Correntropy Based Divided Difference Filtering for the Positioning of Ships

Xi Liu, Badong Chen, Shiyuan Wang, Shaoyi Du

TL;DR

This paper introduces novel "correntropy-based divided difference filters" (CDD1 and CDD2) designed to improve ship positioning accuracy, especially when dealing with noisy sensor data. These filters combine the strengths of traditional divided difference filters with the robustness of correntropy to handle heavy-tailed, non-Gaussian noise, outperforming standard methods in challenging conditions.

What Problem Does This Paper Address?

Accurate ship positioning is crucial for navigation and safety. Modern systems often integrate Dead Reckoning (DR) and Global Positioning System (GPS) to leverage their complementary strengths. DR provides real-time position based on speed and direction but suffers from accumulating errors. GPS offers absolute positioning but has limitations like low sampling rates and susceptibility to interference.

While integrating DR and GPS improves overall performance, filtering these combined signals presents a challenge. Traditional Kalman filters and their extensions (EKF, UKF) work well for linear or mildly non-linear systems with Gaussian noise. However, real-world ship navigation can encounter non-Gaussian noise due to environmental factors (e.g., rough seas) or sensor malfunctions, leading to heavy-tailed distributions. Standard filters, which often rely on minimizing mean squared error, are sensitive to such noise and can lead to inaccurate or unstable estimates. This paper aims to address this by developing robust filtering techniques that can handle these non-Gaussian noise scenarios effectively.

Methods

The core contribution of this paper is the development of two new filtering algorithms: the Correntropy-Based First-Order Divided Difference (CDD1) filter and the Correntropy-Based Second-Order Divided Difference (CDD2) filter.

These filters build upon the Divided Difference Filter (DDF) framework. DDFs approximate nonlinear system dynamics using polynomial expansions, offering advantages like guaranteed positive semi-definiteness of the covariance matrix and fewer parameters than some alternatives like the Unscented Kalman Filter (UKF). The paper utilizes both first-order (DD1) and second-order (DD2) approximations of the DDF.

The novelty lies in incorporating the Maximum Correntropy Criterion (MCC) into the DDF framework. Correntropy is an information-theoretic measure that captures higher-order statistics of data and is known for its robustness against heavy-tailed noise. By reformulating the filtering problem under the MCC, the paper introduces a robust cost function that downweights outliers. This leads to a re-weighting of the covariance matrix and a modified update step for the filters. The authors derive the CDD1 and CDD2 filters by applying this MCC to the nonlinear regression models derived from the state and measurement equations.

For comparison, the paper also evaluates standard filters like the Extended Kalman Filter (EKF), DD1, UKF, DD2, and Huber-based DDF variants (HDD1, HDD2).

Datasets

The paper does not introduce new datasets. Instead, it uses a simulated ship navigation scenario for evaluation. The simulation parameters, including process noise distributions (assumed Gaussian for initial tests) and initial conditions, are provided in Tables 1 and 2. The simulation is run for 1200 seconds with 100 independent Monte Carlo experiments.

Key Insights and Ideas

  • Correntropy for Robustness: The paper successfully demonstrates that incorporating correntropy into DDFs significantly enhances their robustness against heavy-tailed, non-Gaussian noise, a common issue in real-world navigation systems.
  • DDFs as a Solid Foundation: The choice of DDFs as the base framework is justified by their ability to maintain positive semi-definiteness and their relative simplicity compared to some other nonlinear filters.
  • Kernel Bandwidth is Key: The performance of the correntropy-based filters is highly dependent on the kernel bandwidth ($\sigma$). A smaller bandwidth generally leads to better robustness, but too small a value can cause instability. A very large bandwidth makes the CDD filters behave similarly to their non-robust counterparts.
  • CDD2 Outperforms: In the non-Gaussian noise scenarios, the CDD2 filter, with an appropriate kernel bandwidth (specifically $\sigma=2$), achieves the smallest errors, outperforming even other robust filters like HDD2.
  • Practical Implications: The findings suggest that these CDD filters are well-suited for applications like ship positioning where sensor noise can be unpredictable and deviate from ideal Gaussian distributions.

Key Results

The paper presents simulation results comparing the performance of various filters under both Gaussian and heavy-tailed non-Gaussian noise conditions.

  • Gaussian Noise: Under Gaussian noise, the proposed CDD1 and CDD2 filters perform similarly to their non-robust counterparts (DD1 and DD2, respectively). The robust filters (CDD1, CDD2, HDD1, HDD2) generally show slightly worse performance than the non-robust ones when the noise is purely Gaussian, especially with larger kernel bandwidths.
  • Non-Gaussian Noise: This is where the CDD filters shine.
  • The robust filters (HDD1, HDD2, CDD1, CDD2) significantly outperform the non-robust filters (EKF, DD1, UKF, DD2).
  • The CDD2 filter, with a kernel bandwidth of $\sigma=2$, achieves the lowest Total Mean Squared Error (TMSE) for both latitude ($\lambda$) and longitude ($\phi$) positions, showing a substantial improvement over other filters. For example, its TMSE for $\phi$ is around 39.0126, compared to over 91 for EKF and DD1.
  • With a larger kernel bandwidth ($\sigma=20$), the CDD filters' performance degrades and becomes similar to the standard DD filters, highlighting the importance of proper tuning.

Limitations and Future Work

  • Kernel Bandwidth Tuning: The paper acknowledges that choosing the optimal kernel bandwidth is crucial and can be challenging in practical applications. While they show that $\sigma=2$ is effective for CDD2 in their simulation, a systematic method for bandwidth selection is not detailed.
  • Numerical Stability: The paper mentions that in practical applications, Cy can become nearly singular, potentially causing numerical problems for CDD1 and CDD2. They propose a method to address this by conditionally skipping the update step if a certain condition is met.
  • Future Work: The authors suggest several avenues for future research:
  • Combining the proposed filters with adaptive filtering methods.
  • Extending the framework to handle continuous-time systems.
  • Applying the algorithms to other practical examples beyond ship positioning.

Summary updated

Abstract

In this paper, robust first and second-order divided difference filtering algorithms based on correntropy are proposed, which not only retain the advantages of divided difference filters, but also exhibit robustness in the presence of non-Gaussian noises, especially when the measurements are contaminated by heavy-tailed noises. The proposed filters are then applied to the problem of ship positioning. In order to improve the accuracy and reliability of ship positioning, the positioning method combines the Dead Reckoning (DR) algorithm and the Global Positioning System (GPS). Experimental results of an illustrative example show the superior performance of the new algorithms when applied to ship positioning.
BibTeX
@inproceedings{liu2018correntropy,
  author = {Xi Liu and Badong Chen and Shiyuan Wang and Shaoyi Du},
  title = {{Correntropy Based Divided Difference Filtering for the Positioning of Ships}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=4RsKAyXy1n},
}
Figure