Safe and Sound: NIST’s AI-Based Fire Prevention Tech Hears Li-ion Battery Failures Begin

Researchers at the National Institute of Standards and Technology (NIST), led by Wai Cheong “Andy” Tam and Anthony Putorti, have developed a novel AI-powered method for detecting imminent lithium-ion battery fires. By training artificial intelligence to recognize the distinct "click-hiss" sound that occurs when gas escapes through a safety valve in a failing battery, the NIST team aims to provide an early alert in environments where traditional smoke alarms may be too slow to respond.

Growing need for battery safety solutions

Lithium-ion batteries, prized for their high energy density, power countless devices and vehicles, from smartphones to electric cars. But this same energy density poses serious risks, especially in cases of thermal runaway, where overheated or damaged batteries can ignite almost instantly. The challenge is growing more urgent, as demonstrated by the 268 e-bike battery fires in New York City in 2023, resulting in 150 injuries and 18 fatalities.

Traditional fire detection methods are often too slow for lithium-ion incidents, which can emit intense jets of flame at up to 1100°C (2012°F). In contrast to smoldering fires that produce smoke detectable by conventional alarms, lithium-ion batteries tend to release minimal smoke during early stages of failure, making rapid detection critical.

Sound as an early warning signal

The distinctive "click-hiss" sound occurs as pressure builds within a battery’s casing due to chemical reactions. If the battery’s hard casing includes a safety valve, this valve will break to release pressure, producing an audible noise just before ignition. Inspired by observations of this pattern, Tam and his team sought to test and refine sound-based detection as a reliable early-warning method.

To develop the detection system, NIST collaborated with Xi’an University of Science and Technology to record sounds from 38 battery explosions. These recordings, augmented to create over 1,000 unique samples, allowed NIST researchers to train an AI algorithm specifically to recognize this telltale sound. The algorithm has demonstrated 94% accuracy in identifying the “click-hiss” even in noisy environments, such as footsteps or door sounds, which commonly occur in battery-rich settings like warehouses or garages.

Future of AI-enhanced battery fire detection

Tam presented the findings at the 13th Asia-Oceania Symposium on Fire Science and Technology. In tests, the AI system detected the failing battery’s sound approximately two minutes before a catastrophic failure occurred, providing critical time for evacuation or mitigation. With plans to expand the technology to other types of batteries and to further validate the system’s response time, Tam and Putorti are optimistic that AI-enhanced detection could soon be incorporated into specialized fire alarms.

Once fully developed, this system could be deployed across battery-dense facilities, including homes, office buildings, warehouses, and electric vehicle garages, where early warning systems would provide essential protection and peace of mind for those living and working around lithium-ion batteries.