Prediction-Inversion Models of Tunnel Fires by Tunnel Flame Images Under Machine Learning

Abstract

In this paper, a machine learning-based tunnel fire prediction-inversion model is proposed to solve the dynamic evolution relationship model of fire image-ceiling temperature-heat radiation-heat release rate, which is difficult to establish from mathematical relationships. And the caution of fire data is primally due to the difference and high cost associated with conducting real-scale tunnel fire experiments. In order to establish a fire information database, this paper conducted fire experiments in 1:10 scale tunnels, collected fire parameters such as roof temperature, thermal radiation, heat release rate and flame images under different scale tunnel fires, and constructed a fire database. Subsequently, a neural network prediction model for tunnel fires based on machine learning was proposed. The prediction model is able to predict the development of tunnel fires. Meanwhile, the tunnel fire inversion model was established by recognizing the inversion of the prediction results and obtaining other fire parameters such as the heat release rate of the fire source corresponding to the image. The dynamic correlation of information such as flame image-heat release rate-fire temperature-heat flux was realized. The prediction accuracy of the model reaches 90% in terms of indicators such as mean absolute error and structural similarity index. The model can be used as a prediction method to guide fire suppression and rescue operations in tunnel fires.