Compartment Fire Traceability And Prediction Models Based On Neural Network And Large Eddy Simulation

Compartment fires are more threatening to the personal safety than open space fires.A misjudgment of the hazard level in compartment fires is likely to result in injury or death during the rescue process.Therefore,the immediate determination of the hazard level in the compartment during the fire rescue process is of great significance to firefighting and rescue strategy decisions.This paper develops a fire location recognition,intensity traceability,and smoke prediction model for compartment fire scenarios based on neural network and numerical simulation techniques.The location recognition model and the intensity traceability model can output information of the fire location and fire intensity,respectively.Gas temperature,CO₂ concentration,and ventilation condition（door opening width）were used as the input parameters for the location recognition and intensity traceability models.The smoke prediction model can predict the smoke characteristics based on fire location,fire intensity and door opening width,and output a predicted smoke image with a size of 100 x 100.The typical smoke characteristics such as smoke layer height and visibility can be judged from the prediction results to determine the indoor hazard level.First,a compartment fire database containing 165 different compartment fire scenarios was obtained through the fire simulation software FDS.90 of the scenarios were used as the training set and 75 of the scenarios were used as the test set.A training dataset containing 4500 sets of data and 1800 smoke images was formed for the training of the model.Then the three model structures were determined after continuous testing.The location recognition model and the intensity traceability model both used fully connected network as the architecture,and the model structures were 3-layer and 4-layer respectively.TCNN was selected as the architecture of smoke prediction model,and the model structure is a deep neural network with 24 layers.Finally,the reliabilities of the models were verified by test sets A,B and experimental data,respectively.Where test set A contained 45 sets of new fire intensity scenarios and test set B contained 30 sets of new door width scenarios.The kappa coefficients of the location recognition model in the test set A and B validation were 0.706 and 0.75,respectively.the consistency reached a high level of agreement.The R² coefficients of the intensity traceability model were 0.965 and 0.950 in test set A and B,respectively,and the relative errors of all scenarios were less than 10%in the experimental data validation.The smoke prediction models were all able to predict indoor smoke characteristics（smoke layer height,visibility,etc.）well in the validation of the test set,and the average value of visibility errors with the numerical simulation output smoke images were basically within±1.5m.The validation results show that the fire location recognition,fire intensity traceability and smoke prediction models developed in this paper can meet the requirements in terms of accuracy and have strong adaptability to different compartment fire scenarios.It can be used to determine the fire hazard level in the actual fire scene based on the field information and provide support for personnel to make decisions on fire fighting and escape strategies.