Bayesian network model for intelligent recognition of typical compartment fire

Based on advanced fire sensors, the Bayesian network model of recognizing fire size and category intelligently was established, six fire characteristic parameters including upper temperature, lower temperature, CO concentration, CO₂ concentration, O₂concentration and light obscuration were considered as the input variables of the model, fire size and category were considered as the output variables of the model, and the relationship between input and output variables was deduced.Four kinds of fire sources including mattress fire, cable fire, pool fire and spill fire were simulated in living room, combat center, engine room and hangar respectively, 2 880 groups of simulated sample data were generated by using CFAST software, the parameters of the model were trained, and the trained recognition model was validated by using full-scale fireexperimental data.Validation result indicates that when the data of fire sensors are intact, the average recognition accuracy rates are 88.0%, 95.0% and 85.7% for small, medium and large fires respectively, and the average recognition accuracy rates are 90.2% and 81.5% for solid and oil fires respectively.When one sensor can't work because of serious damage or the hit of antiship weapon, the average recognition accuracy rates are 82.4% and 82.7% for fire's size and category, only 8.1% and 2.8% less than the rates with integrated fire sensors data respectively.So, the proposed model has good recognition ability and robustness, and can be integrated into ship damage control supervisory system(DCSS)to assist commanders in timely selecting the most effective firefighting methods and tactics.