Real-time forecast of tunnel fire scenario and hazard based on external smoke images

Abstract

In tunnel fires, conventional detectors are often obstructed by smoke and heat accumulation, impeding decision-making for firefighting and evacuation efforts. This work proposed a real-time prediction of tunnel fire heat release rate (HRR) and location by using external smoke images and deep learning algorithms. A 100-m tunnel was simulated. By varying fire HRR, location, and fuel type, a database of about 1.6 million external smoke images was formed, which was used to train a convolutional neural network (CNN) model and produced a R² of 0.99 and a small MSE less than 0.03. Results demonstrated that this intelligent method can accurately identify transient tunnel fire power and fire source location, with the HRR prediction error being less than 20 % and location error less than 1.5 m, irrespective of the stage of fire development or whether the fire is in a stationary or moving state. Results also showed that there are two ways of model training to ensure prediction accuracy while saving computational cost: 1) training by both front and side views of smoke at a single tunnel gate, and 2) training by only front views of smoke at both tunnel gates. This work contributed an intelligent method to assess tunnel fire scenarios and hazards from a safe distance, showcasing its potential application in smart tunnel fire protection and firefighting systems.