Trustworthy Building Fire Detection Framework With Simulation-Based Learning

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