Forty years of distributed estimation: A review of noteworthy developments

Distributed estimation has been an active research area since the late 1970's, starting with decentralized filtering and estimation for control of large scale systems. Early research focused on optimal fusion to reconstruct the best global estimate by de-correlating the local estimates to avoid double counting of information. This approach was extended to distributed estimation for general sensor networks using information graphs in the 1980's, with channel filter as a special case for acyclic networks. Track fusion for systems with process noise led to a general framework for linear estimate fusion given cross covariances in the 1990's. Fusion rules include maximum likelihood, maximum a posteriori probability, and best linear unbiased estimate (BLUE). BLUE is based on a unified fusion model, which has been used to develop other fusion rules. In addition, covariance intersection handles situations where correlation information is completely unknown. Advances in sensing and communication hardware have resulted in low-cost wireless ad hoc sensor networks around 2000. These require robust distributed estimation algorithms such as consensus-based filters and diffusion-based filters, which have other performance criteria than estimation accuracy. There was recent resurgent interest in distributed estimation algorithms that handle process noise. Examples are distributed Kalman filters using pseudo estimates and distributed accumulated state density filters.

I. Introduction

Many sensing and estimation systems involve multiple sensors and one or more processors. Sensors with complementary spatial or spectral coverage can provide information not available from a single sensor. When multiple spatially distributed sensors are present, choosing the appropriate estimation architecture is an important problem. A centralized architecture with a central processing node can provide the best estimate but requires more communication and has a single point of failure. A distributed architecture with multiple nodes processing local sensor measurements and communicating with other nodes can reduce communication and improve robustness.