TL;DR
This paper introduces an improved Multiple-Model Adaptive Estimation (MMAE) method for integrated navigation systems facing time-varying noise. By combining the Sage-Husa adaptive filter with an enhanced hypothesis testing algorithm and incorporating a forgetting factor as an adaptive parameter, the method significantly boosts robustness and accuracy in challenging noise environments.
What Problem Does This Paper Address?
Accurate navigation systems, particularly those integrating Inertial Navigation Systems (INS) with other sensors like GPS, rely heavily on Kalman filtering techniques. However, these filters are sensitive to noise, and when noise characteristics change over time (time-varying noise), standard filters can degrade in performance or even diverge. Traditional adaptive Kalman filters can struggle with complex noise, and Multiple-Model Adaptive Estimation (MMAE), while robust, faces issues like undesirable model competition and the need for fixed, pre-defined model parameters. This paper tackles the challenge of improving the estimation accuracy and robustness of integrated navigation systems in dynamic, unpredictable noise environments by addressing these limitations.
Methods
The core of this work is an Improved Multiple-Model Adaptive Estimation (MMAE) method. Here's a breakdown of the key components:
- Sage-Husa Adaptive Unscented Kalman Filter (UKF): Instead of using standard Kalman filters within the MMAE framework, the authors replace them with Sage-Husa adaptive UKFs. This allows the filter to estimate noise parameters (specifically, measurement noise covariance) online, adapting to changing noise conditions. The UKF itself is chosen for its superior handling of nonlinear systems compared to the Extended Kalman Filter (EKF).
- Adaptive Forgetting Factor: The paper identifies that a fixed forgetting factor can limit adaptability. They incorporate the forgetting factor as an unknown parameter within the multi-model structure, allowing the algorithm to adaptively adjust it based on system states. This aims to balance adaptability and stability.
- Improved Hypothesis Testing Algorithm: Classical MMAE suffers from "undesirable model competition," where less optimal models can still receive significant weight. The authors propose an improved hypothesis testing algorithm to mitigate this. This enhancement focuses on how model probabilities are assigned, aiming to reduce the influence of suboptimal models and improve the algorithm's parameter identification capabilities.
- Variable-Parameter Multi-Model Structure: By making the forgetting factor and process noise matrix unknown parameters, the proposed method creates a variable-parameter multi-model structure. This reduces reliance on initial noise parameter selection and broadens the coverage of potential system states.
Datasets
The paper utilizes INS/GPS vehicle navigation data collected from a vehicle navigation platform. This data was used to simulate different noise scenarios by artificially adding position noise to the measurement data. The authors do not introduce a new benchmark dataset but rather use real-world collected data and simulate challenging noise conditions to evaluate their algorithm.
Key Insights and Ideas
- Adaptive Forgetting Factor is Crucial: The study highlights that a fixed forgetting factor is suboptimal. Adapting it dynamically based on system states is key to balancing the filter's responsiveness to noise changes with its overall stability.
- Model Competition is a Real Problem: The paper clearly demonstrates how undesirable model competition in classical MMAE can slow down convergence and reduce accuracy, especially in time-varying noise. Their improved hypothesis testing directly addresses this.
- Sage-Husa for Noise Adaptation: Integrating Sage-Husa adaptive filtering within MMAE is a powerful way to handle time-varying noise, as it allows for real-time estimation of noise parameters, a critical factor for accurate state estimation.
- Combined Approach Yields Superiority: The synergy between adaptive noise estimation (Sage-Husa), adaptive parameter tuning (forgetting factor), and improved model selection (hypothesis testing) leads to a more robust and accurate navigation system.
Key Results
The proposed method demonstrated superior performance compared to standard UKF, Sage-Husa UKF, classical MMAE, and an MMAE with an exponentially decaying penalty value function.
- Reduced Position Errors: Figures 5 and 6 show that the proposed method consistently produced lower east and north position errors across different time intervals.
- Lower Mean Squared Error (MSE): Table 4 quantifies this improvement, showing significantly lower MSE values for the proposed method across all tested intervals for both east and north position errors. For instance, in the 1800s-2300s interval, the proposed method achieved an MSE of 6.3463m for east position error, compared to 13.8936m for standard UKF.
- Faster Optimal Model Convergence: Figure 4 illustrates that the optimal model in the proposed method converges much faster than in the classical MMAE, indicating better and quicker adaptation to system changes.
Limitations and Future Work
- Limitations:
- While the method shows significant improvement, the paper acknowledges that a single forgetting factor and process noise parameter matrix in Sage-Husa alone cannot guarantee optimal estimation at every moment, which limits its performance.
- The paper focuses on improving filter stability and adaptability in time-varying noise. Further work is needed to enhance filter stability when dealing with measurement anomalies.
- Future Work:
- The authors suggest that their improved MMAE method can be extended to other applications like pattern recognition, target tracking, and fault detection.
- A key area for future research is how to further enhance filter stability, particularly when dealing with measurement anomalies.
Summary updated