An expert system based on a Bayesian network for fire safety analysis in nuclear area

Abstract For fire safety studies in nuclear installations, IRSN uses the SYLVIA software. The SYLVIA two-zone model was developed by IRSN to simulate a full ventilation network, fire scenarios in a highly confined and mechanically ventilated facility, and airborne contamination transfers inside nuclear installations. In order to take into account the different sources of uncertainty coming from initial and boundary conditions as well as from model parameters, the SYLVIA software is associated with the SUNSET statistical software. However, such a use of SYLVIA software has a major drawback: it requires a large number of runs and a significant statistical analysis what is not always compatible with the requirements of safety assessments in terms of deadlines. To overcome this difficulty, IRSN is currently developing an expert system based on a SYLVIA database. This approach allows deriving the most likely diagnosis or prognosis in a very short time, but also deriving a more complex form of reasoning intertwining prognostic and diagnostic inferences. The proposed expert system is based on the Bayesian Belief Network (BBN) methodology and consists in two steps: First, a large database obtained from SYLVIA runs allows the estimation of Conditional Probability Tables. Then, a message passing algorithm is used to exploit dynamically this data base. The illustrating example is based on the study of pressure effects due to fire scenarios in nuclear facilities and the database is made up of 1,600,000 runs of the SYLVIA software. The goal of this paper is to detail the methodology and process to carry out an expert system for fire safety studies, and is supported by one example showing how it can be used as a decision support tool for fire safety analysis in nuclear area. To our opinion, the development of expert systems represents a new generation of computational tools in the field of probabilistic fire simulation.