Seismic processing via machine learning for event detection and phase picking

Machine learning;Seismic;Phase picking;CNN

Zhu, Lijun ; McClellan, James H. Electrical and Computer Engineering AlRegib, Ghassan Peng, Zhigang Anderson, David V. Newman, Andrew V. ; McClellan, James H.

University:Georgia Institute of Technology

Department:Electrical and Computer Engineering

关键词: Machine learning;    Seismic;    Phase picking;    CNN;   

Others  :  https://smartech.gatech.edu/bitstream/1853/61770/1/ZHU-DISSERTATION-2019.pdf

美国|英语

来源: SMARTech Repository

【 摘 要 】

A feasible solution for seismic event detection and phase picking is prototyped on an embedded system with seismic sensors using a lightweight convolutional neural network (CNN). Event detection and phase picking are essential for locating seismic events and imaging subsurface structures. Real-time detection and picking using an embedded system with seismic sensors are not only valuable for time-sensitive tasks, e.g., earthquake early warning (EEW) systems but also provide the foundation of an interconnected smart sensor network, e.g., internet of things (IoT), for modern seismic acquisition and monitoring. However, existing detection and picking methods are either too simple to achieve the desired accuracy or too computationally complex to be deployed on an embedded system along with the seismic sensors. Accurate offline training of the CNN is demonstrated for small and large datasets, while observing the trade-off between accuracy and computation cost, which gives an efficient online deployment of the neural network. When trained for a small dataset from one area, transfer learning is used to modify the CNN parameters with modest retraining in order to generalize and validate the CNN model for processing in other regions. Simplification of the model and quantization of its parameters is explored to develop a prototype that is suitable for embedded devices. The product of this research is a universal seismic event detection and phase picking tool that is accurate and efficient for processing large volumes of data, as well as lightweight for deployment on an embedded system.