Comparison and Analysis of Centralized, Decentralized, and Federated Filters

Abstract

This paper investigates the implementation of three variations of the Kalman filter—centralized, decentralized, and federated. The underlying theory for these designs is first overviewed, and the designs are then compared on the basis of accuracy, computational efficiency, and fault-tolerance performance. Special attention is paid to the federated filter design and its potential use for fault detection and system recovery. Two numerical examples are included to illustrate the three different filters. The first is the simulated navigation of an object traveling in a straight-line trajectory that is being positioned by two sensors, while the second is a GPS positioning problem using pseudorange and carrier phase observations. Blunders are purposely planted in these two data sets. The numerical results presented support the hypothesis that better fault-tolerance performance can be achieved with a federated-decentralized filter than with a centralized filter approach.