SEDENOSS: SEparating and DENOising Seismic Signals With Dual‐Path Recurrent Neural Network Architecture

Abstract

Seismologists have to deal with overlapping and noisy signals. Techniques such as source separation can be used to solve this problem. Over the past few decades, signal processing techniques used for source separation have advanced significantly for multi‐station settings. But not so many options are available when it comes to single‐station data. Using Machine Learning, we demonstrate the possibility of separating signals for single‐station, one‐component seismic recordings. The technique that we use for seismic signal separation is based on a dual‐path recurrent neural network which is applied directly to the time domain data. Such source separation may find applications in most tasks of seismology, including earthquake analysis, aftershocks, nuclear verification, seismo‐acoustics, and ambient‐noise tomography. We train the network on seismic data from STanford EArthquake Dataset and demonstrate that our approach is (a) capable of denoising seismic data and (b) capable of separating two earthquake signals from one another. In this work, we show that Machine Learning is useful for earthquake‐induced source separation. We provide a reproducible research repository with the algorithms here: https://github.com/crimeacs/source-separation.

EdgePhase: A Deep Learning Model for Multi‐Station Seismic Phase Picking

“…Based on the assumption that a similar pattern of the earthquake signal is shared at neighboring stations, the Edge Convolutional module can be modified to transfer the feature embeddings relevant to the earthquake signals. One can adopt encoders, decoders, and loss functions in the previous single‐station seismograms denoising models (Novoselov et al., 2020; W. Zhu, Peng, et al., 2019) and fine‐tune them as is done in this study.…”

Recent advances in earthquake seismology using machine learning

“…The denoising of seismic waveform records is related to the issue discussed in this section. Several studies have been conducted to extract the target signal from originals containing waves from various sources or separate signals and noise (Zhu et al 2019b;Tibi et al 2021Tibi et al , 2022Dalai et al 2021;Novoselov et al 2022;Yin et al 2022a;Xu et al 2022;Wang and Zhang 2023). These efforts will lead to the effective use of observation data unanalyzed due to a low signal-to-noise ratio.…”

Improved Observations of Deep Earthquake Ruptures Using Machine Learning

“…A skip connection is introduced after the first CNN layer to retain the fine scale of the feature. Compared to the single‐branch prediction of either the earthquake or noise signal (Novoselov et al., 2022; Zhu et al., 2019), our multi‐task model (i.e., two‐branch prediction) depends on the efficiency of feature extraction for both earthquake and noise signals.…”

“…In fact, the time-frequency information may also be utilized implicitly by appropriate convolutional layers considered multi-frequency-band "filters" in the time domain. Using that concept, Novoselov et al (2022) showed that recurrent neural networks could separate overlapping seismic signals produced by distinct sources. Yin et al (2022) combined two-branch encoder-decoder and recurrent neural networks to compose the WaveDecompNet, which has been proven effective in reconstructing local earthquake and noise waveforms.…”

“…Using that concept, Novoselov et al. (2022) showed that recurrent neural networks could separate overlapping seismic signals produced by distinct sources. Yin et al.…”