1. Advanced Metrology Lab., Texas A&M University Collaborative data reduction for sensor power efficiency Aug. 25, 2008 by Chiwoo Park

2. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 2 A Bridge Manager’s problem Frequent sensor replacement can be enormous cost to a bridge manager. How can we increase the lifetime of sensors? The lifetime of bridge is 50 years… The lifetime of sensors’ battery is around 0.5 years 100 times sensor replacements!! Thousands of sensors over a bridge $$$ ??

3. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 3 Possible solutions within sensors Right now, possible solutions are reducing data processing or data transmission. Adjust sampling frequency Replace with energy- efficient sensors Data processing Center X Conceptual diagram of a micro-sensor Micro-sensorMicro-processorMicro-Radio Micro-battery Reduce data processing Replace with energy- efficient processors Reduce data transmission Replace with energy- efficient radio units Area of Interest Increase capacity of batteries X X X X Trade-off

4. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 4 And, he found.. Data reduction on a sensor gives an opportunity to prolong the lifetime of the sensor. Processor Power Consumption ProcessorActive modeSleep mode ATMega 128 (MicaZ)4nJ/instr30μW PXA255(Stargate)1.1nJ/instr20mW Radio Power Consumption Radio moduleTransmission CC2420 Zigbee Radio (MicaZ)430nJ/bit 802.11 Radio (Stargate)90nJ/bit Fact sheet about a sensor 2 x 2850 mAh X 1.5 V = 17550 mWh = 63180 J Batteries’ capacity ( 2 x AA Batteries) Batteries’ lifetime (MicaZ case) Config.ProcessorRadioTotal A0 mJ430 mJ B4 mJ43 mJ47 mJ Sense Data redu ction 1) Transmit 2) Configuration B (Lifetime: 934 days) 1) 10 6 instr / min 2) 10 5 bits / min Sense Transmit 2) Configuration A (Lifetime: 102 days) 2) 10 6 bits / min Power Consumption / min

5. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 5 Manager’s approach The approach is known as “In-network data aggregation”. 1234 5678 Leaf 1Leaf 2 9101112 13141516 Leaf m P-7P-6P-5P-4 P-3P-2P-1P X1X1 X2X2 XmXm A1A1A2A2 A3A4 A5A6 A7A8 Intermediate 1

6. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 6 Time delay matters… As the network structure gets complicated, the time delay for propagating data also increases. 1234 5678 Leaf 1Leaf 2 X1X1 X2X2 9101112 13141516 17181920 21222324 Leaf 3Leaf 4 X3X3 X4X4 25262728 29303132 Intermediate 1Intermediate 2 A1A1A2A2 A3A4 A5A6 A7A8 A9A10 A11A12 A13A14 A15A16 B1B1B2B2 B3B4 B5B6 B7B8 Central Repository 10 sec15 sec12 sec 17 sec 15+20 sec17+23 sec Total Propagation Delay = 40 sec. * The picture is from the site, http://blogs4brownback.files.wordpress.com

7. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 7 Good Thought Reduce data transmission more and remove the propagation delay. Leaf 1Leaf 2 X1X1 X2X2 Leaf 3Leaf 4 X3X3 X4X4 Intermediate 1Intermediate 2 B1B1 B2 B5B7 B6B8 Central Repository 12 sec17 sec14 sec 19 sec 12+5 sec 19+5 sec How about this (MAX DELAY= 24 sec.) B3 B4 B5 B6 B7 B8 B1B1B3 B2B4 17+5 sec 14+5 sec No wait 1234 5678 9101112 13141516 17181920 21222324 25262728 29303132 1234 5678 Leaf 1Leaf 2 X1X1 X2X2 Leaf 3Leaf 4 X3X3 X4X4 Intermediate 1Intermediate 2 A1A1A2A2 A3A4 A5A6 A7A8 A9A10 A11A12 A13A14 A15A16 B1B1B2B2 B3B4 B5B6 B7B8 Central Repository 10 sec15 sec12 sec 17 sec 15+20 sec 17+23 sec Current (MAX DELAY= 40 sec.) Wait for Arrivals 9101112 13141516 17181920 21222324 25262728 29303132

8. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 8 Difficulties How can we make both types of data reduction simultaneously on leaf nodes? Data sharing among sensors requires additional data transmissions. Peer-to-Peer (P2P) Data Sharing Additional Communication burden… Sensor’s memory: only 4KB Data RAM 512KB Flash memory

9. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 9 Our method: overview Our method consists of periodic collaboration and data aggregation. 1. Forward Collaboration 2. Backward Collaboration3. Data aggregation Phase 1: Collaboration (Periodic)Phase 2: Data reduction Leaf 1 5678 Leaf 2Leaf 3 1234 1314 1516 9 101112 2122 2324 17181920 56 134 15 9 1112 22 23 171920 Central Repository 56 134 15 9 1112 22 23 171920 6 14 15 9 22 23 Central Repository Leaf 1 5678 Leaf 2Leaf 3 1234 6 14 1314 1516 9 101112 15 9 2122 2324 17181920 22 23 6 14 15 9 22 23 Analysis and Decision Central Repository Leaf 1Leaf 2Leaf 3 6 14 15 9 22 23 {1,4,6}{9,15}{22,23} Indices of the chosen

10. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 10 Local data reduction Global data reduction Forward Collaboration in detail After tree-based collaboration, we can identify the minimal set of data to be transferred. 1234 5678 Leaf 1Leaf 2 X1X1 X2X2 Leaf 3Leaf 4 X3X3 X4X4 Intermediate 1Intermediate 2 A1A1A2A2 A3A4 A5A6 A7A8 A9A10 A11A12 A13A14 A15A16 A1A1A6 A2A8 A9A14 A15A12 Central Repository 9101112 13141516 17181920 21222324 25262728 29303132 1 2 3 4 5 6 A1 A2 A3 Wavelet Transform PCA (Principal Component Analysis) 1.2 Projection-based data reduction A1 A2 Ridge Regression Lasso Regression Principal Variable A4 A1 A2 A3 A4 1.1 Variable selection (Subset selection) A1A6 A15A8 No dependent variable

11. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 11 Formulation for variable selection For aggressive data reduction, we added the number of the selected elements as a penalty term to the original objective function of “principal variable” method. Key Idea Partial covariance matrix of T given S is a measure of the variation that cannot be captured by S. If the total sum of eigenvalues of is smaller, T is close to be deterministic given S. Find T and S so as to minimize Formulation where If the partial covariance is enough small, T is highly dependent of S. That is, If we know only S, we can almost determine T. A1 A2 A1 A2 A3 A4 S (Selected) T (Not selected)

12. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 12 Solution Procedure for variable selection: Iterative greedy search Iteration1: Initial Settings S= Empty set T= 1234567 Generate Subsets 1 SUBSET 1 = 2 SUBSET 2 = 3 SUBSET 3 = 7 SUBSET 7 = 6 SUBSET 6 = 5 SUBSET 5 = 4 SUBSET 4 = Evaluate each subset Evaluate the criterion for each subset generated C(SUBSET 1) =0.64 C(SUBSET 2) = 0.42 C(SUBSET 3) = 0.35 C(SUBSET 4) = 0.41 C(SUBSET 5) = 0.51 C(SUBSET 6) = 0.20 C(SUBSET 7) = 0.17 Choose the best subset S= SUBSET 1 = T = 1 234567 Iteration1: Initial Settings S = Empty set T= 1234567 Generate Subsets 1 SUBSET 1 = 2 SUBSET 2 = 3 SUBSET 3 = 7 SUBSET 7 = 6 SUBSET 6 = 5 SUBSET 5 = 4 SUBSET 4 = Evaluate each subset Evaluate the objective value for each subset generated C(SUBSET 1) =0.11 C(SUBSET 2) = 0.42 C(SUBSET 3) = 0.35 C(SUBSET 4) = 0.41 C(SUBSET 5) = 0.51 C(SUBSET 6) = 0.20 C(SUBSET 7) = 0.17 Choose the best subset S= SUBSET 1 = T = 1 234567 Iteration 2: Initial Settings S= T= 1 234567 Generate Subsets 1 SUBSET 1 = 1 SUBSET 2 = 1 SUBSET 3 = 1 SUBSET 6 = 1 SUBSET 5 = 1 SUBSET 4 = Evaluate each subset Evaluate the criterion for each subset generated C(SUBSET 1) = 0.88 C(SUBSET 2) = 0.76 C(SUBSET 3) = 0.69 C(SUBSET 4) =0.92 C(SUBSET 5) = 0.82 C(SUBSET 6) = 0.79 Choose the subset of the biggest criterion value S = SUBSET 4 = T= 1 234 5 67 2 4 3 7 6 5 Iteration 2: Initial Settings S= T= 1 234567 Generate Subsets 1 SUBSET 1 = 1 SUBSET 2 = 1 SUBSET 3 = 1 SUBSET 6 = 1 SUBSET 5 = 1 SUBSET 4 = Evaluate each subset Evaluate the objective value for each subset generated C(SUBSET 1) = 0.11 C(SUBSET 2) = 0.10 C(SUBSET 3) = 0.09 C(SUBSET 4) =0.06 C(SUBSET 5) = 0.11 C(SUBSET 6) = 0.10 S = SUBSET 4 = T= 1 234 5 67 2 4 3 7 6 5 Example Add one from T to S Choose the best subset

13. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 13 Case study: real data We split a signal into 14 subgroups to represent multi-sensor data. 224 data points per cycle amplitude 528 sample signals from normal structure 378 sample signals from abnormal structure Signal sampling from a vibration sensor Randomly split to 14 subgroups Sensor 1 Sensor 2 Sensor 14 16 data points

14. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 14 Comparison of three configurations For comparison, we setup the following three experiment configurations. Central repositoryCentral repository (PCA)Central repository (No reduction) I: No data reductionII: In-network data aggregationIII: Our method Highest data reduction capability No information loss No time delay High data reduction Little information loss

15. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 15 Results Our method reduced the amount of data transmissions while it kept most of the features required to detect process faults. Data reduction rateInformation preserving capabilities The original dimension = 224 Our method reduced to 5.1 (on average) dimensions against 15.6 dimensions of individual sensor-based reduction. Our method’s fault detection errors are the best in False Alarm, similar to the best in Miss Detection. Our method are keeping important features with relatively small data transmission. Configuration Data reduction rate Amount of data transmission False AlarmMiss detection I. No data reduction -2240.05340.0417 II. In-network data aggregation 99.55%15.60.05220.0440 III. Our method 97.72%5.10.03430.0588

16. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 16 Summary The major benefits from our method include high power efficiency of sensor’s operation and responsiveness to the outside changes. Tree-based collaboration Apply different types of data reduction methods for leaves and intermediates New collaborative data reduction Reduce all data redundancy on leaves Do nothing on intermediates No propagation time delay Higher data reduction degree than local data reduction Less communication burden for collaboration Periodic execution of collaboration phase Separate collaboration from data reduction

17. Advanced Metrology Lab., Texas A&M UniversityPresented by Chiwoo Park on Aug. 25, 2008 17 Thank you for attention.