Robust Bayesian detection: A case study

Abstract

This paper discusses the use of Bayesian networks in a class of contemporary gas detection/classification problems. In particular, we expose the properties of Bayesian networks which allow creation of detection systems with good performance despite significant deviations between the used models and the underlying true probability distributions. Key to adequate grounding of fusion processes is explicit representation of the locality of causal relations in models of monitoring processes. This provides guidance for a systematic and tractable construction of complex detection systems correlating very heterogeneous information. The resulting Bayesian detection systems are experimentally validated.