Beginning with Project: Data Driven Learning for Earthquake Detection

Motivation

Earthquakes are one of the most destructive natural calamities; they pose a serious threat to life and property. Early warning systems can potentially save several lives and help mitigate damage to property.  An Earthquake Early Warning System is an integrated system of sensors (accelerometers), computers, communication systems and alarm systems that is set up to detect the onset of an earthquake, its estimated magnitude,  and issue a warning before the expected time of arrival in the region for which the system is designed. Such a system can give a warning a few seconds to a few tens of seconds prior to significant ground movement. Although it seems like not much can be altered during this period, prior training and good management can potentially save a lot of lives. Trains can be slowed down, lifts can be halted, industrial systems can be shut down, gas supplies can be shut off, electrical grids can be stabilized, fire engines can be pulled out of the fire station and people can seek shelter under their desks or rush to the nearest open ground [1].

The San Andreas Fault line extends through California, causing thousands of small earthquakes in California every year. Earthquakes of greater magnitude also occur due to the fault, approximately every 150 years, causing huge loss to public life and property [2]. Community Seismic Network (CSN), an ongoing project at Caltech, is an earthquake monitoring system employing a dense array of inexpensive sensors for detection of earthquakes. A large number of such sensors have already been deployed in California, which have recorded data during past few geospatial events. Apart from this, California Integrated Seismic Network (CSIN) has also recorded several earthquake using high quality sensors [3].  Using these gigabytes of data collected from over thirty events, we can use intelligent techniques to build and validate a model to process incoming acceleration data in real time and issue an early warning for the impending seismic event. Supervised Machine learning allows us to use the information extracted from the patterns of the previous earthquakes (labeled data) to be able to detect potential future threats with higher confidence

This SURF project was focused on the supplementing the current earthquake detection with a more precise and validated earthquake detection, hypocenter estimation and Magnitude prediction model. It was shown that with the use of statistical and machine learning tools like the Hidden Markov models, number of false positives in detecting seismic acceleration (a pick) from a sensor could be lowered significantly. Further, using this information, an estimate could be made on the hypocenter within seconds of occurrence of the event. Furthermore, after a analysis, we could classify the severity of the shaking and issue early warning and disseminate this information accordingly.

How does earthquake warning happen?

The method of early warning for an earthquake has been around for quite some time. Here we give a detailed explanation on earthquake early warning systems, and the challenges it faces.

The USGS website mentions how an earthquake early warning works – “Earthquake early warning systems work because the warning message can be transmitted almost instantaneously, whereas the shaking waves from the earthquake travel through the shallow layers of the Earth at speeds of one to a few kilometers per second (0.5 to 3 miles per second). The diagram below shows how such a system would operate. When an earthquake occurs, both compressional (P) waves and transverse (S) waves radiate outward from the epicenter. The P wave, which travels fastest, trips sensors placed in the landscape, causing alert signals to be sent ahead, giving people and automated electronic systems some time (seconds to minutes) to take precautionary actions before damage can begin with the arrival of the slower but stronger S waves and later-arriving surface waves. Computers and mobile phones receiving the alert message calculate the expected arrival time and intensity of shaking at your location.”

Figure 1. Earthquake Early Warning Basics Source: USGS Website

Thus, there are two key factors to any earthquake early warning system that determine how efficacious the system is:

1. How dense is the sensor network
2. How fast information can be processed and warning can be issued in real time

Needless to say, the higher is better in both the cases.

CSN aims at affordable and low cost early warning for all. Due to its cost effective nature, CSN quake catching system can be deployed in many developing countries as well. Hence, Community Seismic Network differs from other early warning systems in two important ways:

1. Cost per sensor: While most early warning systems use high quality sensors and equipment that cost hundreds of thousands of dollar, each CSN deployment cost only $200. To compensate for reduced quality of sensors, CSN has to increase sensor deployment density spatially.
2. Earthquake Detection Algorithm: While most of the expensive sensors can differentiate between S wave and P wave, which is important in determining origin of earthquake, and hence issuing early warning, it is difficult to ascertain the nature of wave from CSN sensors due to their cost effective nature.  Thus CSN needs algorithms to catch earthquakes regardless of the nature of the wave.

CSN’s very reliable and robust system takes care of these issues. With availability of more data every passing moment, these systems can be bettered, made much astute and reliable, upon which this SURF project focusses.