Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RECOVAR: Representation Covariances on Deep Latent Spaces for Seismic Event Detection

Onur Efe, Arkadas Ozakin

TL;DR

An unsupervised method for earthquake detection that learns to detect earthquakes from raw waveforms, without access to ground truth labels is developed, comparable to, and in some cases better than, some state-of-the-art supervised methods.

Abstract

While modern deep learning methods have shown great promise in the problem of earthquake detection, the most successful methods so far have been based on supervised learning, which requires large datasets with ground-truth labels. The curation of such datasets is both time consuming and prone to systematic biases, which result in difficulties with cross-dataset generalization, hindering general applicability. In this paper, we develop an unsupervised method for earthquake detection that learns to detect earthquakes from raw waveforms, without access to ground truth labels. The performance is comparable to, and in some cases better than, some state-of-the-art supervised methods. Moreover, the method has strong \emph{cross-dataset generalization} performance. The algorithm utilizes deep autoencoders that learn to reproduce the waveforms after a data-compressive bottleneck and uses a simple, cross-covariance-based triggering algorithm at the bottleneck for labeling. The approach has the potential to be useful for time series datasets from other domains.

RECOVAR: Representation Covariances on Deep Latent Spaces for Seismic Event Detection

TL;DR

An unsupervised method for earthquake detection that learns to detect earthquakes from raw waveforms, without access to ground truth labels is developed, comparable to, and in some cases better than, some state-of-the-art supervised methods.

Abstract

While modern deep learning methods have shown great promise in the problem of earthquake detection, the most successful methods so far have been based on supervised learning, which requires large datasets with ground-truth labels. The curation of such datasets is both time consuming and prone to systematic biases, which result in difficulties with cross-dataset generalization, hindering general applicability. In this paper, we develop an unsupervised method for earthquake detection that learns to detect earthquakes from raw waveforms, without access to ground truth labels. The performance is comparable to, and in some cases better than, some state-of-the-art supervised methods. Moreover, the method has strong \emph{cross-dataset generalization} performance. The algorithm utilizes deep autoencoders that learn to reproduce the waveforms after a data-compressive bottleneck and uses a simple, cross-covariance-based triggering algorithm at the bottleneck for labeling. The approach has the potential to be useful for time series datasets from other domains.
Paper Structure (16 sections, 1 equation, 3 figures, 1 table)

This paper contains 16 sections, 1 equation, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Overview of the appproach
  5. The CNN autoencoder
  6. The loss
  7. The architecture
  8. Computing the cross-covariance
  9. Triggering: classification metrics
  10. Training and Testing Procedures
  11. Datasets and preprocessing
  12. Training
  13. Testing
  14. Results
  15. Examples of Latent Representations
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Outline of the Single autoencoder method. The encoder and decoder blocks are seen at the top. After the autoencoder is trained, the waveform enters the encoder and the latent representation is fed into batch normalization and an autocovariance profile is obtained. A simple metric measuring the prominence of the autocovariance peak is used for triggering the detector.
  • Figure 2: Two different ways of forming representation ensembles which can be used for earthquake detection. The Augmented autoencoder method (left) involves obtaining multiple augmented raw waveforms which are encoded by the same encoder while the Multiple autoencoders method (right) encodes the same raw waveform using different encoders to obtain ensemble of representations.
  • Figure 3: Earthquake (top) and noise (bottom) waveforms, and their latent representations for single autoencoder and multiple autoencoders (for the multiple autoencoders case, we show one typical representative from multiple autoeconders). The $x$-axis in the representation plots is "compressed time", the $y$-axis is the channel index, and the color coding represents the activation level of the relevant channel at the given instance. Earthquake signals lead to temporal "phase transitions" in latent space representations and a strong covariance profile, in contrast to noise signals. The covariance in latent representations is seen to be significantly more discriminative than the covariance of raw waveforms.