| Abstract | Flashover occurs in building fires whereby all the combustible materials in an enclosure begin to ignite rapidly. This detrimental phenomenon occurs because of the intense heat radiating from the hot smoke layer accumulated near the ceiling, which is one of the major causes of firefighter fatalities. Thus, forecasting flashover in real-time is of great importance to firefighters’ safety. Previous empirical and theoretical analysis of the flashover phenomenon provides criteria for predicting flashover on-set and fire size, yet they are based on the thermal sensor data in the room. It is not straightforward to transfer the flashover criteria and understanding from the previous studies in application to actual fire grounds since thermal sensor data are not readily available. As a novel non-invasive prediction technique of flashover onset, this investigation explores the potential of recognizing vision indicators of flashover, such as flame extrusion or rollover. Adopting artificial intelligence (AI) technologies for pattern recognition and feature extraction, we proposed to build an optimized convolutional neural network named “FlashoverNet”, a modern AI algorithm, to detect and forecast flashover occurrence in enclosure fire incidents. FlashoverNet was trained and validated by data from several full-size actual room fire experiments. FlashoverNet imitates the approach of flashover onset prediction by firefighters by monitoring fire growth and looking for early visible indicators in a room fire. Our investigations demonstrate that FlashoverNet can successfully detect and forecast flashover onset in real-time with more than 94% accuracy in full-scale actual room fire tests. |
|---|
