A computer-vision-based deep learning model of smoke diffusion

Abstract

Smoke is a lethal hazard in fire accidents and the biggest obstacle to emergency rescue. Predicting its diffusion range is important in fire command and emergency rescue. This article utilizes computer-vision-based deep learning methods to learn the characteristics of smoke diffusion and achieve real-time prediction of smoke-temperature fields. Firstly, a tunnel fire was simulated to verify the reliability of the deep learning model in predicting smoke-temperature fields. Thirty-six sets of fire simulation conditions were established with various values of the fire source location and heat release rate (HRR) and longitudinal ventilation airflow velocity. This created a database of 17,280 smoke-temperature fields. Based on windless scenario with the fire source located in the middle of the tunnel, it was determined that the SiLU activation function provides the best nonlinear modelling ability and stability in training neural networks. Based on the smoke-temperature-field database, a U-Net network was constructed to predict the smoke diffusion process of the tunnel fire. Under different scenarios, the U-Net model achieved 90 % accuracy in smoke-temperature-field prediction. The proposed U-Net model was then used to predict smoke dispersion in a petroleum storage depot scenario. The results show that this approach can help to delineate safe areas during fire emergencies.