Application Research Of Bayesian Machine Learning In Fire Forecasting And Source Intensity Back-Calculation

The high temperature caused by fire poses a great threat to personal and property security.After fire occurs,the promptly grasp of the fire source make some sense to fire control.However,due to the influence of high temperature,firefighters and emergency personnel cannot directly enter the fire site to obtain relative parameters.Thus,it is of great significance to develop a scientific fire back-calculation framework based on the measured data around the fire site.Once the on-site personnel obtain the fire source information,they can calculate the temperature change in the fire environment through physical models,which is conducive to the correct prediction of the fire development,so as to make a better fire control strategy.But meanwhile,the timeliness requirement of emergency management also brings a challenge to the rapid response of physical models.This shows that the development of high-accuracy and low-cost models for fire forecasting and source intensity back-calculation can provide scientific support for fire emergency management.At the same time,the physical processes of these two pro and con problems are opposite to each other,and their solutions can provide some reference for the all-round development of fire science.The prediction of fire temperature generally depends on numerical simulation methods,but the existing ones cannot simultaneously meet the low-cost and high-accuracy requirements for emergency management.With the rise of artificial intelligence,various numerical surrogate models have been successfully applied to the fire research,but these methods only become effective if they are trained with lots of high-fidelity data.In order to explore the possibility of further reducing the modeling cost of surrogate models,this paper introduced a nonparametric Bayesian regression model,namely Cokriging model,into the temperature prediction of a single-room fire.Computational fluid dynamics software FDS and fire zone simulation software CFAST provided high-and low-fidelity training data for the model,respectively.The results of the leave-one cross validation showed that this model had been effectively trained with only small amount of high-fidelity data,while taking the vent size,heat release rate of fire and ambient temperature as input parameters,and the temperature of the upper/lower layer as output variables.To further explore the model performance in emergency prediction,this paper also carried out a detailed analysis from two aspects of time cost and prediction accuracy.Compared with the commonly used numerical simulation methods,the prediction accuracy of Cokriging was much higher than that of CFAST,and its prediction results were very close to the simulations of FDS,whereas the model response time was only about 1 second.This indicates that the Cokriging model can make a rapid and accurate prediction of fire temperature,which can be regarded as an effective fire numerical surrogate model.Through comparative analysis,the multi-fidelity model developed in this paper performed as well as single-fidelity surrogate models like ANN and Kriging,but the modeling time cost was only 1/10 of the latter,which satisfied the rapid modeling requirements under emergency conditions.Finally,on the premise of keeping the total number of training data unchanged,the prediction results under different proportions of high-fidelity and low-fidelity data-were compared and tested.The results show that even with very small amount of high-fidelity data,the CoKriging model can also make a relatively accurate prediction of the smoke layer temperature,which reflects its powerful capability in data fusion.The back-calculation of fire source is often highly nonlinear.However,the existing solutions either do not work well under complex fire scenes,and the calculation results are highly uncertain,or they can only make a rough estimation of the fire location,which are unable to provide information about source intensity.Different from the optimization algorithm,the Bayesian inversion framework can take full account of the error caused by experimental observation and model simplification,and finally presents the credibility of the inversion results in the form of probability distribution.In order to explore the feasibility of Bayesian inversion framework in pool fire,two back-calculation models respectively estimated for pool diameter and heat release rate were developed based on the literature data from toluene and n-heptane combustion experiments.Two thermal radiation models developed by Shokri and Beyler were introduced respectively as forward prediction methods for the two research examples.The results show that after 50,000 iterations,the inversion framework can estimate the pool diameter accurately within the error range,and the posterior sampling results approximate to a low-dispersion Gaussian distribution.However,for the inversion study of heat release rate,the Bayesian framework worked out two posterior peaks,and the sampling result corresponding to one peak was close to the real value,which show the ill-posedness of the fire inversion problem.It is found that the number of the observed data also have a certain impact on the inversion results:when the observed data were too few(less than 3 sets),the model were unable to calculate for pool diameter.In addition,when the observed data were more than 3 sets,the posterior mean value of the inversion results did not change significantly with the increase of observed data,but the dispersion degree of the posterior distribution gradually decreased.