Variational Learning Data Fusion With Unknown Correlation.

Area of Science:

• Signal Processing
• Data Fusion
• Statistical Inference

Background:

• Accurate state estimation in data fusion is challenged by unknown and time-varying correlations between data sources.
• Existing Bayesian methods often struggle with dynamic correlation structures, limiting their applicability.
• Understanding and identifying these correlations is crucial for robust data fusion performance.

Purpose of the Study:

• To develop a Bayesian framework for simultaneously estimating system states and identifying unknown, time-varying data correlations.
• To address complex correlation structures including single-sensor, multi-sensor, and local estimate correlations.
• To improve the accuracy and robustness of data fusion systems under uncertain correlation conditions.

Main Methods:

• Proposed a novel Bayesian learning framework incorporating variational Bayesian inference.
• Represented unknown correlations as a randomly weighted sum of positive semi-definite matrices.
• Derived a closed-form iterative solution by minimizing Kullback-Leibler divergence to jointly estimate states and correlation weights.

Main Results:

• Successfully derived the joint posterior distribution of the state and unknown weights in a closed-form iterative manner.
• Demonstrated superior performance in root-mean-square error for both state estimation and correlation identification across three simulation cases.
• Validated the method's effectiveness in handling diverse and dynamic correlation scenarios.

Conclusions:

• The proposed variational Bayesian approach offers a robust solution for joint state estimation and correlation identification in data fusion.
• The method effectively handles unknown and time-varying correlations, outperforming existing techniques.
• This work advances the field of data fusion by providing a more accurate and reliable estimation framework.