AI could come to the rescue of future firefighters

In firefighting, the worst flames are the ones you don’t see coming. Amid the chaos of a burning building, it’s hard to notice the signs of an impending flashover – a deadly fire phenomenon in which nearly every combustible element in a room suddenly ignites. Flashover is a leading killer of firefighters, but new research suggests artificial intelligence (AI) could provide first responders with much-needed warning.

Researchers from the National Institute of Standards and Technology (NIST), Hong Kong Polytechnic University and other institutions have developed a Flashover Prediction Neural Network (FlashNet) model to predict fatal events seconds before they occur. burst. In a new study published in Artificial intelligence engineering applications, FlashNet has shown up to 92.1% accuracy on more than a dozen common US residential floor plans and outperforms other AI-based flashover prediction programs .

Flashovers tend to erupt suddenly at around 600 degrees Celsius (1,100 degrees Fahrenheit) and can then cause temperatures to rise further. To anticipate these events, existing research tools rely either on constant streams of temperature data from burning buildings or on machine learning to fill in missing data in the likely event that heat detectors succumb to high temperatures.

So far, most machine learning-based prediction tools, including the one the authors previously developed, have been trained to work in a single, familiar environment. In reality, firefighters do not have the right to such a luxury. When charging into hostile territory, they may know nothing of the floor plan, the location of the fire, or whether the doors are open or closed.

“Our previous model only had to accommodate four or five parts in a layout, but when the layout changes and you have 13 or 14 parts, it can be a nightmare for the model,” said the mechanical engineer of the NIST, Wai Cheong Tam, co-first. author of the new study. “For real-world application, we believe the key is to move to a generalized model that works for many different buildings.”

To deal with the variability of real fires, the researchers augmented their approach with graphical neural networks (GNN), a kind of machine learning algorithm capable of making judgments based on graphs of nodes and lines, representing different data points and their relationships to one another.

“GNNs are frequently used for Estimated Time of Arrival, or ETA, in traffic where you can analyze 10 to 50 different routes. It’s very complicated to properly use this type of information simultaneously, so it’s where we got the idea to use GNN,” said Eugene Yujun Fu, assistant research professor at Hong Kong Polytechnic University and co-first author of the study. “With the exception of our application , we look at parts instead of roads and predict flashover events instead of ETA in traffic.”

Researchers digitally simulated more than 41,000 fires in 17 building types, representing the majority of US residential building stock. In addition to the layout, factors such as the origin of the fire, the types of furniture and the opening or closing of doors and windows varied throughout. They provided the GNN model with a set of nearly 25,000 fire cases to use as study material, and then 16,000 for fine-tuning and final testing.

Across all 17 house types, the accuracy of the new model depended on how much data it had to analyze and how much time it was looking to provide to firefighters. However, the model’s accuracy – at best, 92.1% with 30 seconds of delay – outperformed five other machine learning-based tools, including the authors’ previous model. Critically, the tool produced the fewest false negatives, dangerous cases where models fail to predict an impending flashover.

The authors ran FlashNet in scenarios where it had no prior information about the specifics of a building and the fire burning inside, similar to the situation firefighters often find themselves in. Given these constraints, the tool’s performance was quite promising, Tam said. However, the authors still have a long way to go before they can get FlashNet across the finish line. As a next step, they plan to test the model with real rather than simulated data.

“In order to fully test the performance of our model, we actually need to build and etch our own structures and include real sensors in them,” Tam said. “Ultimately, it’s essential if we’re going to deploy this model in real-world fire scenarios.”

Leave a Reply

Your email address will not be published. Required fields are marked *