Energy Efficient Data Gathering in Sensor Networks F

Energy Efficient Data Gathering in Sensor Networks F. Koushanfar, UCB N. Taft, Intel Research M. Potkonjak, UCLA A. Sangiovanni-Vincentelli, UCB Sampling • • Methodology Goals – Extend the lifetime of a network by Sensor Data reducing energy consuming activities such as sensing & communication. Non-parametric Statistical Modeling • Evaluate Model Learn – Use a data-driven approach. Explore different energy saving approaches: Inter-node Sample Rate Compression – Sampling, Compression, Prediction Selection Prediction & Sleeping ILP for Domatic Compressed Sampling – Consider these individually and jointly Rate Compression • Data Sets Why? Temporal correlation between signals is high Approach: – examine 1/m – use linear interpolation to predict unsampled points – select sampling rate based on a target error rate Take-away points Sampling 1 in 2 1 in 5 rate (min) – More reduction possible for Temp temperature & humidity, 0 0 (errors) less for light. Humidity 0 0 – Different sensors have (errors) different minimal sampling rates Light nonunifo sam pling rm (errors) Sleeping Schedule 1% 4% 1 in 10 (min) 1 in 20 (min) 0 1% 5% 7% frequency Small differences btwn successive samples more frequent Prediction 1/5 1/10 1/15 1/20 1/25 than large differences => huffman coding attractive Sampling rate (30 sec time unit) • Idea: take advantage of temporal and spatial correlations Temperature so that one node can be used to predict a few others. • Approach: non-parametric method – Build histograms of conditional probabilities P(n 2=y/n 1=x) (pair-wise • Results (percentage of bits needed) prediction) – Temperature 45%, Humidity 40%, Light 20% – Prediction: use average of this distribution (minimizes mean squared error rate • Take-away points error) Light – Light is most compressable modality – Model Validation: resubstitution methods. – Savings uniform across all nodes Build model using 6 days, evaluate on next 21 days. – Huffman very close to optimal • Take-away points – For temperature & humidity, most nodes can be easily predicted by others to within 5% accuracy. error rate Sleeping Coordination – Light is more difficult to predict. • The weight of a directed edge ni nj – Prediction ability is often not symmetric. shows the conditional prob. P(nj|ni) Number of disjoint dominating sets for each value of error Sleeping Coordination (Results) • Edges are included in the graph when Avg. error Temp Humidity Runtime probabilities are above a threshold (seconds) • Problem: Find the maximal number of disjoint dominating sets 0. 01 1 1 0. 04 • Can be formulated as an Integer Linear Program n 2 Sample dominating sets n 5 0. 02 2 1 0. 06 • Take-away points 0. 03 5 2 0. 09 n 1 n 4 e 1 P(n 2|n 1) n 7 0. 04 6 5 1. 65 – Short run times, optimal e 2 P(n 4|n 1) n 3 n 6 0. 05 9 6 3. 70 – Works well for temperature & humidity, but not for light : e 8 n 2 n 5 e 1 : – Works differently for differently modalities e 2 n 1 e 8 P(n 2|n 5) n 4 n 7 – If willing to tolerate 5% error rate in prediction, can extend lifetime 5 -10 times n 3 n 6 – For light, it’s harder to find disjoint dominating sets November 18, 2004