Introduction
In the realm of modern navigation, the integration of Strap-down Inertial Navigation Systems (SINS) and Global Navigation Satellite Systems (GNSS) is the industry standard for determining vehicle position and velocity. While SINS provides high-frequency, autonomous data, it suffers from drift over time. GNSS provides absolute positioning but is susceptible to signal blockages and multipath interference. To fuse these sources, engineers typically rely on the Cubature Kalman Filter (CKF). However, standard CKF implementations often struggle in real-world conditions where noise statistics are unpredictable and non-Gaussian. This paper introduces a robust alternative: the Adaptive Generalized Maximum Correntropy Cubature Kalman Filter (AGMC-CKF).
The Challenge: Non-Gaussian Noise and Process Uncertainty
Traditional CKF designs are built on the Minimum Mean Square Error (MMSE) criterion. This mathematical framework assumes that sensor noise follows a "Gaussian" or normal distribution (the classic bell curve). In controlled environments, this assumption holds, but real-world navigation is rarely so tidy.
Two primary issues degrade standard filter performance:
- Process Uncertainty: When a vehicle undergoes sudden, sharp maneuvers, the internal motion models used by the SINS can become inaccurate, leading to "process noise" that does not fit the Gaussian model.
- Measurement Outliers: In "urban canyons" or under heavy foliage, GNSS signals can bounce off buildings or be partially blocked. This results in "glitchy" data points—outliers that are far outside the expected range.
Because the MMSE criterion squares the error, these large outliers exert a disproportionate influence on the final estimate, causing the navigation system to "jump" or lose track entirely.
The AGMC-CKF Solution
The proposed AGMC-CKF addresses these vulnerabilities by replacing the MMSE criterion with an Adaptive Generalized Maximum Correntropy (AGMC) framework.
1. From MMSE to Correntropy
Instead of minimizing the square of the error, the AGMC-CKF uses a correntropy-based cost function. Correntropy acts as a local similarity measure. By employing a Generalized Gaussian kernel, the filter effectively "ignores" large, improbable errors (outliers) while focusing on high-probability data. This creates a saturating effect where extreme errors are capped, preventing them from corrupting the state estimate.
2. Statistical Linearization and Regression
To maintain computational efficiency while handling the complex nonlinearities of SINS/GNSS fusion, the authors utilize a statistical linearization approach. This transforms the filtering problem into a linear regression model, allowing the predicted state and new measurements to be processed simultaneously in a batch-mode format.
3. Adaptive Forgetting Scheme
The "Adaptive" component of the filter is driven by a Recursive Least Squares (RLS) algorithm that estimates a fading factor ($\lambda_k$) in real-time. This allows the filter to dynamically adjust its reliance on past data. When the vehicle's motion becomes unpredictable, the filter "forgets" old, potentially irrelevant process noise statistics and adapts to the current dynamics.
4. Outlier Identification
To further harden the system, the researchers incorporated a hypothesis test based on the Mahalanobis distance. By calculating the mathematical "distance" of a new measurement from the expected state, the filter can identify and reject outliers before they enter the update cycle.
Experimental Validation
The researchers validated the AGMC-CKF through both high-fidelity numerical simulations and real-world car-mounted experiments in Taiyuan, China. The test vehicle was driven for nearly eight minutes along a challenging route designed to induce vibrations and signal occlusions.
Key Performance Metrics:
- Velocity Accuracy: In real-world trials, the AGMC-CKF maintained a velocity error between 0.06 and 0.07 m/s. In contrast, the standard CKF produced errors as high as 0.39 m/s.
- Position Accuracy: The improvement in positioning was significant. For North-axis positioning, the AGMC-CKF achieved a Root Mean Square Error (RMSE) of 3.80 meters, nearly doubling the precision of the standard CKF, which recorded an RMSE of 7.74 meters.
- Computational Efficiency: Despite its increased complexity, the AGMC-CKF remains practical for real-time use. Through mathematical optimization that reduced the number of required matrix inversions, the filter's processing time was kept to approximately 0.438ms per step, only slightly higher than the basic CKF's 0.408ms.
Conclusion
The AGMC-CKF represents a significant step forward for robust navigation in unpredictable environments. By moving beyond the limitations of the MMSE criterion and incorporating adaptive forgetting schemes, the filter provides a stable, accurate solution for SINS/GNSS integration. Future work will focus on automating the selection of kernel parameters to make the system more adaptable to diverse operational environments without manual tuning.
Summary updated