Computational monitoring systems are often the first line of defense in terms of detecting leaks from cross country pipelines. However, the alarms provided by these systems often require considerable review, assessment and analysis by pipeline control, operations and engineering support personnel to winnow out the few real leaks that are actually embedded in a much larger number of alarms. Detection of such breaches of integrity may require the integration of large amounts of knowledge from multiple sources via the man-machine interface to assess the true probability that the pipeline is actually experiencing a leak. Knowledge integrated by the control room staff may include data from multiple instruments as well as in inherent understanding of how at risk the pipeline is as a result of leaks of varying size. This paper investigates the application of Bayesian Belief Networks (BBNs) to address these issues. These networks integrate disparate sources of information as well as prior probabilities of events as part of an effort to detect leaks, reduce false positives, provide improved event summaries and optimize leak detection thresholds. Bayesian systems provide the unique ability to detect leaks by using prior estimates of the probability for important events such as leaks, the logical structure of impacts that such leaks have on instrument outputs, and re-estimation of the prior leak probabilities based on changes to the leak evidence when expressed in terms of the SCADA system outputs, as well as standard mass balance system computational parameters. The paper consists of three parts. The first section discusses the theoretical background for the application of Bayesian belief networks to this problem. This includes a high level discussion of Bayesian theory, belief node structure, and the application of engineering analysis to provide prior estimates of leak incident and other probability distributions. The second part of the paper provides a discussion of the theoretical application of BBNs to the problem of leak detection, and a description of their application via a case study. This includes a discussion of applicable equations, prior probabilities and their sources, appropriate net structures and a description of the real world pipeline. Results of the analysis are discussed in the third part of the paper. This includes performance results for all of the BBNs, including leak detection sensitivities, false positive rates, PLDS efficiencies, and a comparison of the tested net topologies. The paper concludes with a discussion of the advantages and disadvantages of this approach, and a brief summary of proposed future work. BAYESIAN BELIEF NETS THE GAME SHOW PARADOX
Imagine for a moment that you are a contestant on a game show. At the climax of the show the game show host gives you the choice of three doors. Behind two of the doors the host has tied up a goat. However, the third door has the prize: a suitably expensive and impressively gas-guzzling SUV.