Robust Central Difference Kalman Filter With Mixture Correntropy: A Case Study for Integrated Navigation

Kaiqiang Feng, Jie Li, Debiao Zhang, Xiaokai Wei, Jianping YinIEEE Access 2021 · DOI

Introduction

In the field of integrated navigation, particularly for Micro-Electro-Mechanical Systems (MEMS) based Strapdown Inertial Navigation Systems (SINS) fused with Global Navigation Satellite Systems (GNSS), the Kalman Filter (KF) serves as the foundational estimation framework. However, the efficacy of the standard KF relies heavily on two primary assumptions: the system model is linear, and the noise follows a Gaussian distribution.

In real-world operational environments, these assumptions are frequently violated. High-dynamic maneuvers, such as aggressive cornering or rapid acceleration, introduce significant process uncertainty. Simultaneously, GNSS signals are often compromised by multipath effects or signal occlusions in urban canyons, leading to non-Gaussian measurement noise characterized by heavy-tailed distributions and outliers. Under these conditions, conventional filters like the Extended Kalman Filter (EKF) or the Central Difference Kalman Filter (CDKF) can suffer from performance degradation or even divergence. This paper introduces a Robust Central Difference Kalman Filter (RCDKF) designed to mitigate these challenges simultaneously.

Methodology

1. Statistical Linearization

To address the nonlinearities inherent in navigation systems without the complexity of Jacobian matrices required by the EKF, the proposed RCDKF utilizes a statistical linearization approach. This method transforms the nonlinear state and measurement equations into a linear-like regression model. This transformation is crucial as it facilitates the application of robust cost functions that are typically easier to optimize in a linear framework.

2. Adaptive Maximum Mixture Correntropy Criterion (AMMCC)

The core innovation of the RCDKF is the replacement of the standard minimum mean square error (MMSE) criterion with the Maximum Mixture Correntropy Criterion. Correntropy is a local similarity measure that is inherently robust against outliers. While standard correntropy-based filters use a single Gaussian kernel, the RCDKF employs a mixture of two Gaussian kernels. This "mixture correntropy" provides a more flexible cost function that can be tuned to be highly sensitive to small errors (maintaining precision) while remaining remarkably insensitive to large, non-Gaussian outliers.

3. Adaptive Fading Factor

To combat process uncertainty—errors arising from the vehicle's motion not perfectly matching the mathematical model—the researchers integrated an adaptive fading factor. Derived using innovation orthogonal theory and Cholesky triangular decomposition, this factor dynamically adjusts the weight of past observations. During periods of high maneuverability or model mismatch, the fading factor increases the influence of the most recent measurements, preventing the filter from becoming "sluggish" and ensuring it tracks the vehicle's true trajectory accurately.

Experimental Validation

The performance of the RCDKF was evaluated through both high-fidelity simulations and real-world field tests.

Simulation Study

A complex 488.6-second trajectory was simulated, incorporating 18 distinct motion phases including climbs, descents, and sharp turns. To test robustness, the researchers injected non-Gaussian noise (a mixture of Gaussian distributions with high variance) into the measurement data. The RCDKF was compared against several benchmarks: the standard CDKF, a fading CDKF (FCDKF), the Huber-based Mixed Filter (HMF), and other state-of-the-art robust filters.

Car-Mounted Field Test

A real-world experiment was conducted using a vehicle equipped with a custom MEMS-SINS unit. The vehicle traversed a "bumpy road" environment with intermittent GPS occlusions caused by surrounding foliage and buildings. A high-precision LCI-1 tactical-grade IMU and a Propak satellite receiver provided the ground truth for error analysis.

Results and Discussion

The experimental results consistently demonstrated the superiority of the RCDKF in terms of both accuracy and stability.

  • Estimation Accuracy: In the field tests, the RCDKF achieved a Root Mean Square Error (RMSE) for East Velocity of 0.105 m/s, a significant improvement over the 0.651 m/s recorded by the standard CDKF. Similar improvements were noted in position and attitude estimation.
  • Attitude Stability: The filter maintained high precision in pitch and roll (errors of approximately 0.203° and 0.106°, respectively), even when subjected to mechanical vibrations and sensor glitches that caused other filters to drift.
  • Computational Efficiency: A critical requirement for navigation filters is real-time feasibility. The RCDKF's computational overhead is minimal; the average execution time per step was 0.452 ms, compared to 0.432 ms for the standard CDKF. This negligible increase makes it highly suitable for deployment on embedded hardware.

Conclusion

The Robust Central Difference Kalman Filter (RCDKF) presented in this study offers a comprehensive solution for integrated navigation in challenging environments. By combining statistical linearization, a mixture correntropy cost function, and an adaptive fading factor, the algorithm effectively suppresses both process uncertainty and non-Gaussian measurement noise. While the current implementation requires manual tuning of kernel parameters, future research will focus on automated parameter selection to enhance the filter's adaptability across diverse sensor platforms and environmental conditions.

Summary updated

BibTeX
@article{feng2021robust,
  author = {Kaiqiang Feng and Jie Li and Debiao Zhang and Xiaokai Wei and Jianping Yin},
  title = {{Robust Central Difference Kalman Filter With Mixture Correntropy: A Case Study for Integrated Navigation}},
  journal = {IEEE Access},
  year = {2021},
  doi = {10.1109/access.2021.3083963},
}
Figure