Predicting smoke hazards and burning fuel via smart building fire sensor network and dual-agent deep learning

Abstract

A smart building should quantify real-time fire hazards in complex-built environments and support real-time emergency response. This work introduced a smart framework using dual-agent deep learning to predict real-time fire and smoke hazards in a 30-m corridor by feeding point-sensor data. After validation with four full-scale corridor tests, a numerical database was established, consisting of 100 fire scenarios with varying fire locations, intensities, growth curves and fuel materials. The Point Model, using fully connected neural network (FCNN), can read sensor data of temperature, extinction coefficient and CO concentration to predict real-time fire location, heat release rate and type of burning fuels. The Field Model, consisting of long short-term memory (LSTM) and transposed convolutional neural network (TCNN), inputs past 10-s point readings of temperature, extinction coefficient or carbon monoxide (CO) concentration to predict their 2D field and evaluate smoke hazards. The overall prediction accuracy was above 96% even if two out of five sensors failed, showing a high system resilience. Moreover, the dual-agent model successfully predicted the inverse smoke stratification phenomena, demonstrating its capacity to address complex fire incidents. The proposed dual-agent method can provide crucial building fire information with existing fire sensors, which can support firefighters in making decisions and reduce fire casualties.