ABSTRACT
This chapter describes a sensor fusion approach that builds a model of failure and derives a fusion rule that is sensor failure robust by minimizing Bayesian risk under the failure model. It examines both the cases of dependent and independent sensor failures. The chapter demonstrates how one can model the failed sensor as drawing from some distribution of characteristics. It then focuses on three numerical experiments that demonstrate the robust fusion methods. The first example examines how robust fusion tempers the trust of a stronger classifier in a pool of weaker classifiers because of its failure rates. The second example simulates the placement of a grid array of sensors to detect a target. This example demonstrates dependent sensor failures as well as unknown failure characteristics. The third example models artifact conditions as sensor failures to make better decisions in a brain–computer interface using real data.
