Bayesian Networks for Source Intrusion Detection

Abstract

Bayesian belief networks are graphical probabilistic analysis tools for representing and analyzing problems involving uncertainty. The problem of monitoring the propagation of a contaminant in a water distribution system can be represented by using Bayesian networks (BNs). The presented methodology proposes the use of BN statistics to estimate the likelihood of the injection location of a contaminant and its propagation in the system. A clustering method, previously developed by the authors, is first applied to formulate a simplified representation of the distribution system based on nodal connectivity properties. Given evidence from clusters, information is combined through probabilistic inference using BNs to find the most likely source of contamination and its propagation in the network. The conditional independence assumptions with the BNs allow efficient calculation of the joint probabilities and diagnostic and predictive queries (e.g., the most likely event given evidence or the probability of an outcome given starting conditions). In addition, a theoretic information measure is suggested to evaluate the significance of the clusters relying on the BN model of the system and possible optimal sensor locations. The proposed methodology is developed and tested on two water supply systems.