In The Netherlands in particular for example, safety regulations on tunnel design have put emphasis on functional features leaving a number of issues to be determined by the group of organisations concerned in one way or another with the construction. Better informed manners for analysing the risks and communicating such analyses at design stage are being investigated. The aim is to produce methods or tools to help evaluate the impact of design parameters and safety facilities in different scenarios while estimating their likelihood. A Bayesian network has been produced using HUGIN to evaluate its suitability in engineering risk analysis and in the decision making at planning stage. As an example of this application a preliminary object-oriented network was made related to tunnel fires with the purpose of estimating the heat flux probability distribution of a fire in a 1 km length tunnel. HUGIN seems particularly suitable to simulate the event of a tunnel fire due to the rarity of such happening, the difficulty in performing enough experimental tests and therefore the lack of significance of statistical evidence. HUGIN allows the inclusion of objective and subjective probability, in the form of frequency-based and expert-based estimates. The model presented here is a Bayesian network that could be transformed into an influence diagram for greater application in the decision making. The network consisted of five sub-networks and one main network encompassing all. These are as follows:

• Fire load Involved (a measure of the amount of energy to be released by the materials involved in a fire). The CPT here was constructed by estimating the probability distribution of the type of vehicles likely to be inside the tunnel and their fire loads. The estimation was made both using statistical information and subjectivity.
• HRR (heat release rate or power released by the fire in the form of heat). Using expert knowledge, the highly complex phenomenon of maximum HRR development was simplified into a CPT specific for a tunnel fire. The resulting CPT is a simple version but considered valid nonetheless for the purposes here. The parents of the HRR node are the fire load sub-network, the tunnel ventilation and the presence and type of sprinklers. The resulting probability distribution is of course conditional on the assumptions made.
• Losses . This sub-network estimates the probability distribution of the heat fluxes in the tunnel, varying with distance from the fire and at two moments in time, five minutes and ten minutes after the start of the fire. The last node is based on the HRR maximum sub-network and its relationship with other nodes representing other physical parameters involved in the heat transfer phenomenon according to known engineering calculations. Eventually other sub-networks could be included in this model specifying the position of people inside the tunnel to estimate their exposition to heat fluxes and to smoke.
• Yearly fire rate. This sub-network estimates the probability distribution of the number of fires in this tunnel, based on a Poisson distribution. The parameter of the Poisson distribution is made of the number of vehicles crossing the tunnel per year, the probability of a fire per vehicle kilometer in a tunnel, the length of the tunnel (but here this has been fixed to 1 km) and the probability of the fire being mitigated with a handheld fire extinguisher of fire hose.
• View Factor. This sub-network calculates the view factor, parameter that depends entirely on the geometry of the fire and the receiver of the heat, and their relative positions. This sub-network is then used as an object in Losses.
• Main Network. Establishes the causal relationship between HRR and Losses. This sub-network represents the entire network, running this network provides the user with the joint probability distribution of all the nodes in all the sub-networks.

Building a tunnel fire belief network might require more time than producing a more traditional risk analysis diagram such as a fault tree or an event tree. This is due largely to the inclusion of uncertainty in the form of conditional probability tables and the somewhat more detailed understanding of a phenomenon that may be required to build a Bayesian network and/or to simplify the domain of the problem. However the positive difference is in the quality and scope of the information obtained in comparison with the other methods. The Bayesian network works as an intelligent system, it is much more versatile and robust because it deals with causal relationships and their associated uncertainties and the fact that the links also can emulate the reasoning of a human expert in the subject. Therefore the quality of the information obtained from this kind of model -in particular an object-oriented network, capability offered by HUGIN- is excellent in communicative and informational quality. In tunnel construction design, it is potentially very useful for better informed decision making on tunnel design of safety facilities such as fire emergency exits.

This project was completed at TNO Bouw, The Netherlands and presented to the University of Aberdeen, Scotland to fulfill partial requirements for the degree of Master of Science in safety engineering and risk management.