Separation and denoising of seismically-induced ground-motion signals with dual path residual neural network architecture.

Abstract:

Source separation is an important task in signal processing, with application in music, speech, but also in seismic signals. Multiple signals can be added together to obtain a mixture, but a mixture can also be separated back into individual signals. In this work, we present a technique from the Machine Learning domain, called a dual-path residual neural network that utilizes the ability of the neural network to process source mixtures End-to-End in the time domain, reconstructing original sources from mixtures. We recorded data with a Raspberry Shake and Boom seismic sensor widely used by Citizen Science Community (24 bits ADC, Sigma-Delta, 144 dB dynamic range, flat frequency response from ~2 seconds to ~40 Hz) at the University of Vienna, with a sampling frequency of 100 Hz. The sensor is located in direct vicinity of railway tracks (S40, U4, Spittelau station). We trained the network to produce “clean” individual signals from “mixed” waveforms, and we demonstrate that even though predicted signals contain under-suppressed features from each other, they do correspond well to their target counterparts. One of the most stunning features of such separation is rather precise time localisation of individual signals. We also show a technique know as Transfer Knowledge, that allows a pre-trained network first to denoise signals and second to perform tasks outside of its training domain - such as P- and S- wave arrival picking, allowing virtually endless applications of the same architecture. This work proves the concept and steers the direction for further research of earthquake-induced source separation.