AI predicts fire "flashover" within 30 seconds with an accuracy of 92.1%, which may help save firefighters' lives in the future

During this period, the weather is extremely hot and dry, and it is also a time when fires are most likely to occur.

There have been some fire incidents one after another recently, and the news of the sacrifice of firefighters is heartbreaking. The greatest threat to firefighters in fires is actually deflagration. In just two days, there were two accidents in which firefighters died due to deflagration.

This kind of deflagration phenomenon is often caused by the burning of indoor fires inside the building, which fills every room in the room. Deflagration occurs when combustible gases and incompletely burned gases reach a certain concentration.

At this time, the combustibles in other rooms in the room that were not exposed to the fire were also ignited and burned. That is, with a "boom", all the combustibles in the room were ignited, so the entire The process is also called "flashover."

It is precisely because of the unpredictability of "flashover" that it is so dangerous. ​

This week, a study leveraged graph neural networks (GNN) to build a system to learn relationships between different data sources (represented as nodes and edges) in simulated fires , thereby predicting in advance whether the "flashover" phenomenon will occur in the next 30 seconds.

This research is expected to help firefighters determine whether "flashover" will occur in indoor buildings, thereby saving lives. The paper was published in "Artificial Intelligence Engineering Applications".

Paper address: https://www.nist.gov/news-events/news/2022/08/ ai-may-come-rescue-future-firefighters

Predict "flashover" within 30 seconds, with an accuracy of 92.1%

Generally speaking, firefighters have to rely on their own experience to judge whether such a "flashover" will occur:

1. Produce radiant heat that burns human skin, and radiate after a few seconds. Thermal intensity can reach 10kw/m².

2. The hot air flow in the room makes people unable to stand, and the convection temperature in the room is close to 450℃.

3. The door is so hot that the temperature of the wooden part exceeds 320℃ on average.

4. The flames jumping out from the door almost reached the ceiling, and a large amount of radiant heat was reflected from the ceiling to the combustible materials in the room.

5. The smoke drops to about 1m above the ground, and the thermal layer in the air occupies the upper air, driving the thermal decomposition products to fall.

To better help firefighters predict "flashover", researchers collected a variety of data, ranging from building layout, surface materials, fire conditions, and ventilation configurations , the location of smoke detectors, and the temperature distribution in the room, 41,000 virtual fires in 17 different building types were simulated. A total of 25,000 fire cases were used to train the model, and the remaining 16,000 cases were used for fine-tuning and testing.

Across 17 different homes, the new model's accuracy depended on the amount of data it needed to process and the lead time it sought to provide firefighters.

In the end, the model’s accuracy (best 92.1% at 30 seconds ahead) surpassed five other machine learning-based tools, including the project team’s own Importantly, the tool produced the fewest false negatives, i.e., dangerous situations where the model fails to predict "flashover."

This model, called FlashNet, puts FlashNet into scenarios where FlashNet has no prior knowledge of the specific conditions of the building and the fire conditions inside the building. , which is similar to what firefighters often encounter.

"Considering these limitations, the performance of this tool is quite promising," said Tam, the author of the paper. However, the author still has a long way to go before leading FlashNet across the finish line. As a next step, they plan to field-test the model with real-world data rather than simulated data.

From 4 to 5 rooms to more than a dozen rooms, the prediction difficulty is Max

Flashover general tendency Yu suddenly erupts at about 600 degrees Celsius (1,100 degrees Fahrenheit), which can then cause temperatures to rise even further.

Previous forecasting tools have either relied on a constant temperature data stream from a burning building or used machine learning to fill in data that could be lost when heat detectors are affected by high temperatures.

To date, most machine learning-based prediction tools, including one previously developed by the authors, have been trained to operate in a single, familiar environment. But in reality, firefighters face an extremely complex environment. When they rush into a fire area, they may have no idea about the scene, the location of the fire, or whether the door is open or closed.

"Our previous model only had to consider four to five rooms in one building layout, but when the building layout switches and you have 13 to 14 rooms, it can be a nightmare for the model," Tam said , “For real-world applications, we believe the key is to build a general model that applies to many different buildings.”

GNN as a method is good at making judgments based on graphs of nodes and lines Machine learning algorithms that can represent different data points and their relationships to each other are ideally suited for such tasks.

"GNN is often used to estimate time of arrival, or ETA, in traffic, which you can analyze (with GNN) 10 to 50 different roads. It is very complicated to rationally utilize this kind of information at the same time, so we came up with the idea of ​​using GNN," said paper author Yujun Fu, a research assistant professor at the Hong Kong Polytechnic University.

In addition to the National Institute of Standards and Technology (NIST), Google and the Hong Kong Polytechnic University, China University of Petroleum also participated in this research.

Related reports:

https://www.theregister.com/2022/08/14/ai_firefighter_prediction/https://www.sciencedirect.com/science /article/abs/pii/S0952197622003220https://baike.baidu.com/item/flashover/1869756?fr=aladdin#2​

The above is the detailed content of AI predicts fire "flashover" within 30 seconds with an accuracy of 92.1%, which may help save firefighters' lives in the future. For more information, please follow other related articles on the PHP Chinese website!