Real-Time Earthquake Detection and Phase Picking Using Temporal Convolutional Networks

Date: 4/24/2019

Time: 09:00 AM

Room: Elliott Bay

Real-time earthquake monitoring has high requirements on both accuracy and speed. The STA/LTA method for phase arrival identification is widely used by seismic networks due to its low computational cost, but it suffers from both missed events and false predictions. Recently, deep-learning-based methods have achieved significantly higher accuracies for earthquake detection. Although these algorithms perform much faster than search-based methods, they are not still fast enough for real-time applications. Sliding windows are needed for basic convolutional neural networks to achieve real-time prediction. This will introduce extensive computations due to the overlap among sliding windows, and making up for this by using a short sliding window will decrease neural networks’ performance. Moreover, basic convolutional neural networks do not focus on the edges of the window, which reduces sensitivity to new events emerging from the front end of the time series. In this work, we build a deep neural network for real-time earthquake detection and phase picking using a dilated causal convolutional architecture. The network maps a continuous seismic waveform to a sequence of three classes: P, S, and noise. The network is designed to make real-time prediction for each new data point from a real-time data stream, making it hundreds of times faster than basic convolutional neural networks. The network learns long-range temporal dependencies and features based on a large receptive field, which achieves higher prediction accuracy than STA/LTA method. Our network has a good balance of speed and accuracy and can be used in real-time earthquake monitoring.

Presenting Author: Weiqiang Zhu

Authors