS), and Synthetic Aperture Radar (SAR). While each sensor has strengths, they also possess critical weaknesses: INS suffers from cumulative error over time, GPS is susceptible to electromagnetic interference or jamming, and SAR requires distinct topographic features for effective matching.
To mitigate these issues, the Federated Kalman Filter (FKF) is the industry standard. However, standard FKF implementations are limited by the requirement for homogeneous mathematical models across local filters and the use of simplistic scalar information-sharing factors, which can lead to sub-optimal data fusion or excessive computational loads if the system requires frequent resets.
Methodology: The Random Vector Space Approach
The proposed method enhances the traditional FKF by replacing the global fusion step with a Random Vector Space (RVS) geometric approach.
In this architecture, the INS serves as the core reference system. Local filters process data from GPS and SAR. Instead of using scalar weights to merge local estimates into a global state, the RVS approach treats state predictions and estimations from all modules as orthogonal basis vectors in a random vector space. The "true" state is then represented as a linear combination of these bases. By applying a Lagrange Function to minimize the trace of the covariance matrix, the system identifies the most stable and accurate combination of these vector bases. This geometric interpretation allows the filter to capture multi-dimensional relationships between position, velocity, and attitude more effectively than scalar-based methods.
Simulation and Results
The algorithm was evaluated using a 300-second mathematical simulation (from 1700s to 2000s) based on realistic UAV flight data. Sensor noise models included specific biases for accelerometers (50µg) and gyroscopes (1°/h) to reflect real-world hardware constraints.
The RVS-FKF approach demonstrated significant performance gains over the standard FKF:
- Velocity Accuracy: Improvements of 19.9% (North), 13.3% (Up), and 17.6% (East) were observed.
- Position Accuracy: Improvements were more modest but consistent, at 1.59% (North), 1.53% (Up), and 1.29% (East).
- Computational Efficiency: The processing time remained virtually identical, with the RVS-FKF requiring 3.56 seconds compared to 3.52 seconds for the standard FKF. This negligible 0.04s difference makes the approach highly suitable for embedded UAV hardware.
The substantial improvement in velocity compared to position is attributed to the fact that position is often an indirect derivation in these systems; thus, the refined velocity data provides the most immediate evidence of the RVS fusion's efficacy.
Conclusion
The Random Vector Space approach offers a computationally efficient way to enhance the accuracy of integrated navigation systems for UAVs. By moving from scalar information sharing to a vector-based geometric fusion, the system achieves up to a 20% improvement in velocity estimation without requiring additional hardware resources. Future work will explore the application of RVS to more heterogeneous sensor suites and its robustness in high-maneuverability flight scenarios where sensor non-linearity may be more pronounced.
Summary updated