An Adaptive UKF Method for Geomagnetic/Celestial Autonomous Navigation

Chang Su, Baohua Li, Qiang Zhang, Qi Sun2016 · DOI

Abstract

An adaptive unscented Kalman filtering (UKF) method for geomagnetic/celestial autonomous navigation is presented to solve the problem that the conventional UKF is sensitive to the initial value and declines in accuracy and further diverges due to the system model inaccuracy. Firstly, the residual sum of squares of observations is taken as the criteria to determine the trend of filter divergence. When the residual sum of squares of observations is more than theoretical prediction, an adaptive factor is chosen to adjust the covariance matrixes of state vectors and observed vectors adaptively in order to reduce the impact on filtering caused by initial value and inaccuracy of the system model. Simulation results show that adaptive UKF can restrain filter divergence efficiently, ensure convergence and stability of the filter and improve the precision of the autonomous navigation method using geomagnetism/celestial information under the condition of large initial errors or system model inaccuracy.

Introduction

Autonomous navigation for near-surface aircraft—those flying below 300km—faces unique challenges. Unlike high-altitude satellites, these aircraft cannot rely on infrared horizon sensors due to atmospheric blurring. Furthermore, star sensors alone are often insufficient because the aircraft's motion is not fast enough to produce significant apparent changes in star positions within an inertial frame over short periods.

To address these limitations, researchers have turned to integrated Geomagnetic/Celestial (GNS/CNS) systems. By measuring the angle between a known navigation star and the Earth's magnetic field vector, an aircraft can determine its position. However, the underlying mathematics are highly nonlinear, typically requiring an Unscented Kalman Filter (UKF). Standard UKFs are notoriously sensitive to initial state estimates and model inaccuracies, which can lead to filter divergence and total navigation failure.

Methodology

The proposed Adaptive UKF (AUKF) improves upon the standard UKF by incorporating a real-time "health check" based on the innovation sequence—the difference between actual sensor measurements and model predictions.

The system calculates the included angle ($\alpha$) between the starlight vector and the geomagnetic intensity vector. The starlight vector is obtained from a star sensor, while the geomagnetic vector is derived from an onboard magnetometer and the International General Reference Field (IGRF) model.

Navigation geometry showing the relationship between the aircraft, navigation star, and Earth's geomagnetic vector.

The Adaptive Factor ($\sigma_k$)

The core innovation is the adaptive factor. The algorithm constantly monitors the residual sum of squares of observations. If these residuals exceed a predefined theoretical threshold, indicating that the filter is starting to drift, the adaptive factor is triggered. This factor automatically adjusts the covariance matrices of the state and observed vectors, effectively rebalancing the weight between the mathematical model and real-world sensor data to pull the filter back toward convergence.

Experimental Results

The AUKF was evaluated through numerical simulations of an aircraft in maneuver flight for 3000 seconds. The simulation included 150nT of geomagnetic noise to account for real-world IGRF model inaccuracies. Performance was compared against a standard UKF under two conditions: "Small Initial Offset" and "Large Initial Offset."

Comparison of navigation errors between standard UKF and Adaptive UKF (AUKF) under different initial conditions.

As shown in the results, the AUKF demonstrated superior stability and accuracy. In scenarios with large initial offsets, the standard UKF exhibited significant fluctuations, whereas the AUKF converged rapidly. Specifically, the synthesis velocity error in the large offset scenario dropped from 16.03 m/s to 9.61 m/s, representing a dramatic improvement of approximately 40%. Positioning accuracy also saw significant gains, with latitude error dropping from 0.069° to 0.045° in the large offset case.

Conclusion

The Adaptive UKF effectively addresses the sensitivity of standard nonlinear filters in autonomous navigation. By using the innovation sequence as a diagnostic tool to dynamically adjust filter weights, the system maintains high precision and stability despite model inaccuracies or poor initial state estimates. Future work will focus on testing the algorithm against real-world magnetic anomalies and exploring its integration with other sensors like GPS or LIDAR for enhanced resilience.

Summary updated

Abstract

An adaptive unscented Kalman filtering(UKF) method for geomagnetic/celestial autonomous navigation is presented to solve the problem that the conventional UKF is sensitive to the initial value and declines in accuracy and further diverges due to the system model inaccuracy. Firstly, the residual sum of squares of observations is taken as the criteria to determine the trend of filter divergence. When the residual sum of squares of observations is more than theoretical prediction, an adaptive factor is chosen to adjust the con-variance matrixes of state vectors and observed vectors adaptively in order to reduce the impact on filtering caused by initial value and inaccuracy of the system model. Simulation results show that adaptive UKF can restrain filter divergence efficiently, ensure convergence and stability of the filter and improve the precision of the autonomous navigation method using geomagnetism/celestial information under the condition of large initial errors or system model inaccuracy.
BibTeX
@misc{su2016adaptive,
  author = {Chang Su and Baohua Li and Qiang Zhang and Qi Sun},
  title = {{An Adaptive UKF Method for Geomagnetic/Celestial Autonomous Navigation}},
  year = {2016},
  doi = {10.2991/amsm-16.2016.59},
}
Figure