The best listening experience is on Chrome, Firefox or Safari. Subscribe to Federal Drive’s daily audio interviews Apple Podcasts or PodcastOne.
Since there are fires, firefighters have to worry about the dangerous phenomenon known as flash drives. Researchers at the National Institute of Standards and Technology have now developed an artificial intelligence tool that promises to help predict flash drives. NIST mechanical engineer Andy Tam talks about more details Federal driving with Tom Temin.
Tom Tem: Dr. Tam, it’s good that you’re at work.
Andy Tam: Thanks, Tom. I’m glad to be here.
Tom Tem: And let’s talk about the flash itself. What exactly is in the first place?
Andy Tam: This particular term may not be familiar to too many of us. So, maybe, let me take a minute or so to explain what flash is and why it’s so dangerous. So flashover is an extreme fire event. When this happens, almost all directly exposed combustible materials, such as [a] sofa, mattress, wooden cabinet, even carpets in the compartment as [a] living room, bedroom or even kitchen can be lit at the same time. Therefore, when everything burns at once, the temperature of the gas in this compartment will rise dramatically in a very short period of time. And the temperature usually exceeds 100 degrees C, so in such a high temperature environment, even when firefighters are fully equipped, survival will be very low. And unfortunately, this type of rapid fire progression, such as a flash drive, is the leading cause of death and injury to firefighters. So we hope that if we can do something, then we can save many lives.
Tom Tem: And it is difficult to observe conditions that lead to change. And so you have developed an algorithm or prediction model. Tell us what you have included in this model, tell us how it works.
Andy Tam: So our focus is to develop a model that can actually be used in real life and we will find scenarios. The idea behind it is to use information from fire extinguishers, such as temperature readings from heat detectors in the building. However, there are two main challenges to this type of realistic application. The first challenge is that there are no perfect sensors or thermal detectors to withstand this type of high temperature environment. And usually the heat detector would probably stop working without giving reliable temperature signals at about 150 degrees C. Again, we must keep in mind that the potential flash usually occurs when the temperature of the gas in the upper layer reaches about 600 degrees C. So this is a challenge that we have to overcome. And the second challenge is that we don’t actually have any information about the location of the fire, what object is burning or how big the fire is, and it’s also very important that we don’t know any information about the state of opening doors and windows for buildings. However, we believe that if our model is well trained, it should be able to recognize the flash based on the limited temperature information from the heat detector. Having said that, then the next challenge will be – we need to find a way to get the data.
Tom Tem: In other words, if the heat detector has some increase before it stops, and you extrapolate from this phenomenon, it may give the impression that it will get really hot and cause a flash.
Andy Tam: Right.
Tom Tem: Okay. And I guess my question is, how many rooms and homes usually have heat detectors in the first place?
Andy Tam: So let me focus on another aspect of it, because every home will be different. But when we try to develop a machine-based model, it is well known that we need a large amount of relevant data. Before our case, this meant that we had to burn a building structure many, many times. And if we think about it, conducting hundreds or even thousands of this kind of large-scale experiment will be relatively difficult. And that’s because it will be very time consuming and expensive. Also, there is no guarantee that the data we receive is 100% useful. So to overcome this problem with the lack of data, we use what we call learning through a synthesis approach. And that means we’re going to do the experiment on a full scale, virtually using a NIST fire model called CFAST. For this particular model development, we choose a popular single-family building structure. And we conduct a large number of digital experiments with a wide range of fire conditions from burning a wooden chair to burning a full-size mattress. In our experiment, we also look at the conditions of different opening doors and windows to try to mimic various fire events that occur in real life, such as a door open due to evacuation and a broken window due to a high temperature difference. But one thing I would like to point out here is that we have carried out more than 20,000 different fire cases. But we only have about 5,000 instances of flash that we can use to train our model. So, if we had to physically conduct the experiment ourselves, it would be a very difficult job. But we were very happy to find a way to overcome it and get synthetic data for the development of our model.
Tom Tem: Yes, I guess it’s better than burning an entire neighborhood to check it out. We’re talking to Dr. Andy Tam. He is a mechanical engineer in the fire research department at the National Institute of Standards and Technology. Okay, so the model has been tested with real and simulated fire data. How does this work operationally? I’m a fire platoon. I go to a three-bedroom burning ranch. How do I know that a flash show can happen?
Andy Tam: So in the first place, this model is still proof of the concept stage. However, we expect such questions. So when we try to develop our models, one question we will always try to ask ourselves is that since our model is developed based on synthetic temperature data, can it be believed that it will be used in real fire scenarios. So to address this concern, we tested our model against full-scale experimental data obtained from the laboratories of insurers with the same building structure. Here I would like to emphasize that our model never sees any of these experimental data during the training process. Yes [it] is a very good reality check for our model and see if the model will actually work in real fire conditions. The results are very promising and we have an overall accuracy of about 85%. However, they are still additional work that we need to do to improve the performance of the model. So in the coming years we will conduct our own large-scale experiments with the integrations of our model in the fire alarm control unit and test it against a real fire.
Tom Tem: So you’re going to build a few houses and burn them down?
Andy Tam: Well, first it will take time to build the house.
Tom Tem: Yes, but that sounds like the fun part of the whole experiment. Ultimately, it would be an application that could exist on the radio or say a smart device that a firefighter would carry with him on fire.
Andy Tam: Yes, we anticipate that this model will be implemented in a building structure with wireless fire extinguishers. In a fire scenario, we believe that fire data, such as temperature, can be data from, say, thermal detectors. And this information can then be processed locally to provide valid information to our firefighters, such as when or where there will be a change in the building structure. We believe that if firefighters can have this type of information before they arrive at the scene of the fire, then they can plan in advance for the fire attack as well as search and rescue strategies. The difficult part, however, is that the mechanism that will be used to transmit the data to the firefighter has not yet been resolved. Overall, research is still ongoing. But then the emphasis will be that the system itself must be accessible, it cannot be in the order of thousands of dollars, and so on. In addition, it should only use existing commercial devices. And that’s because we believe that wireless technology is advancing so fast. And we are afraid that we are very adapted to it. For example, I can control my thermostat away from home. So we believe that reliable data transmission can be achieved in the near future. And this is the direction our research is heading, and we hope that what we are doing in our NIST group can help raise awareness of the situation, as well as fire safety for our firefighters.
Tom Tem: And just out of curiosity, how did you, as a mechanical engineer, get involved in this firefighting and fire prevention deal?
Andy Tam: This is a very good question. And that’s because if we look back, this problem is one of the main problems associated with deaths and fire injuries. I happen to have experience, and I also did my research using a neural network. And then, to address this problem or this problem in real time, traditional methods may not be very applicable, because performing this type of flight simulation will take a long time. And the use of machine learning will help us overcome this numerical difficulty, which means that we did a lot of training, getting all the data in advance and then inventing a neural network model, and when we use it, we just call neural network architecture or mathematical formulations. , then usability will be very easy.
Tom Tem: Dr. Andy Tam is a mechanical engineer in the Department of Fire Research at the National Institute of Standards and Technology. Thank you so much for joining me.
Andy Tam: Thank you very much for accepting me.