Hierarchical multiple sensor fusion using structurally learned Bayesian network

Multiple sensor fusion is very important for wireless health monitoring since a single type of sensor usually can only provide limited aspects of the health condition while multiple sensors of different types hopefully can complement each other and yield more comprehensive aspects of the health condition. Many existing sensor fusion approaches are based on a flat structure, where multiple sensor features are treated as in the same layer and are fused by the feature-level fusion. In this paper we present a systematic approach using a structurally learned Bayesian Network (BN) for sensor fusion. The BN serves as a powerful framework that can integrate multiple sensor features in a hierarchy that is automatically learned via supervised learning. We present a hybrid structure learning approach that includes four steps and consists of both systematic global and local structure learning, as well as random perturbation for structure learning. Subsequent to the feature selection, we first learn an Augmented Bayesian Classifier (ABC) and it is followed by an extended K2 structure learning to search for a better structure in another structure subspace. Random structure learning is then performed to perturb the structure learning so as to avoid getting stuck in a local optimum. Finally, we perform local structure learning with hill-climbing by reversing or removing each link between features. The proposed hierarchical sensor fusion solution outperformed some conventional approaches such as Naïve Bayesian Classifier and Support Vector Machine classifier that integrate multiple sensor features by a flat feature-level fusion.