AI Earthquake Tracker Is Inspired by Speech Recognition Technology
A Washington DC-based company has developed an artificial neural network that can detect and locate even very minor tremors.
The state of Oklahoma has witnessed a stunning rise in the frequency of earthquakes, which has been linked to an increase in the use of fracking technology in the oil and gas sector. Starting in 2009, the annual number of quakes measuring above magnitude 3.0 in the state exploded from fewer than three to as many as 903 in 2015.
Now, all this seismic activity has prompted scientists to develop a new tool for tracking it — drawing on speech recognition technology.
The result is a system dubbed ConvNetQuake that’s designed to detect even the tiniest earthquake against background geological noise in the same way that a smartphone can discern a human voice inside a car that’s rumbling down the highway.
The system represents an upgrade in sensitivity and detection-speed from current methods, according to its designers. When tested against historical field data, the new approach uncovered 17 times more quakes than were recorded in the Oklahoma Geological Survey standard earthquake catalog.
“We’ve trained the algorithm to understand what’s just noise and what’s an earthquake, and also where the earthquake is coming from,” Thibaut Perol, lead author of a new paper describing the system, told Seeker. Perol works on voice-recognition and artificial intelligence at a startup in Washington DC called *gramLabs.
Fracking is a relatively new form of crude oil and natural gas production that’s dramatically revived US hydrocarbon output.
The process involves blasting chemical-laced water below ground to fracture rock formation and withdraw oil or natural gas, opening up previously inaccessible reserves. But excess water is seeping out into dormant faults, and is thought to be causing them to slip, resulting in earthquakes.
Most existing earthquake-detection methods are designed to detect moderate-to-large events. As a consequence, they miss many low-magnitude earthquakes that get masked by background seismic noise.
But picking up the smaller quakes allows researchers to paint a more precise picture of all the earthquake activity in a place like Oklahoma, yielding a better understanding of the location of the quakes, whether they might be shifting, and whether the frequency is rising or falling. The extra data could eventually yield insight into whether a big one is coming, Perol said.
That’s because the art of predicting earthquakes remains essentially one of modeling likely future risk based on the patterns that have come before. In spite of some promising new research in the field of earthquake forecasting, the state-of-the-art is still limited, essentially, to an understanding of how many quakes have come before, and how often.
Existing platforms for detecting earthquakes use three stations to triangulate the source of the rumbling. The new method isn’t just more sensitive, but requires only one detection location.
To develop their system, Perol and his colleagues used a machine learning technique that funneled in both signal data and background noise then tweaked the algorithm until it could tell the difference. The data from real, recorded earthquakes were seeded into the system against backdrops of both real, field-recorded underground geological noise and synthetic blasts of artificially generated fuzz.
Using that method, “we trained the algorithm to detect, in real time, whether it’s noise or an earthquake,” Perol said.