Single Room Fire Traceability and Prediction Models Based on Deep Learning Methods

ABSTRACT

When a fire occurs, fast recognition on key information of the fire based on limited data available and predict fire smoke motion trends from limited fire information is desirable to develop effective emergency strategies. The proposed models include fire intensity traceability model and fire position traceability model based on Back Propagation Neural Network (BPNN). Smoke prediction model is proposed based on Transpose Convolutional Neural Network (TCNN). The numerical model is first validated by the single-room fire test, and then a numerical database of 165 transient room-fire scenarios is established under various fire locations, fire sizes, and vent sizes. The model test is conducted in new fire intensity and vent size scenarios, with R² coefficients of 0.965 and 0.95, respectively. The experimental data validation shows that relative errors were less than 10%. The position traceability model has Kappa values of 0.758 and 0.75. The average visibility error values of the smoke prediction model’s output images in the validation of the test set generally fall within the range of ± 1.5 m. The results indicate that the models have good accuracy and strong adaptability to different compartment fire scenarios which can provides effective support for fire rescue and emergency strategy development in fire scenes.

KEYWORDS:

• Fire safety
• transpose convolutional
• fire simulation
• neural network
• fire traceability
• fire intensity
• fire position

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work is supported by the National Key R&D Program of China (Grant Number: No. 2018YFC0810600) and the Key Projects of Science and Technology Research and Development Plan of China State Railway Group Co., Ltd. [No. N2021G033].