Separating and denoising seismic signals with dual-path recurrent neural
network architecture
Abstract
Separation of overlapping signals is an important task in signal
processing, with application in music, speech, and seismic signal
processing. We show that separation is possible also for seismic
recordings, using techniques from machine learning (and even those
recorded with a single sensor).
This may have an impact on seismic applications such as
ambient noise tomography, induced seismicity, earthquake analysis,
aftershock analysis, nuclear verification, and
seismoacoustics/infrasound.
The machine learning 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.
We train the network on seismic data produced by trains, and recorded
with a Raspberry Shake sensor at the University of Vienna. We
demonstrate that the network predicts the signals from a synthetic
mixture very well.
We then use a transfer learning approach to fine-tune this pre-trained
network for earthquake signals and denoise them. We also perform a task
outside of its initial training domain - a P- and S- wave arrival
picking, demonstrating the wide potential for applications of such a
network. Furthermore, we argue that a network built this way can serve
as a Bidirectional Encoder Representation (BERT) pre-training step in
waveform Machine Learning applications, thus reducing necessary training
time for potential applications. This work proves the concept and steers
the direction for further research of earthquake-induced source
separation. We have therefore aimed to describe the technicalities in
detail. We provide a reproducible research repository with the
algorithms and datasets.