Inverse Tracing of Multi-room Fire Sources Based on CFD Simulation, Neural Network and Bayesian Optimization Algorithms

Abstract

In this work, a forward model for fire temperature field reconstruction and an inverse model for fire source tracing were established using CFD simulation, neural networks, and Bayesian optimization to assist in fire investigations. Several hundred datasets were generated through CFD simulation of two- and three-room indoor fires. These datasets were filtered, processed, and then used to train the neural network. To enhance the efficiency and accuracy of the hyperparameter selection for the BP network, Bayesian optimization was employed to determine the optimal hyperparameter. As a result, the two-room forward model achieved a prediction accuracy exceeding 90%, with a confidence interval of 10%. The two-room inverse model reaches accuracies of 95% and 99% under threshold settings of 0.15 m and 0.25 m, respectively, demonstrating remarkable generalizability when dealing with unfamiliar data. The three-room inverse model was trained on the dataset of the ‘I’-type three-room model, resulting in an accuracy of 94.3%. This model was further used to trace fire sources in the ‘L’-type three-room configuration, achieving an accuracy of 92.4%. Although the accuracy decreased, it remained at a satisfactory level, indicating strong extensibility and generalizability.