300 likes | 320 Views
This research paper discusses the use of model-based event detection in sensor networks. The researchers propose a method that utilizes principal component analysis (PCA) to model observed quantities and detect events by identifying deviations from expected behavior in a reduced feature space. Experiments and results are presented, and future work is discussed.
E N D
Jayant Gupchup, Andreas Terzis, Randal Burns, Alex Szalay Model Based Event Detection in Sensor Networks The Johns Hopkins University
Outline • Motivation • Data & Model • Experiments and Results • Discussion The Johns Hopkins University
Motivation The Johns Hopkins University
“Event starts” Detect Event Increase Sampling Frequency/Trigger Alarms “Event ends” Return to steady behavior Importance of detecting events • Fixed Sampling: • High Freq => too much data • Low Freq => miss temporal transients • Detect Events: Adaptive Sampling • (increase % of usable data) • Conserve Energy • Alarm Triggers • Correlate events and observed • phenomena in large databases The Johns Hopkins University
Rain Event Non-Event Days Sample Event The Johns Hopkins University
Solution: Rough Sketch • Model observed quantities using Principal Component Analysis (PCA). • Project original data on a “feature space” (reduce dimensionality) • Look for observations “deviating” from Average/Expected behavior in the feature space The Johns Hopkins University
Principal Component Analysis (PCA) X : Points original space O : Projection on PC1 Variable #2 First Principal Component Variable #1 PCA : • Finds axes of maximum variance • Reduces original dimensionality (In e.g. from 2 variables => 1 variable) The Johns Hopkins University
Motivation for Using PCA Typical day: “Fits model well” Event day: “Large residuals” The Johns Hopkins University
Why Not Soil Moisture ? Reaction to event Reaction to event The Johns Hopkins University
LifeUnderYourFeet Data & Model Preparation The Johns Hopkins University
LifeUnderYourFeet data • 10 MICAz Sensors • Air Temperature (AT) • Soil Temperature (ST) • Soil Moisture • Photo Sensor The Johns Hopkins University
Air Temp vs. Soil Temp Notice the phase lag for Soil Temperature The Johns Hopkins University
Data Preparation t=10 t=20 … t=1440 1 day, 10 sensors • Model built on Air temperature and Soil Temperature. Size of matrix : [(# of days x 10) X 144] The Johns Hopkins University
PCA Bases (AT & ST) Eigenvector1 Is the Diurnal cycle similarity eigenvector1 for ST & eigenvector2 for AT The Johns Hopkins University
Methods and Results The Johns Hopkins University
Methods Three methods 1) Basic Method • Projections on the first principal component for AT 2) Highpass Method • Removes seasonal drift by looking at sharp changes in the local neighborhood. 3) Delta method • Makes use of the inertia of the soil and seasonal drift The Johns Hopkins University
Test Data • Test Period : 225 days between September, 2005 – July, 2006 • 48 major events were known to occur (taken from the BWI weather station, http://www.wunderground.com/US/MD/Bwi_Airport.html) • Offline Analysis The Johns Hopkins University
Method 1 : Basic Method • Considers only Air Temperature. • First Basis Vector covers 55% of variation in the data First Basis Vector (PC1) = X 1 day Average Day n Day 1 Day 2 The Johns Hopkins University
Method 1: Basic Method (cont.) Results : Drawback: • Does not consider seasonal drift • Does not make use of the inertia information of the soil. The Johns Hopkins University
Method 2 : Highpass Method • Again, Considers only Air Temperature • Highpass filter on ‘E1’ series. Call this series ‘S1’ • Highpass filters detects sharp changes by considering the local neighborhood only => Removing seasonal drift • Threshold on ‘S1’, values below the threshold are tagged as events. The Johns Hopkins University
Method 2: Highpass Method (cont.) Results : Drawback: • Does not make use of the inertia information of the soil. The Johns Hopkins University
Method 3 : Delta Method • Considers Air Temperature and Soil Temperature • Create E1 series for AT and E1 series for ST separately as discussed before • Highpass filter on AT_E1 & ST_E1 => AT_S1 & ST_S1 • Delta = AT_S1 – ST_S1 for all days. • Set a threshold on the Delta series. The Johns Hopkins University
Method 3: Delta Method (cont.) Results : The Johns Hopkins University
Event detection for 12/13/2005 – 01/02/2006 Due to the inertia of the soil, ‘Delta method’ shows sharper negative peaks for event days. The Johns Hopkins University
Discussion The Johns Hopkins University
Future work • Implement “Online event detection” • Compute Basis vectors from historic data. • Load the ‘basis vectors’ and ‘threshold’ values on the motes. • Apply technique for faulty sensor detection • Detect localized events by forming clusters of motes with similar eigencoefficients. • Consider variants of PCA (Gappy-PCA, online-PCA). The Johns Hopkins University
Acknowledgements • Ching-Wa Yip 1 - PCA C# library and Discussions. • Katalin Szlavecz 2 & Razvan Musaloui-E 3 • Domain expertise and data collection. • Jim Gray 4 & Stuart Ozer 4 • Online database 1 : JHU, Dept of Physics & Astronomy 2 : JHU, Dept of Earth and Planetary science 3 : JHU, Dept of Computer Science. 4 : Microsoft Research The Johns Hopkins University
Future work • Online event detection on the motes • Apply this method for faulty sensor detection • Detect localized events by forming clusters of motes with similar eigencoefficients. • Consider incomplete days using Gappy-PCA. • Explore incremental & robust PCA techniques. The Johns Hopkins University
Training Set (Air Temp) • Seasons exhibit “Diurnal Cycles” around their daily mean (DC component) • Construct Zero-Mean Vectors for each Sensori for each day (remove DC Component) • Remove outliers using a • simple median filter to • build the training set X The Johns Hopkins University