Multi-sensor fault recovery in the presence of known and unknown fault types

This paper proposes an efficient online, hybrid, Bayesian multi-sensor fusion algorithm for target tracking in the presence of modelled and unmodelled faults. The algorithm comprises two stages. The first stage attempts to remove modelled faults from each individual sensor estimate. The second stage de-emphasises estimates which have been subject to unanticipated faults and are still faulty despite undergoing the Stage 1 fault recovery process. The algorithm is a computationally efficient and decentralisable hybrid of two standard approaches to fault detection, namely model-based fault detection and majority voting. The algorithm is tested on two distinct simulated scenarios (1) when the target process model does not match reality and (2) in the presence of simultaneous modelled and unanticipated faults.