Capturing Sensor-Generated Time Series with Quality Guarantees

1
Capturing Sensor-Generated Time Series with
Quality Guarantees

• Iosif Lazaridis
• Sharad Mehrotra
• University of California, Irvine
• ICDE 2003, Bangalore, India

2
Talk Outline

• Sensors and QUASAR
• Archival Vs. On-line applications
• Poor Mans Compression
• Using Prediction
• Experiments
• Conclusions

3
Sensors

• Sensors are becoming smaller, cheaper, and more
  configurable
• Systems incorporating large numbers of them will
  be feasible
• Main problems limited energy, limited bandwidth
• Goal limit communication
• Benefits data recipient as well

4
Quality Aware Sensing Architecture (QUASAR) _at_ UC
Irvine

• Tradeoff between data/application accuracy, and
  system performance
• This paper how to capture streaming time series
  with a given error tolerance

5
Archival Vs. Online Applications (I)

• Online applications discard the time series
  history (e.g., intrusion detection)
• Archival aspect is important because
• Data may be precious (e.g., once-in-a-lifetime)
• Unforeseen uses of the data (e.g., new mining
  applications)
• Roll-back feature (e.g., what caused an accident
  ?)

6
Archival Vs. Online Applications (II)

• Online applications have
• timing requirements
• time-variable needs (e.g., sensor S1 may become
  more, or less popular with time)
• Archival
• timing less important
• constant needs (archived time series must meet
  some overall quality criteria)

7
Preliminaries

• Let P lt p1, p2, , pn gt be a sequence of
  environmental measurements (time series)
  generated by the producer, where n now
• Let S lts1, s2, , sngt be the server side
  representation of the sequence
• A within-? quality data collection protocol
  guarantees that
• for all i error (pi, si) ? ?

8
Piecewise Constant Approximation (PCA)

• Given a time series Sn s1n a piecewise
  constant approximation of it is a sequence
• PCA(Sn) lt (ci, ei) gt
• that allows us to estimate sj as
• s j ci if j ?ei-11, ei
• c1 if j ? e1

9
Poor Mans Compression

• Goal Given time series of sensor values,
  generate a within-?capt PCA representation of it
• Poor Mans Compression - Midrange (PMC-MR)
• Maintain m, M as the minimum/maximum values of
  observed samples since last segment
• On processing pn, update m and M if needed
• if M - m gt 2?capt , output a segment ((mM )/2,
  n-1)

Value
Variant PMC-Mean Uses the mean of the observed
points
?capt 1.6
Time
1 2 3 4 5
10
Optimality of PMC-Midrange
CLAIM A PCA representation BETTER with KltK
segments violates ?capt
Hence, the segment of BETTER must contain
ek1 Since PMC-MR output a new segment after ek
? range of values in ek-11,ek1 is gt 2 ?capt
? Segment of BETTER violates ?capt tolerance
11
Why Predict?
Future (no data)
Recent Past (precise data in sensor)
History (compressed data in archive)
Time
n ( now )
n - nlag
12
Who Should Fit the Predictive Model?

• Server
• Long-haul models possible
• No extra communication
• - No prediction accuracy guarantee
• - Server only sees approximate time series
• Sensor (Data Producer)
• Short-term adaptation (but )
• Sensor sees precise time series
• Prediction accuracy ?pred guarantee

13
Producer-Side Prediction

• M is the predictive model, and ? its parameters
• Producer sends (M, ?) to server
• Server estimates time series value spredj using
  (M, ?)
• Producer generates new (M, ?) when
• error(pj,spredj)gt ?pred

14
Issues in Prediction

• Choice of M is domain-specific
• Goal is to minimize communication
• Prediction accuracy error(pj,spredj) lt
  ?pred
• Parameter size ?
• Adaptive Model Selection
• E.g., constant velocity or constant acceleration,
  for predicting a moving objects location

15
Combining Prediction with Compression

• Imagine that ?pred lt ?capt
• No need to do extra work for archival
• the sequence of (M,?) suffices
• In general, compression should exploit the
  information given in the sequence (M,?)
• Solution
• djpj-spredj (error time series)
• Compress d1n within- ?capt

16
Combining Prediction with Compression (Example)
17
Experiments

• Data sets
• Synthetic Random-Walk
• x1 0 and xixi-1si where si drawn
  uniformly from -1,1
• Oceanographic Buoy Data
• Environmental attributes (temperature, salinity,
  wind-speed, etc.) sampled at 10 intervals from a
  buoy in the Pacific Ocean (Tropical Atmosphere
  Ocean Project, Pacific Marine Environment
  Laboratory)
• GPS data collected using IPAQs (not in paper)
• Experiments to test
• Compression Performance of PMC
• Benefits of Model Selection
• Query Accuracy over Compressed Data
• Benefits of Prediction/Compression Combination

18
Compression Performance
K/n ratio number of segments/number of samples
19
Query Performance Over Compressed Data
How many sensors have values gtv? (Mean
selectivity 50)
20
Impact of Model Selection

• Objects moved at approximately constant speed (
  measurement noise)
• Three models used
• locn c
• locn cvt
• locn cvt0.5at2
• Parameters v, a were estimated at sensor over
  moving-window of 5 samples

K/n ratio number of segments/number of samples.
?pred is the localization tolerance in meters
21
Combining Prediction with Compression
K/n ratio number of segments/number of samples
22
GPS Mobility Data from Mobile Clients (iPAQs)
QUASAR Client Time Series
Latitude Time Series 1800 samples
Compressed Time Series (PMC-MR) Accuracy of 100
m 130 segments
23
Conclusions and Future Work

• We motivated the importance of real-time
  applications to co-exist with data archival
• We showed how compression can be used to reduce
  the size of the archived time series optimally
• We investigated how prediction can be used to
  limit communication by allowing the database to
  estimate values ahead of time
• We noted the interplay between prediction and
  compression and showed how they can be combined
• In the future
• Adaptive algorithms for model selection
• Exploiting inter-sensor correlation for further
  reducing communication