Development and Application of Collective Anomaly Detection methods to Electromagnetic Satellite Data

  • Vyron Christodoulou

Student thesis: Doctoral Thesis

Abstract

Collective anomaly detection (AD) is the problem of detecting and identifying change when it deviates from what is considered normal. In this work, the focus is on anomalies, known as seismic precursors, that appear in the Earth’s electromagnetic field (EMF) before a major seismic event. The hypothesis investigated in this work is whether seismic events can cause such anomalies that can be detected. A novel and robust workflow in the form of a prototype tool that streamlines data processing from satellites, performs data cleaning and presents the results of collective AD both qualitatively and quantitatively is proposed. As part of the tool, two new methods are developed, a mixed CUSUM-EWMA (CE) and a Fuzzy-based method (FSB) and are applied in real-world data along with two well established off-the-shelf methods, HOT-SAX and 1D-SAX. No prior knowledge is assumed for the anomalies and the methods developed are evaluated in ECG time series signals. ECG signals were used as a good proxy for EMF anomalies because of their similar nature and variation. The knowledge acquired by the benchmark data was transferred and applied to the real-world data from ESA’s SWARM satellites. Equally importantly, a solution to compare collective AD methods is proposed in the form of a new measure, called the R-Measure. The R-Measure is used to assess the performance of each of method before their application in the real-world data. The overall workflow, findings and future pathways for the correlation of seismic events and precursors are presented in this work.
Date of AwardSep 2020
Original languageEnglish
SponsorsEuropean Space Agency - ESA
SupervisorGeorge Wilkie (Supervisor) & Yaxin Bi (Supervisor)

Keywords

  • Machine learning
  • SWARM Satellites
  • Symbolic Approximation
  • Fuzzy Encoding
  • Electromagnetic Data

Cite this

'