Towards self-organizing Kalman filters

Abstract-Distributed Kalman filtering is an important signal processing method for state estimation in large-scale sensor networks. However, existing solutions do not account for unforeseen events that are likely to occur and thus dramatically changing the operational conditions (e.g. node failure, communication deterioration). This article presents an integration solution for distributed Kalman filtering with distributed self-organization to cope with these events. An overview of existing methods on both topics is presented, followed by an empirical case study of a self-organizing sensor network for observing the contaminant distribution process across a large area in time. I. INTRODUCTION A current trend in estimation is to connect different sensor nodes via a scalable network topology and thereby, create "onthe-fly" (ad-hoc) networks for monitoring large-area processes with a high spatial accuracy. This trend is mainly a result from the widely available sensor nodes for setting-up such a (wireless) sensor network. However, sensor networks exhibit special features, which makes their application challenging. For one, sensor nodes are typically battery powered and/or have an energy scavenging device to acquire the consumed energy. In addition, due to power and space considerations, the onboard computational and communication capabilities of sensor nodes are seriously limited. Yet, they are frequently deployed in harsh, even hostile environment, where node failures are not exceptions but part of the normal operation. For example, the changing environment in which nodes exchange data induces dynamics in the network capacity and call for adaptive communication strategies to assure the availability of communication resources. These aspects make the application development extremely challenging, as a stable and predictable system performance should be delivered on a dynamically changing configuration with computational, communication and energy constraints. Distributed data interpretation algorithms with reconfiguration capabilities constitute an important and promising way to address these challenges. In the following, the state estimation problem (based on Kalman filtering) is investigated in this context. Several distributed solutions addressing the Kalman filtering problem have been explored and aim to make use of the local processing elements that are already present in each node. Such distributed Kalman filters typically process the sensor measurements locally at each node rather then communicating them to a single, central node. Characteristic approaches to distributed Kalman filtering are presented in [1]-[4] and the references therein. The proposed methods perform a modified Kalma