Trustworthy Building Fire Detection Framework With Simulation-Based Learning

With the difficulty of collecting desirable training data due to the heterogeneities of IoT sensors in various buildings and the scarcity of fire events, it is time consuming and expensive to apply data-driven deep learning approaches to fire detection systems in specific building environments. Simulation-based learning has been actively researched to mitigate data scarcity problems by reproducing potential fire events. Since simulation-based learning mainly depends on synthetic training data, trained deep learning models may generate erroneous predictions in real-world scenarios that are unlike any of the training samples. In this paper, we propose a trustworthy building fire detection framework based on a multioutput encoder-decoder network, named MEDNet, which is designed for the practical usage of simulation-based learning in building fire detection. The fundamental steps of our approach are (1) modeling and simulating fire events to create realistic synthetic data that reflect data from actual buildings, (2) predicting a fire event and dissimilarities between real input data and synthetic training data based on the trained MEDNet model, and (3) operating a switching mechanism to use a knowledge-based method that does not depend on synthetic training data when dissimilarities exist. Finally, we perform simulation experiments based on a real building compartment where the proposed framework is compared with conventional time-series classification networks on various evaluation datasets. The proposed framework is trustworthy in practical usage because MEDNet with a switching mechanism achieves a 36.65% higher F1-score than conventional time-series classification networks and generates false-positive predictions lower than 0.02% even in unpredictable scenarios.

Automated summary

This study introduces a reliable building fire detection framework based on simulation learning, aiming to mitigate data scarcity issues by simulating fire events, ultimately achieving a higher F1 score and lower false positive rate compared to conventional time-series classification networks.