AI may help protect firefighters

A ‘flashover’ is a nightmare firefighting scenario in which all combustible material in a room suddenly ignites, due to extremely high local temperatures. As one of the leading causes of firefighter deaths, any forewarning that can be given to firefighting teams will be extremely valuable.

New research suggests that artificial intelligence (AI) may help provide this, with the aid of the Flashover Prediction Neural Network (FlashNet) predictive model.

Published in Engineering Applications of Artificial Intelligence, a study by a groups of researchers from the National Institute of Standards and Technology (NIST), Hong Kong Polytechnic University and other institutions has shown FlashNet to outperform existing AI-based flashover forecasting tools, thanks to its accuracy of up to 92.1 percent across more than a dozen 'typical' residential floorplans in the US.

Machine learning-based prediction tools have been trained in a single, familiar environment, unlike firefighters who operate in diverse and complex scenarios on a regular basis. Wai Cheong Tam, co-first author of the new study, said that moving from a four or five-room model to 13 or 14 rooms “can be a nightmare for the model."

"For real-world application, we believe the key is to move to a generalised model that works for many different buildings," he said.

To cope with the variability of real fires, the researchers used graph neural networks (GNN), an ML algorithm that uses graphs of nodes and lines, to represent different data points and their relationships with one another.

The researchers simulated over 41,000 fires in 17 kinds of buildings that aimed to represent most common layouts of US residential buildings. Other factors included the origin of the fire, types of furniture and whether doors and windows were open or closed throughout. The model was trained in an initial data set of approximately 25,000 fire cases, with a further 16,000 used to fine-tuning and testing.

The model provided 92.1 percent predictive accuracy and outperformed five other machine-learning-based tools. Significantly, the tool produced the least false negatives, dangerous cases where the models fail to predict an imminent flashover.

Proving the model’s theoretical utility is not the end, however, as real-world testing still needs to be done.

"In order to fully test our model's performance, we actually need to build and burn our own structures and include some real sensors in them," Tam said. "At the end of the day, that's a must if we want to deploy this model in real fire scenarios.