An Ensemble Learning Approach to Covariance Intersection Fusion

Covariance intersection (CI) is an algorithm used for data fusion that combines uncertain information from multiple sources. The algorithm does not demand knowledge of the correlation between these sources. This paper proposes a novel approach to CI fusion using a machine learning ensemble framework, leveraging both polynomial regression (PR) and eXtreme gradient boosting (XGBoost) regression models. A synthetic multi-sensor dataset using the traditional CI algorithm has been generated, driven by an optimal weight parameter ω. The input feature vector for this framework includes spatial information and its associated covariance matrices from multiple sensors. The study aims to fuse these multi-sensor input feature vectors using machine learning techniques to obtain fused spatial information and covariance. This information is obtained through the CI fusion algorithm. This research considers both the PR and XGBoost models together. The ensemble weights are optimized by minimizing the mean squared error (MSE) to achieve minimum loss. This leads to optimal fusion results. The ensemble model infers that while PR gets the higher weights to spatial information, XGBoost regression prioritizes covariance data. This observation aligns with the strengths of each model, where PR captures linear relationships effectively, and XGBoost handles non-linear dependencies as well as complex interactions.