Information Integration Through Events

Funded by NSF Science and Engineering Informatics and Information Integration

The objectives of this project are to explore methods for the integration of heterogeneous sensor data streams. The application context is the Gulf of Maine GoMOOS data systems. GoMOOS collects data on multiple variables from different locations and depths.

Our integration approach is based on abstracting "events" from time series. Events represent an abstraction over sensor data streams into units with some common property (e.g. constancy or linearity). Events are information objects in their own right that explicitly represent change (or absence of change). Events are collected and stored as a new data type with spatial and temporal footprints. The abstraction creates an event data type for analysis and display. Events as topological changes (splits and mergers) have been formally defined. 

Project participants: Kate Beard (PI), Neal Pettigrew (CoPI), Mike Worboys (CoPI), Tony Stefandis (CoPI), Sue Elston (post doc), Heather Deese (PhD), Jixiang Jiang (PhD), Avinash Rude (MS)

Time series event extraction:

Within the GoMOOS data system there are 10 moored buoys observing 15 different variables at several depths. We extract events that are specific to the observed variables (e.g. temperature, salinity, wind speed), the domain and application context (e.g. inflow, outflow, stratification, mixing, and advection events in the oceanographic context).

The figure below illustrates an individual time series for salinity at Buoy B for 3 depths (1meter, 20meters and 50 meters). Observations are taken hourly. An annual cycle is fit to the time series using a least squares fit to the annual and semi annual harmonics for the hourly data from 8/2001-8/2004. One set of events of interest for the oceanographers are the deviations from the annual cycle. The black line in the second graph illustrates the deviations in the 1meter depth time series from the fitted annual cycle and the red and blue segments indicate the derived anomaly events:  periods of positive and negative deviations from the annual cycle.

In the next figure, positive and negative deviation events from the annual cycle are compared across several buoy locations in the Gulf (A, B, E, I and L). From this event band display, a strong negative anomaly appears Gulf wide starting in summer 04.  The color value in the event bands encodes the mean anomaly value over the duration of the event. The gray bands indicate periods of no data.

Another phenomenon of interest to oceanographers is the timing and duration of the mixed season (period when temperature, salinity, and density are uniform with depth). Mixing season events were extracted using piecewise fitting to the difference between the density measures at 1m and 20 meters and 1 meter and 50 meters. The top graph illustrates the time series for density (sigma-T) measured at 1, 20 and 50 meters at Buoy I. In the second graph the green line indicates the differences between 1 and 20 meters. The black line segments are a piecewise fit to the green line and the breaks indicate event start and end locations for mixing (density is same with depth) and stratification (density varies with depth) events. The third graph is the same for differences between 1 and 50 meters.

The next figure illustrates comparison of mixed season events across locations and time (between July 2001 and July 2006). The gray bars represent missing data events, which were more prevalent in the early deployment period. This display supports comparison of the onset and duration of the mixed season across the gulf by buoy location, year, and depth.  During winter 2002-2003 and 2004-2005 the onset and duration are similar across location and depth. In 2003-2004 the onset of the mixed season is delayed with depth and 2005-2006 shows wide variation in mixed season duration with location.