Title of Presentation: A Time Series Analysis Method for Geophysical Sensor Networks

Primary (Corresponding) Author: Robert Granat

Organization of Primary Author: NASA Jet Propulsion Laboratory


Abstract: Modern sensor networks are collecting enormous volumes of observational information about geophysical systems. These sensor arrays, including GPS networks, ocean buoys, and seismic networks, are often designed to detect one particular phenomenon, but also capture signatures generated by a wide variety of other processes. We present a method for identifying, detecting, and cataloging these signals.

Our method is based on the use of hidden Markov models (HMMs). HMMs are statistical models for time series data; fitting an HMM allows use to describe the statistics of the data in a simple way that ascribes discrete modes of behavior to the system. Our HMM fitting algorithm, which uses a regularized deterministic annealing expectation-maximization (RDAEM) procedure, allows for far greater solution stability than other methods, even when the model is unconstrained.  This has two benefits.  First, the method can be used in situations in which the underlying system model is unavailable or unreliable, as is the case in analysis of many complex physical and engineering systems.  Second, the method can be used in applications where the stability and reliability of results is paramount.

This approach provides several analysis capabilities for time series data.  These capabilities include:

  • Signal Search: a user identifies a signal of interest in streaming or stored data; matches to this data ranked according to relevance are retrieved from stored data.
  • Signal Detection: a user identifies a signal of interest; incoming data that matches that signal triggers an alert.
  • Anomaly Detection: nominal data is used to train the method; incoming or stored data that differs significantly from the nominal case triggers an alert.
  • Mode Segmentation: a time series containing unknown activity is segmented into statistically significant behavior modes.
  • Sensor Network Change Detection: mode segmentation analysis is applied to each sensor in a network; when modes change simultaneously across the entire network or within a significant sub-network, an alert is triggered.

These analysis capabilities have been implemented in our software, called RDAHMM (Regularized Deterministic Annealing Hidden Markov Models), and integration with the SCIGN and SOPAC GPS networks through the QuakeSim project web portal environment is ongoing.