A new study has found that computers can be trained to better detect distant nuclear explosions, chemical explosions and volcanic eruptions by learning from man-made explosion signals.
The study was published in the journal ‘Geophysical Research Letters.’
Witsil, at the Institute of Geophysics’ Alaska Wilson Technical Center, and colleagues created a library of synthetic infrasonic burst signals to train computers to recognize the source of the infrasound signal. Infrared waves are at frequencies too low for humans to hear and travel farther than high-frequency sound waves.
“We used modeling software to generate 28,000 synthetic ultrasonic signals that, although generated in a computer, can be recorded using ultrasonic microphones deployed hundreds of thousands of times,” says Witsil. kilometers from a big bang”.
The man-made signals reflect changes in atmospheric conditions, which can alter the signal of a regional or global explosion as the sound waves travel. Such changes can make it difficult to detect the source and type of an explosion from a long distance.
Why make artificial sounds of explosions rather than using real world examples? Since explosions don’t happen at every location on the planet, and the atmosphere is constantly changing, there aren’t enough real-world examples to train generalized machine learning detection algorithms.
“We decided to use synthetics because we can model several different types of atmospheres through which signals can propagate,” says Witsil. “So even though we don’t have access to any of the explosions that happened in North Carolina, for example, I can use my computer to model the North Carolina explosions and build build a machine learning algorithm to detect the explosion signal there.”
Today, detection algorithms often rely on infrasound arrays consisting of multiple microphones in close proximity. For example, the International Comprehensive Test Ban Treaty Organization, which monitors nuclear explosions, has deployed infrasound arrays worldwide.
“That’s expensive, it’s hard to maintain, and much more can break,” says Witsil. Witsil’s method improves detection using hundreds of single-element ultrasonic microphones already available around the world. That makes detection more cost-effective.
Machine learning approaches extend the usefulness of single-element ultrasonic microphones by making them capable of detecting more sophisticated burst signals in near-real time. Single-part microphones are currently only useful for retroactive analysis of known and often high-amplitude signals, as they did with the massive January eruption of the Tonga volcano.
Witsil’s method can be deployed in an operational environment to protect the nation or mitigate natural hazards.