Building Earthquake Early Warning Networks With Low Cost, Off-the-Shelf Components

Ryan Logsdon, Robert L. Walker, & Sean Gibbons

Submitted August 15, 2018, SCEC Contribution #8815, 2018 SCEC Annual Meeting Poster #052

We seek to demonstrate the ability to use minimal cost, off-the-shelf components to create disparate networks for earthquake detection. These networks seek to draw a balance between fully crowd-sourced data gathering, such as that generated through smartphones, and dedicated, scientific operations, as exemplified by conventional seismometer networks. The resulting system is intended to provide at least a significant fraction of the continuity afforded by traditional seismometer networks, at a fraction of the price, and orders of magnitude increase in the number of streaming data sources.

Individual sensors are built around open standards and previous projects that have demonstrated an optimal balance between data quality, flexibility of use, and cost. While our individual "stations" are designed in a manner as to minimize cost, they are similarly intended to allow for easy expansion with more advanced and robust tools as they become available, allowing for an agile development lifecycle. At present, data gathering is focused around MEMS accelerometers which are ubiquitous in today’s hardware designs, especially in the Internet of Things (IOT) space. Utilizing these sensors comes with the drawback of having much less reliable input data. However, since drift and noise filtering are largely well-understood problems, these drawbacks are easily combated. The main benefits of using IOT components are that we can warn the public without reliance on telephony, and the cost to produce our seismic sensors is so low, whole metropolises such as Los Angeles and San Diego can be fully protected.

Data gathered is treated initially using standard signal processing and seismological techniques, allowing for meaningful detections of P and S wave arrivals from what would otherwise be largely noisy and unreliable data. By employing a novel recurrent neural network (RNN), we may further analyze the live data by assigning each seismic sensor ephemeral probabilities of accuracy and signal confidence. The resultant data is used to identify and locate earthquakes in real time, allowing us to warn members of the public prior to the arrival of damaging effects in their geographic location. The learning model begins as an isotropic velocity model. It is augmented with a localized model of peak ground acceleration. It is then updated aposteriori with the velocities measured by our sensors during times of seismic activity.

Key Words
earthquake early warning, ots, off-the-shelf, neural networks, rnn

Logsdon, R., Walker, R. L., & Gibbons, S. (2018, 08). Building Earthquake Early Warning Networks With Low Cost, Off-the-Shelf Components. Poster Presentation at 2018 SCEC Annual Meeting.

Related Projects & Working Groups
Earthquake Forecasting and Predictability (EFP)