← Back to Sessions

Ditingtools and Ditingbox: Seismic Big Data Processing via Edge and Cloud Computing

Description: 

In the era of seismic big data, traditional “centralized” data processing methods are increasingly facing bottlenecks in data transmission, data storage, and peak data processing capabilities. We suggest that "decentralized" data processing via edge and cloud computing will be a solution to that dilemma and would become the next-generation tool in observational seismology. To achieve this goal, two major points should be considered in practical applications: (i) sufficiently robust data processing algorithms that can handle the variety of big seismic data recorded in different regions; (ii) edge devices that can deploy these algorithms for real-time data processing on the instrument side.

Based on a large-scale Chinese seismic benchmark dataset (the so-called “DiTing” dataset) and several other big datasets such as STEAD and INSTANCE, we have developed and trained several deep learning neural networks for seismic analysis, which are collectively referred to as “DiTingTools”: the U-net architecture "DiTingPicker" for earthquake detection and P and S arrival-time picking, the Holistically-Nested Edge Detection architecture "DiTingMotion" for further P first polarity determination, and the Convolutional Network "DiTingAzi" for single station azimuth estimation. These models have good generalization ability, and more importantly, they can be deployed on edge devices with limited computing resources. We also designed an edge device named "DiTingBox" to deploy “DiTingTools”, which can predict earthquake phases, azimuths, and magnitudes in less than 5 seconds after receiving an earthquake signal with a power consumption of only 1 watt. Finally, we can build a big data processing system by combining “DiTingBox” with cloud computing, which gathers outputs from individual stations and conducts multiple-station analysis on the cloud including phase association, earthquake location, and focal mechanism inversion. This system could be used in rapid earthquake cataloging, earthquake nowcasting, as well as aftershock monitoring.

Session: Opportunities and Challenges for Machine Learning Applications in Seismology

Type: Oral

Date: 4/19/2023

Presentation Time: 11:15 AM (local time)

Presenting Author: Miao Zhang

Student Presenter: No

Invited Presentation: 

Authors