decentralised Data Fusion: A Graphical Model Approach (Summary)
Data fusion is a task in which various types of data are collected from multiple sources, which are then combined to form meaningful information. Consequently, this information will be used as the input of reasoning systems for achieving inference. It is clear that this form of information gathering from multiple platforms is more efficient and applicable than if they were obtained accumulatively from different sources <ref> Reference 1 </ref>. The process of fusing data is categorized to centralized and decentralized data fusion.
In centralized data fusion, the data is collected and combined in a central node, and then fused information is inferred using the central node. On the other hand, decentralized data fusion is a task in which the process of forming, communication and assimilation of data at remotes sites is in a way that guarantees efficient reliable operation; In fact, the main problem of DDF is finding a method to understand and formulate this process. Having the significant advantages of modularity, scalability and robustness, decentralized data fusion has kept the attention of many scholars, and is the main contribution of this paper. This paper addresses the problem of Decentralized Data Fusion (DDF) using Graphical Models techniques. Decentralized Data Fusion problem can be defined as fusing data coming from different platforms, i.e. sensors, which are attempting to estimate a common state of interest that might be partially observed by each of the platforms. As a solution, graphical models incorporates a uniform network-like model of state and observation relations as well as communication and assimilation algorithms which are aimed on distributed message passing and local inference. The authors in this paper applied the concepts from graphical models to both the problem formulation and the solution of decentralized data fusion problem. For formulation part, they have used the well-known Channel Filter algorithm, and moving through solution, they have proposed two ideas: graphical models based on 1) cloned variables and 2) variable cliques. The main part of this work is based on the decentralized graphical model architecture introduced by Paskin and Guestrin <ref> Reference 1 </ref> that is applied to the problem of data fusion here. The formulation begins with representing the physical communication links between platforms with a connectivity graph. Furthermore, it is assumed that that data fusion can be done in both static and dynamic models with static topology. Figure below shows two examples of static and dynamic models.
where in the left depiction, [math]\displaystyle{ \ x }[/math] is the state of interest and observations are shaded ovals, and for the right one,