Type analysis of laboratory seismic events by convolutional neural network

Abstract

In this work, we successfully identified seismic events (observations of earthquakes) in seismograms using a Convolutional Neural Network (CNN). In accordance with past (analogue) seismogram interpretations, we did not treat digital seismograms as a time series, as per the general method, but, rather, converted them into time snaps of continuous data flow. Multichannel seismograms were represented with a time-frequency domain in the form of multilayer images, with each signal channel forming one image layer. Images were then exposed to CNN (composed of three convolutional blocks whose architecture design was justified using Bayesian optimization). To improve reliability, we evaluated the posterior type function (PTP) as a combination of the probabilities of all of the considered classes of signal types (five in our case) which increased robustness of the identification. For data, we used records of acoustic emission (AE) events. The events were generated during laboratory loading experiments originally performed to study material/rock properties. As known, AE events may be studied in the same manner as natural earthquakes and treated in other ways as laboratory earthquake models. AE events are less complex compared to natural earthquakes where many of the physical parameters are known or may be controlled. Based on our results, we concluded that the successful identification of AE events is a necessary step prior to applying the proposed methodology for identifying natural earthquakes in seismograms.