Automatic Phase Identification of Earthquake Based on the UBDN Deep Network

Abstract

\n Seismic data obtained from seismic stations are the major source of the information used to forecast earthquakes. With the growth in the number of seismic stations, the size of the dataset has also increased. Traditionally, STA/LTA and AIC method have been applied to process seismic data. However, the enormous size of the dataset reduces accuracy and increases the rate of missed detection of the P and S wave phase when using these traditional methods. To tackle these issues, we introduce the novel U-net-Bidirectional Long-Term Memory Deep Network (UBDN) which can automatically and accurately identify the P and S wave phases from seismic data. The U-net based UBDN strongly maintains the U-net’s high accuracy in edge detection for extracting seismic phase features. Meanwhile, it also reduces the missed detection rate by applying the Bidirectional Long Short-Term Memory (Bi-LSTM) mode that processes timing signals to establish the relationship between seismic phase features. Experimental results using the Stanford University seismic dataset and data from the 2008 Wenchuan earthquake aftershock confirm that the proposed UBDN method is very accurate and has a lower rate of missed phase detection, outperforming solutions that adapt traditional methods by an order of magnitude in terms of error percentage.