An Automated Station Assessment Based on Deep Learning

Date: 4/24/2019

Time: 06:00 PM

Room: Grand Ballroom

As the number of seismic stations increasing, it helps to improve the completeness of earthquake catalogue and the understanding of seismic hazard. However, it has become more difficult to ensure the quality of seismic data and, thus, equally important to develop a method for allowing only good data to be used in real-time earthquake monitoring system, especially in earthquake early warning system. Therefore, an automated method for quality control of seismic data is required. In this study, we proposed a new method for seismic station performance assessment based on deep learning.

This study used the power spectral density (PSD) of seismogram for the assessment, which has been used for investigating characteristics of background seismic noise over a wide range of frequencies. The shape of the PSD also varies depending on the condition of a station, which can be used for training data of deep learning-based pattern recognition. We used 2 years of PSDs of the Korea Meteorological Administration as our training data, which were estimated every 30 minutes from seismograms of broadband seismic stations. The PSDs were manually classified into three categories: quiet or normal condition for background seismic noise, bad or abnormal condition for system transient or instrumental glitch, and event condition for local or teleseismic earthquake. The total number of data is 8298; 2922 PSDs for good condition, 3000 PSDs for bad condition, and 2376 PSDs for event condition.

Supervised learning model was implemented by the convolutional neural network, which has produced extremely promising results for various fields such as image recognition, natural language understanding, and even recently seismic phase detection. The training was performed with 70 % of data and the rest was used for validation. It is found that the precision and recall are very high (more than 99 %), indicating that most of PDSs were classified very well. Therefore, it is expected that this can be used for a real time station assessment tool for reporting of quality issues to network operators.

Presenting Author: Dong-Hoon Sheen

Authors