Abstract: In order to help fire investigators to determine the fire origin more accurately and efficiently, a BP neural network-based fire point prediction model is proposed in this paper. The soot deposition database of wall soot deposition under 59 different fire origin scenarios is constructed by numerical simulation of single compartment fire, and the wall soot deposition characteristics under representative fire origin scenarios are analyzed, which indicates a strong correlation of the fire origin location with the mass of wall deposition and the average value of the maximum concentration. The above two parameters are selected as input, and the fire origin location is used as output for network training. And the new data is used for prediction. The results show that the maximum absolute error of the predicted value is 0.65 m, the minimum absolute error is 0.03 m, and the average absolute error is 0.37 m, indicating that the proposed model can achieve the prediction of fire source location with relatively high accuracy and is a good alternative method for fire investigation.

Key words: BP neural network, soot deposition, numerical simulation, single compartment, fire origin