TL;DR
This paper introduces a novel hybrid adaptive Cubature Kalman Filter (CKF) that effectively handles systems with unknown measurement noise variance. By fusing estimates from multiple noise variance estimators and adaptively adjusting the filter, the proposed method significantly improves filtering accuracy and stability compared to traditional CKF, especially in challenging nonlinear scenarios.
What Problem Does This Paper Address?
Accurate state estimation in nonlinear systems is crucial for many applications, from navigation to target tracking. Traditional Kalman filters, including the popular Cubature Kalman Filter (CKF), rely on accurate knowledge of both system dynamics and noise statistics. However, in real-world scenarios, the variance of measurement noise is often unknown or only imprecisely known. This uncertainty can lead to significant degradation in filtering performance, including divergence. Existing adaptive filtering techniques often struggle with the complexity of nonlinear systems and may not provide robust estimates of noise variance. This paper aims to bridge this gap by developing an adaptive filtering approach that can accurately estimate and adapt to unknown measurement noise variance in nonlinear systems.
Methods
The core of this paper is the development of a Hybrid Adaptive Cubature Kalman Filter (HASCKF). The approach involves several key steps:
- Weighted Fusion Estimation of Measurement Noise Variance: Instead of relying on a single estimator, the paper proposes a novel method to fuse estimates from multiple noise variance estimators (specifically, MAP and Variational Bayesian methods). This weighted fusion approach aims to leverage the strengths of different estimators and provide a more robust and accurate estimate of the measurement noise variance.
- Adaptive Filtering Structure: The HASCKF algorithm integrates this fused noise variance estimate into the standard CKF framework. Specifically, it uses a fading factor adjustment (FFA) technique to adapt the measurement update part of the CKF. This allows the filter to dynamically adjust its reliance on measurements based on the estimated noise variance.
- Information Filter Variant (HASCIF): For improved numerical performance and potentially simpler implementation, the authors also present an information filter version of their adaptive algorithm, the Hybrid Adaptive Square-Root Cubature Information Filter (HASCIF).
- Theoretical Analysis: The paper provides theoretical analysis of the HASCKF's stability and filtering accuracy, drawing upon existing theorems for CKF convergence and the Cramer-Rao Lower Bound (CRLB) for performance evaluation.
Datasets
Not discussed in the paper. The paper uses simulated data for its examples.
Key Insights and Ideas
- Fusion is Better Than Single Estimation: Combining estimates from multiple noise variance estimators (MAP and VB) through a weighted fusion approach yields a more accurate and stable estimate of the measurement noise variance than any single estimator alone. This is particularly true for time-varying noise.
- Adaptive Noise Handling is Crucial for Nonlinear Systems: The performance of nonlinear filters is highly sensitive to the accuracy of noise variance estimates. Adapting the filter based on estimated noise variance is essential for maintaining accuracy and preventing divergence, especially when the true noise variance is unknown or changes over time.
- Hybrid Approach Balances Accuracy and Stability: The proposed HASCKF algorithm successfully combines the strengths of adaptive noise estimation with the robust numerical properties of square-root filtering (specifically, square-root CKF), leading to improved stability and accuracy.
- Information Filter Offers Numerical Advantages: The HASCIF variant demonstrates that an information filter formulation can achieve similar performance with potential benefits in numerical stability and implementation simplicity.
Key Results
The paper presents simulation results for two examples:
- Example 1 (Noise Estimator Comparison): The weighted fusion noise estimator (WF-NE) consistently outperformed individual MAP-NE and VB-NE estimators in estimating both constant and time-varying measurement noise variances. WF-NE showed smaller mean absolute errors and better tracking of the true variance, especially in challenging situations. While WF-NE had a slightly higher CPU time cost, it was deemed acceptable.
- Example 2 (HASCKF vs. SCKF): The HASCKF algorithm demonstrated significantly better performance than the standard SCKF when dealing with inaccurate prior measurement noise variance. The standard SCKF showed large deviations in state estimation after a certain time, whereas HASCKF maintained accurate tracking. The mean absolute errors for HASCKF were orders of magnitude smaller across all state variables (position and velocity) compared to SCKF.
Limitations and Future Work
- Limitations:
- The paper acknowledges that while WF-NE generally performs better, there might be specific time intervals or noise realizations where individual estimators (MAP-NE or VB-NE) might momentarily outperform the fusion method due to the inherent randomness of noise.
- The theoretical analysis of stability and convergence relies on certain assumptions about the system and noise characteristics, which might not always hold in highly complex or extreme scenarios.
- Future Work:
- The authors suggest further research into designing fusion methods within centralized and distributed fusion frameworks.
- Investigating the functional equivalence among various fusion algorithms is also proposed.
- Exploring hybrid adaptive filtering in more complex nonlinear centralized fusion frameworks and distributed fusion frameworks.
Summary updated