Artificial intelligence: Deep learning to save firefighters' lives

(Hier finden Sie die deutsche Version des Beitrags)

In the chaos of a burning building, it is extremely difficult to recognise the signs of an impending "flashover". Because if the heat of the fire from individual burning objects, the pyrolysis gases that form and other objects heat up far enough, the fire can spread abruptly - which repeatedly leads to the death of firefighters.

With the help of computer models, experts investigate the dynamics of such fires to better assess critical situations in advance. However, physical fire simulation requires a lot of computing time and capacity. For several years, researchers worldwide have therefore been working on software that uses machine learning methods to warn firefighters of critical developments during operations.

Researchers at JPL, for example, presented an approach in 2018 in which they use so-called GANs (Generative Adversial Networks) to enhance the images from firefighters' helmet cameras- this allows the smoke to be extracted from the images. They then generated the actual prediction of a possible flashover based on the enhanced images of the fire. Researchers at the National Institute of Standards and Technology (NIST) used machine learning to reconstruct data from failed temperature sensors. And Tianhang Zhang from Hong Kong Polytechnic University developed a method in 2021 that uses LSTM (Long Short Term Memory) networks to predict the time course of fires - but only for specific conditions at a time.

When does the fire leap over?

Now Wai Cheong Tam, also from NIST, and colleagues from Hong Kong Polytechnic University, China University of Petroleum and Google have developed software that can predict a flashover with up to 30 seconds' warning and 90 per cent accuracy for a variety of floor plans and boundary conditions.

In order to do justice to the variability of real fires, the researchers use neural networks based on graphs, so-called GNN. GNNs represent the available data - in this case, temperature curves in different rooms - in the form of mathematical graphs, i.e. structures that can represent relationships between the data (temperature curves from non-adjacent rooms, for example, are not directly linked). The GNN learns from examples how individual fire sources in a flat or house influence each other. In a second part of the model, the researchers then use convolutional neural networks to predict the temporal course of the temperature at individual sensors - and thus also an impending flashover.

However, the software has so far only been trained and tested with data from simulated fires. The researchers simulated more than 41,000 fires in 17 building types representing much of the typical housing stock in the US. In addition to the floor plan, they also varied factors such as the origin of the fire, the type of furniture and whether doors and windows were open or closed. The accuracy of the model - 92.1 per cent at best with 30 seconds of lead time - outperformed five other machine learning-based tools, including the authors' previous model. As a next step, they plan to test the model with real rather than simulated data.

(pavb)