Model-Driven Data Acquisition in Sensor Networks - Amol Deshpande et al., VLDB ‘04

1. Model-Driven Data Acquisition in Sensor Networks- Amol Deshpande et al., VLDB ‘04 Jisu Oh March 20, 2006 CS 580S Paper Presentation

2. Problems – Sensornet is not a database! • Databases - complete, authoritative sources of information - Answer a query “CORRECTLY” based upon all the available data • Sensornets - Misrepresentations of data :only samples acquired, not random :need to complement the sensornet readings - Inefficient approximate queries :Existing query processing from a completist’s approach is costly

3. Solutions – Model-driven data acquisition • Use a statistical model which maps the raw sensor readings onto physical reality - in order to robustly interpret sensor readings - and provide a framework for optimizing the acquisition of sensor readings

4. Proposed approaches- Architecture • Architecture for model-based querying in sensor networks

5. Proposed approaches (cont.)- Probability density function (pdf) • Probability density function (pdf) • Based on time-varying multivariate Gaussians • Estimates sensor readings in the current time period • Properties • : correlation between different attributes • : cost differential • - Constraints: need historical data and training with them

6. Key concepts – Probabilistic queries • BBQ query processing (static probabilistic model) • A user requests a range query that ask if an attribute Xi is in the range [ai, bi] with confidence (1-α) • Marginalize a prior density (probability density function), p(Xi, …, Xn) to a density over only attribute Xi, p(xi) • Compute P(Xi ∈[ai, bi]) = ∫aibip(xi)dxi • Answer true if p>1-α, false if p> α • Otherwise, move on conditioning step

7. Key concepts – Probabilistic queries (cont.) • BBQ query processing (conditioning) • Acquire new sensor readings, o, a set of observations • Marginalize a posterior density (conditional probability density function), p(Xi, … Xj-1, Xj+1, … Xn | xj) to a density over only attribute Xj p(xj|o) • Compute P(Xi ∈[ai, bi] | xj) = ∫ai, bip(xi|o)dxi • Answer true if p>1-α, false if p> α

8. Key concepts – Probabilistic queries (cont.) • BBQ query processing (dynamic probabilistic model) • describe the evolution of the system over time • for time evolved attributes • use transition model p(X1t+1, .. Xnt+1 | X1t .. Xnt) for conditioning • marginalize transition model then obtain p(X1t+1, .. Xnt+1 | O1…t)

9. Key concepts (cont.)– Choosing on observation plan • Choose a data acquisition plan for the sensornet to best refine the query anser. • Cost(O) = Ca(O) + Ct(O) • Ri(O), statistical benefit of acquiring a reading • Minimize O ⊆{1, …, n} C(O), such that R(O) ≥ 1-α

10. Contribution • integrate a database system with a correlation-aware probabilistic model • build the model from historical readings and improve it from current readings • answer approximately SQL queries by consulting the model + shield from faulty sensors + reduce # expensive sensor readings + reduce # radio transmissions

11. Experiment results

12. Experiment results

13. Conclusions • Probabilistic model-driven data acquisition • Help to provide approximations with probabilistic confidences, significantly more efficient to compute in both time and energy • Strong assumptions and constraints • The model must be trained • based on static network topology • all sensor nodes are well synchronized

14. Questions?