Adaptive Stream Resource Management Using Kalman Filters Aug UCLA DB seminar
Paradigm Base station passively wait for sensors update UCSB Stanford (STREAM) U Maryland Brown (Aurora) U Pennsylvania Cornell (Cougar) Base station can actively contact specific sensors Berkeley (TinyOS / TinyDB) Brown (Aurora)
Motivation Reduce communication cost Reduce power consumption Reduce bandwidth Reduce computation cost at base station Tradeoff : imprecise answer
Basic approach Base station keep a stale copy of sensors reading Sensors update only when reading fall out of boundary
Improvement Sensors readings are predictable Location of moving objects power usage Temperature Heart-beat rate Network traffic Precipitation?
Kalman filter Prediction of discrete time linear system
Kalman filter I x – state u – user input a – relation between successive states b – relation between input and state
Kalman filter II w - noise
Kalman filter III z – measurement v – measurement noise h – relation between measurement and state
Kalman filter IV
Kalman filter V
Dual Kalman Filter (DKF) Base station and sensors maintain the same Kalman filter
Architecture of DKF model
Experiment – moving object
Result – communication cost
Experiment – power load
Result – communication cost
Conclusion Shift “intelligence” (computation) to sensors Compressing Historical Information in Sensor Networks A. Deligiannakis, Y. Kotidis, N. Roussopoulos in SIGMOD 2004 Optimization of Online, In-Network Data Reduction J. M. Hellerstein, W. Wang in International Workshop on Data Management for Sensor Network 2004