Abstract
In practical SINS/GPS integrated navigation, severe maneuvers and GPS measurement anomalies often lead to unknown process noise statistics and non-Gaussian measurement noise. These conditions degrade the performance of traditional Cubature Kalman Filters (CKF). This paper presents a Robust Cubature Kalman Filter (RCKF) based on adaptive information entropy theory. By integrating the Maximum Correntropy Criterion (MCC) with a multi-channel adaptive fading strategy, the proposed filter effectively suppresses both process uncertainty and non-Gaussian outliers. Experimental results from simulations and real-world car-mounted tests demonstrate significant improvements in estimation accuracy and robustness over state-of-the-art methods.
Introduction
The integration of Strapdown Inertial Navigation Systems (SINS) and Global Positioning Systems (GPS) is a cornerstone of modern navigation. While SINS offers high-frequency, autonomous data, it suffers from time-dependent drift. GPS provides absolute positioning to bound this drift but is susceptible to signal blockages and multipath interference in urban environments. The Cubature Kalman Filter (CKF) is widely used for this nonlinear fusion task, yet its reliance on Gaussian noise assumptions makes it fragile in real-world "dirty" environments characterized by heavy-tailed noise and unpredictable vehicle maneuvers.
Methodology: The RCKF Framework
The proposed Robust Cubature Kalman Filter addresses the limitations of standard CKF through two primary mechanisms:
1. Maximum Correntropy Criterion (MCC)
Traditional filters minimize the Mean Square Error (MSE), which only considers second-order statistics (mean and variance). In contrast, the RCKF utilizes a cost function based on Maximum Correntropy. Correntropy is a localized similarity measure that captures higher-order moments of the error distribution. By employing a Gaussian kernel, the filter becomes inherently insensitive to large outliers (non-Gaussian noise), effectively "weighting down" abnormal GPS measurements that would otherwise skew the navigation solution.
2. Multi-Channel Adaptive Fading Strategy
Process noise statistics often change during high-dynamic maneuvers, leading to model mismatch. While standard adaptive filters apply a single global fading factor to the entire covariance matrix, the RCKF introduces a multi-channel approach. This strategy applies distinct scaling factors to different components of the 21-dimension state vector (e.g., separating velocity, position, and attitude errors). This granular adaptation prevents the filter from over-correcting stable states while aggressively addressing those affected by maneuvers, providing surgical precision in tactical navigation.
3. Numerical Stability and Efficiency
To ensure the algorithm is suitable for real-time applications, the authors derived a simplified Kalman gain and implemented an innovation sequence detection scheme. These improvements prevent numerical instabilities, such as matrix singularity, which often plague robust nonlinear filters during extreme measurement anomalies.
Experimental Validation
The RCKF was validated through two distinct scenarios:
- Simulation: A 500-second trajectory was subjected to "mixed-Gaussian" noise (a combination of 90% standard noise and 10% high-variance outliers). The RCKF maintained high precision where standard filters diverged.
- Field Test: A car-mounted experiment was conducted in Taiyuan, China, using a tactical-grade IMU (LCI-1) and a Propak satellite receiver. The route included "dump roads" with significant occlusions from trees and buildings, providing a realistic testbed for the non-Gaussian interference the filter was designed to handle.
Key Results and Discussion
The RCKF demonstrated a marked improvement in performance:
- Accuracy: In field tests, the RCKF reduced the Root Mean Square Error (RMSE) across velocity, position, and attitude by at least 36% compared to existing robust filters like the H-infinity or Huber’s M filter.
- Computational Load: The per-step execution time was measured at approximately 0.416ms, making it nearly twice as fast as more complex Interactive Multiple Model (IMM) filters and well within the requirements for high-frequency real-time hardware.
- Robustness: The filter remained stable during sudden maneuvers where the standard CKF experienced significant error spikes, demonstrating superior "inherited advantages" from the adaptive correntropy framework.
Conclusion
The Robust Cubature Kalman Filter provides a powerful solution for SINS/GPS integration in challenging environments. By combining the outlier-rejection capabilities of Maximum Correntropy with the precision of multi-channel adaptive fading, it achieves superior accuracy and stability. While the filter is sensitive to the choice of kernel bandwidth (σ), future research will focus on generalized kernels and evolutionary optimization to further enhance the filter's autonomy across varying environmental conditions.
Summary updated