A deep learning approach for suppressing noise in livestream earthquake data from a large seismic network

Abstract

SUMMARY Detecting and analyzing small earthquakes is important for many seismological studies. Signals of small earthquakes are often obscured by noise. Recent advances in signal processing and deep learning along with available computing resources provide a great opportunity to address this challenge. In this study, we present a time domain method of suppressing noise for processing livestream earthquake data from a large seismic network by applying a deep neural network Real-time Denoiser (RTDenoiser). This neural network is able to attenuate a variety of colored noise and non-earthquake signals and suppress noise in the overlapping frequency bandwidth with signals. Because of its simplicity in time domain without domain transformation and subsequent processing, the method is able to process continuous livestream three-component data from several hundreds of seismic stations simultaneously. We create ‘noise-free’ samples by scaling down the waveforms of relatively large events from ML 3.5–5.0 to ML 1.5–3.0 according to the Richter scaling relationship. We also select noise samples from the same seismic station and add to ‘noise-free’ data to generate samples at different signal-to-noise ratio (SNR) levels. These data samples are randomly split into training, validation, and test sets. We verify the trained network to process data recorded in Sichuan and Yunnan, China from 2013 to 2018. Results show that the RTDenoiser can help improve SNRs from 5 dB to 15 dB in averag. The number of detected small events at magnitude between ML 1.0 and 3.0 has been increased by 58.8 per cent. The method is currently applied in a seismic network of 300 stations in Sichuan and Yunnan, China for continuous processing. It takes about 10 ms on average to process three-component 60-s data from 300 seismic stations on a single GPU.