ken innetwork sensor inference

1
kenin-network sensor inference
Networked Systems Design Implementation2-4 May
2005 Boston, MA
David Chu
Problem
Approach
Report data only if it differs significantly
from what is expected.

• Procedure
• For each sensor node, select model estimate
  parameters.
• Synchronize model between base station and node.
• At each time step...
• Perform inference at node and base station.
• If the predicted model value matches the actual
  sensor value, node does not report. Base station
  provides the predicted model value to the user.
• Otherwise, the node reports the actual sensor
  value. Base station relays this to the user.

• Communication is costly. Nodes should minimize
  messages.
• Users prefer all the data SELECT FROM
  Sensors
• Anomaly detection requires all the data.
• How do we resolve these conflicting requirements?

Multi-node cliques

• Allow multiple nodes (a clique) to collect data
  in-network at a clique root and perform inference
  over multiple sensor readings.
• Clique root decides which readings (if any) to
  send back to base station.
• Exploits spatial correlations.
• But how do we choose the right cliques? We
  developed an approximation algorithm for disjoint
  cliques case.

• Properties
• At a minimum, the user receives estimated
  readings guaranteed within a user-tolerated
  range.
• Nodes always report anomalous readings, an event
  detection primitive.
• Base station can calculate expected error and
  time to report, possibly detecting unresponsive
  nodes.
• Exploits time correlations.

Evaluation
Generalized in-network models

• Graphical model selection Markov model with time
  varying transition matrices. The always observed
  hour variable ht selects the transition matrix.
  The sensor reading xt may be observed depending
  upon the report function. Figure to left shows
  the graphical model of a 3-clique.

• Aggregate value model (above)
• Collect all values, compute aggregate value (e.g.
  average) at root, then disseminate computed
  aggregate.
• Benefits Composite value (average, median, etc.)
  may be strong predictor of sensor value.
• Overlapping cliques
• Allow multiple cliques to contain same node
• Benefits Most general model, though hard to find
  optimal.

Related work

• UC Berkeley Botanical Gardens
• 11 motes, single-hop network
• 100 training hours, 201 test hours
• Strong space correlations, strong time
  correlations

• TinyDB Madden, et al. data service for sensor
  networks
• BBQ Deshpande, et al. Base station selects
  node tour that satisfies user-requested
  confidence level.
• Distributed Inference Paskin Guestrin
  Distributed junction tree in-net
• Synopsis diffusion Gibbons, et al. Redundant
  summarizations in-net
• Approximate caching Olston, et al. model-less
  caching
• Kalman filters on streams Jain, et al. Kalman
  filters on single nodes

• Intel Research Lab, Berkeley
• 49 motes, multi-hop network
• 100 training hours, 482 test hours
• Moderate space correlation, moderate time
  correlation.
• Intra-clique setup costs dominate inter-clique
  costs.

Thanks
Amol Deshpande, Wei Hong, Joe Hellerstein Mikhail
Traskin, Carlos Guestrin
Conclusions future work

• Inter-attribute modeling
• Additional gains when modeling across multiple
  attributes at a single node.
• Savings even for weakly correlated attributes.

• Modeling sensors saves significant communication
  costs while the user still gets estimates of all
  the data.
• Bounded error allows anomaly detection
• Rich space of possible in-network models to
  explore.