| Work Detail |
Machine learning techniques have been trained to recognise the characteristic sound of safety valves opening, providing an early warning system for battery fires.
Researchers at the National Institute of Standards and Technology (NIST) have developed an algorithm capable of detecting the sound of a lithium-ion battery cell overheating. This technology provides a critical early warning system against catastrophic battery fires.
Using a convolutional neural network, NIST researchers created a computer program capable of identifying with 94 percent accuracy the characteristic noise of an overheating lithium-ion battery cell.
The research builds on previous studies of “venting acoustics” in lithium-ion batteries, which identified the sound of a safety valve breaking with 92% accuracy. This is a “click” sound emitted when a battery’s safety valve breaks, typically about two minutes before ignition. The solar industry has also studied sound analysis for early detection of inverter faults.
For the study, “ Development of a Robust Early-Stage Thermal Runaway Detection Model for Lithium-ion Batteries,” NIST researchers collaborated with the Xian University of Science and Technology in China to conduct 38 “thermal abuse tests” on single-cell lithium-ion batteries.
These experiments simulated extreme conditions, such as overheating, to observe the progression of thermal runaway events. The study focused on 18,650 lithium-ion batteries, commonly found in laptops and electric vehicles. Each battery tested had a graphite anode, a LiNiCoAlO2 cathode, a nominal capacity of 3.2 Ah, and a voltage of 3.7 V.
The team meticulously recorded the sounds generated during each thermal runaway event, focusing on the characteristic “click-hiss” sound of the safety valve breaking. By studying the sound characteristics of the 38 recorded events, the researchers identified unique acoustic patterns that reliably indicate an impending thermal runaway.
To ensure the algorithm could distinguish these sounds from background noise, the team applied advanced machine learning techniques. They trained the convolutional neural network using a diverse dataset that included recorded drum sounds and upscaled samples. By adjusting the pitch and speed of the recordings, they created more than 1,000 unique audio samples to simulate various real-world scenarios.
“We tried to confuse the algorithm with all kinds of noises, from recordings of people walking to doors closing or Coca-Cola cans being opened,” explains Dr. Wai Cheong “Andy” Tam, one of the lead researchers. Despite these difficulties, the algorithm was able to identify the sound of a safety valve breaking at 94% accuracy, making it a promising candidate for integration into early warning systems.
In the image above, the acoustic signals from Experiment 30 show how the algorithm distinguishes the different sounds of the critical event of a safety valve breaking. The top graph (3a) shows several noises: whispering, flipping a light switch, moving the camera, closing a door, using a hammer, and turning on a battery. The safety valve breaking is characterized by its high amplitude and long oscillation, which distinguish it from the sharp, singular spikes of a switch and the more complex oscillations of a hammer blow.
The researchers have applied for a patent on their device. They are also studying improvements, such as testing with different types of batteries and incorporating alternative microphones to increase the systems precision and versatility.
The need for advanced battery fire detection systems is becoming more urgent, especially after high-profile incidents such as the battery fire in Arizona in 2019. In that incident, a shipping container full of batteries exploded, injuring eight firefighters and a police officer. The blast knocked the entire team unconscious, and one firefighter was thrown more than 20 meters by the force of the explosion.
One safety innovation developed in response to such incidents is remote “outgassing” sensing, which allows gases to be vented from battery storage containers before they can ignite. In the NIST experiments, the outgassing process is visible on video after the safety valve ruptures. This buildup of gases was the primary cause of the Arizona explosion when the container was opened. |
