Anomaly Detection and Algorithm Development

First, we studied anomalies detected prior to earthquakes by the ESP stations. Here are a few examples:

Example 1

Close up Kodiak magnetometer readings with earthquake overlays of the November 27th 2014 earthquake. Green dot indicates start of anomaly, red dot indicates end of anomaly. The vertical red line indicates the seismic event. Please note that this chart and analysis was performed in AK UTC time zone, to match the sensor’s time zone. Previous anomalies were performed in UTC, while both time zones were listed.

Statistical data taken from the day prior to the anomaly compared with statistics of magnetometer readings during the November 27th anomaly. All measurements in gauss.   

Table1

Details: November 27, 2014 earthquake event at 08:17:36 (UTC -8 AK), (16:17:36 UTC)

Kodiak, Alaska Station

Begin Anomaly: 2014-11-26 at 01:30 AK (09:30 UTC)

End Anomaly: 2014-11-26 at 21:00 AK (05:00, 2014-11-27 UTC)

Anomaly duration = 19.5 hours

Anomaly start time prior to earthquake = 31 hours    


Example 2

On January 19, 2015 USGS reported a 4.2 earthquake approximately 498 kilometers away from the sensor station. This event produced a significant anomaly in all axes, and a unique return to baseline. All three axes changed values significantly, returned to baseline, then changed again prior to the earthquake. This whole region also has a sharp jagged pattern. This initial anomaly occurred 33 hours before the event. 

Statistical data taken from the day prior to the anomaly compared with statistics of magnetometer readings during the January 19th anomaly. All measurements in gauss.   

Table2


Details: January 19, 2015 earthquake event at 10:36 (UTC), 02:36 (AK)

Kodiak, Alaska Station

Begin Anomaly: 2015-01-18 at 06:00:00 UTC

End Anomaly: 2015-01-19 at 13:00:00 UTC

Anomaly duration = 31 hours

Anomaly start time prior to earthquake = 28.5 hours


Magnetic Field Data Analysis Processes

This project analyzes the magnetic field data collected in Alaska, the most seismically active region in the world.

1. Magnetic field strength (x-y-z) is measured and filtered through multi-signal noise-canceling algorithms to enable ‘clean’ observation of the baseline magnetic signal.

2. Anomaly detection algorithms are applied to the data to identify unusual activity that may be indicative of an impending earthquake.

3. A separate algorithm clusters the anomalous points to identify statistically significant activity. Machine-learning algorithms then evaluate the anomalous features extracted from historical data to identify “Earthquake Signal Precursors.” The goal is to forecast, in real time, the increasing risk of an imminent earthquake.

Processes Data Structure - Raw Magnetic Data (50 - 123 Hz) has a Diurnal Signal Pattern


Filtering - Utilization of a Bandpass Filter Allows Finer Analysis for How Earthquake Signals Affect Magnetic Field Vectors

Anomaly Detection - Each Anomaly Spike is Characterized by Studying Values along the Magnetic Field Duration

Forecast an Earthquake (Theory)

By analyzing the data in this fashion, we should be able to observe perturbations in the magnetic field vectors, allowing for machine learning to build a confidence level on earthquakes occurring in the near future.


Back to NASA Europa 2016 main page

© Trillium Learning 2017        info@trilliumlearning.com        AWorldBridge.com        +1 973-907-2332