1. IIT Bombay Tracking Dynamic Phenomena: Sensor Networks to the Rescue Krithi Ramamritham Dept of Computer Sc. & Engg. Indian Institute of Technology Bombay www.cse.iitb.ac.in/~krithi

2. 2 More and more of the information we consume is coming from sensors in the field…….

3. 3 Traffic Management # Vehicles expected on Pacifica Highway above threshold ?

4. IIT Bombay Early Warning System For Landslide Prediction using Sensor Networks

5. 5 Major Landslides Major Landslide Prone Areas –Himalayas –Western ghats

6. 6 Wireless agri sensors for crop disease forecasts wind speed wind direction air temperature relative humidity solar radiation evaporation rate Upload updates periodically

7. 7 Sensornet Apps…. redwood forest microclimate monitoring smart cooling in data centers http://www.hpl.hp.com/research/dca/smart_cooling/ condition-based maintenance and More… bridge structural integrity

8. 8 Types of Sensors Weather Vibration –2 or 3 axis accelerometers Tracking –Microphone (for ranging and acoustic signatures) –Magnetometer GPS RFID Reader EEPROM  512K off chip, 32K on chip Many megabytes in the future Many megabytes in the future  Writes at disk speeds, reads at RAM speeds  Interface : random access, read/write pages

9. 9 Power Consumption and Lifetime Power typically supplied by a small battery Power typically supplied by a small battery  1000-2000 mAH Lifetime, power consumption varies by application Lifetime, power consumption varies by application  Mica2 processor: 5mA active, 1 mA idle, 5 uA sleeping 5mA active, 1 mA idle, 5 uA sleeping  CC1000 Radio 5 mA listen, 10 mA xmit/receive, ~20mS / packet 5 mA listen, 10 mA xmit/receive, ~20mS / packet  Sensors 1 uA -> 100’s mA, 1 uS -> 1 S / sample 1 uA -> 100’s mA, 1 uS -> 1 S / sample

10. 10 Queries over Sensors Query sensornet through a (remote) base station Query sensornet through a (remote) base station Sensor nodes have severe resource constraints Sensor nodes have severe resource constraints  Limited battery power, memory, processor, radio range…  Communication is the major source of battery drain  transmitting a single bit of data is equivalent to 100s of instructions base station (root, coordinator…) http://www.intel.com/research/exploratory/motes.htm

11. 11 Local storage policy: store data on nodes + simple +data collection is cheap -queries are flooded, costly base queries

12. 12 In-network Computation in Trees Goal is for root to compute a function of data at leaves Goal is for root to compute a function of data at leaves Trivial solution: push all data up tree and compute at base station Trivial solution: push all data up tree and compute at base station – Strains nodes near root: batteries drain, disconnecting network – Very wasteful: no attempt at saving communication Can do much better by “In-network” query processing – Simple example: computing max – Each node hears from all children, computes max and sends to parent (each node sends only one item)

13. 13 Master Control Room Control Room at nearby station Schematic Mumbai Railway Station GPRS Sensor node with strain gauge Cluster head With GPRS

14. 14 Distributed Processing BS-1 CH-1 CH-(m-1)CH-3 CH-4 CH-2 BS-2 12 n 12n 12 n 12n 12n CH-m 12n In-network aggregation Probabilistic estimation

15. 15 Statistical Techniques Approximations, summaries, and sampling based on statistics and statistical models Approximations, summaries, and sampling based on statistics and statistical models Applications: Applications:  Limited bandwidth and large number of nodes -> data reduction -> data reduction  Lossiness -> predictive modeling  Uncertainty -> tracking correlations and changes over time  Physical models -> improved query answering

16. 16 Monitoring Dynamic Physical Phenomena A disaster management team is interested in tracking an oil spill with the help of sensors. Sensors track the perimeter of” spill. Will the spill hit the shore before 5pm?

17. 17 Remote/Range Sensing of Boundaries Solution Space Characteristics 1.Static sensors 2.Range sensing 3.Approximate location 4.Dynamic Boundary

18. 18 Sensing Model

19. 19Assumptions Random distribution of n static sensors Random distribution of n static sensors Sensors know their own locations Sensors know their own locations Boundaries are smooth : belong to C 2 class Boundaries are smooth : belong to C 2 class  Second derivative continuous No abrupt changes of boundary along x axis No abrupt changes of boundary along x axis Errors in sensor observations independent and mean zero Errors in sensor observations independent and mean zero Boundaries are curve instead of contour Boundaries are curve instead of contour

20. 20 Problem Definition Compute confidence band at all times such that actual boundary all times such that actual boundary lies within the interval with a lies within the interval with a confidence confidence Width of the intervals should not exceed a certain value δ exceed a certain value δObjective Update with minimal communication overheads communication overheads to increase sensornet lifetime to increase sensornet lifetime Confidence band should have low loss of coverage low loss of coverage

21. 21Overview Motivation Motivation Problem: Estimation of boundary Problem: Estimation of boundary Solution Approach Solution Approach  Spatial Estimation through in-network aggregation  Estimate at minimal locations along the boundary  Tracking dynamics using Temporal Correlation Experimental Results Experimental Results Future Work Future Work

22. 22 Problem Formulation Static Boundary Tracking n, δ, ε m ON nodes, k conf. intervals Dynamic Boundary Tracking n, δ conf. band Entire Boundary Tracking k, ε Metric : Loss of Coverage

23. 23 Estimate Boundary and Confidence Interval at any x  Gather observations from sensors from sensors within h-neighborhood within h-neighborhood  Use Spatial Aggegation to estimate CIs for entire boundary δ h x h

24. 24 continuous, bounded and symmetric continuous, bounded and symmetric Non-negative Non-negative k has support [-1, 1] k (-1) = k (1) = 0 Kernel Properties k has support [-1, 1] k (-1) = k (1) = 0 Kernel Properties Normalized – integrates to one Normalized – integrates to one Defines shape of weights Defines shape of weights Boundary at any x x i values: predictor vars y i values: response vars Relationship modeled as non-parametric regression

25. 25 δ h x h

26. 26 Kernel Smoothing d(x) is estimated as avg. of d(x) is estimated as avg. of kernel weighted y’s measured by sensors in a neighborhood of x kernel weighted y’s measured by sensors in a neighborhood of x Choices of weight sequences Choices of weight sequences Nadaraya-Watson [Nadaraya 1964], Nadaraya-Watson [Nadaraya 1964], Gasser-Mϋller Gasser-Mϋller what should be the neighborhood of x? what should be the neighborhood of x? weight is non-zero only if |x-x i | < h weight is non-zero only if |x-x i | < h K(u) =.75 × (1 − u 2 ) I(|u| ≤ 1).

27. 27 Optimal (h) selection Accuracy of estimation depends on h Accuracy of estimation depends on h Avg. Square Error Avg. Square Error = Bias + Variance = Bias + Variance At optimal h, bias and variance balance At optimal h, bias and variance balance Penalty function and ASE h

28. 28 Estimation of Confidence Intervals Estimation of Confidence Intervals How to find the pointwise confidence intervals of d(x)? How to find the pointwise confidence intervals of d(x)? = conditional variance of Y | X=x = conditional variance of Y | X=x Require Distributed Estimation Require Distributed Estimation

29. 29 Partial Aggregates of Partial Aggregates of sum of kernel x observation sum of kernel sum of kernel x observation sum of kernel x observation 2

30. 30 Nodes organized into clusters Nodes organized into clusters CHs form a tree CHs form a tree Nodes send Observations to Cluster Heads (CHs) Nodes send Observations to Cluster Heads (CHs) CHs perform local aggregations CHs perform local aggregations BS computes final Confidence Intervals BS computes final Confidence Intervals Data Dissemination Scheme BSBS CH S1S1 S2S2 N 1, D 1, Q 1 N 2, D 2, Q 2 CH

31. 31 Partial Aggregates of Partial Aggregates of sum of kernel x observation sum of kernel sum of kernel x observation sum of kernel x observation 2 BSBS CH S1S1 S2S2 N 1, D 1, Q 1 N 2, D 2, Q 2 CH

32. 32 Satisfying δ Criterion p1p1 p2p2 p3p3.44.6.56.49 Which nodes to turn ON/OFF? Which nodes to turn ON/OFF? Any order of preference for nodes? Any order of preference for nodes? CH maintains score list CH maintains score list  score = # boundary points in h-neighborhood with width > δ Should maintain connectivity Should maintain connectivity.46 var (avg)= var/n

33. 33 Estimation using Real Sensors Experiment using robot with range sensors Error Variance changes with x Confidence Intervals cover the boundary if # Obs > 100 if # Obs > 100

34. 34 Confidence Intervals Increase in confidence implies wider interval Increase in confidence implies wider interval Increase in noise variance implies wider interval Increase in noise variance implies wider interval

35. 35 DBTR: Dynamic Boundary Tracking Can periodic update scheme work? Assumption: constant mean velocity and Gaussian Noise and Gaussian Noise Process model to track the dynamicsProcess model to track the dynamics Linear dynamics can be modeled usingLinear dynamics can be modeled using Kalman Filter Kalman Filter state = distance, velocitystate = distance, velocity        ),( ),( ),( i i i txd txd txs 

36. 36 Tracking the Dynamics new state = F x old state + noise new state = F x old state + noise new position = old position + velocity x t s + noise new position = old position + velocity x t s + noise new velocity = old velocity + noise new velocity = old velocity + noise observation = H x state + noise observation = H x state + noise

37. 37 When to update Confidence Intervals? When to update Confidence Intervals? Spatial estimation involves communication overheads but gives more accurate information Spatial estimation involves communication overheads but gives more accurate information Temporal estimation gives the changes in boundary Temporal estimation gives the changes in boundary at specific location at specific location Minimize the frequency of updates Minimize the frequency of updates Update the boundary only if it has changed by cδ Update the boundary only if it has changed by cδ

38. 38 Block Diagram for TE & SE Block Diagram for TE & SE Use Spatial Estimation as a feedback Use Spatial Estimation as a feedback Feedback improves the accuracy of Temporal Estimation Feedback improves the accuracy of Temporal Estimation PredictionUpdate Regression changed by cδ yes no

39. 39 Block Diagram for TE & SE Block Diagram for TE & SE Use Spatial Estimation as a feedback Use Spatial Estimation as a feedback Feedback improves the accuracy of Temporal Estimation Feedback improves the accuracy of Temporal Estimation PredictionUpdate Regression changed by cδ yes no

40. 40 Heuristic for Optimal k Variance indicates spatial variation Variance indicates spatial variation Estimate at more locations in the region of high variance Estimate at more locations in the region of high variance

41. 41 Simulation Results Randomly distributed nodes in 100x100 field Randomly distributed nodes in 100x100 field TOSSIM as well as MATLAB TOSSIM as well as MATLAB Real sensor traces Real sensor traces # of boundary points k = 10 to 50 # of boundary points k = 10 to 50 σ 2 =.5 to 2.0 σ 2 =.5 to 2.0 Metrics Metrics – Accuracy of Estimation - LOC – Communication Overheads

42. 42 Loss of Coverage DBTR better than individual techniques DBTR better than individual techniques Spatial Est. better than temporal for lower δ Spatial Est. better than temporal for lower δ Temporal Est. improves for Temporal Est. improves for larger δ larger δ

43. 43 Loss of Coverage vs. change in y Better Coverage with more frequent updates Increasing δ helps in improving coverage

44. 44 Communication Overhead ( Dynamics of Boundary ) δ = 1.2 Velocity # of updates 0.5 unit/sec 44 44 1 unit/sec 64 64

45. 45 Communication Overhead (Points changing at different velocity) DBTR adaptively updates based on velocity of boundary. More Communication overheads for higher velocity

46. 46 Comparison with a Periodic Scheme DBTR has less communication overheads Has comparable loss of coverage

47. 47 Heuristic for Minimal # Estimation Pts Estimating at more locations reduces Interpolation Error Prediction Error Function shows the same trend as LOC with variation of k Increase estimation points until Prediction Error Function stabilizes

48. 48 Communication Overhead – (Spatial) Distributed scheme does not change much with network size scalable solution Value of h reduces with network density Distributed performs ~20-50% better than centralized for k = 10

49. 49 Verification of Heuristic for deriving k Prediction Error Function stabilizes at k=12 LOC < 4 % at k=12

50. 50 Effect of Variance of Sensing Angle on LOC

51. 51Comparisons Sensing Model ON/OFFAccuracy Disseminatio n Scheme Characteriz ation of boundary Nature of Boundary OursRange/remoteyes Confidence Intervals Cluster-based Non- parametric Smooth Nowa k et al. Point sensing All ON LB MSE Hierarchical Cluster based Staircase like approximation Smooth Guestr in et al. Point sensing All ON Depends on the # of Basis functions Specialized Data Structure (Junction Tree) Parametric Field Modeling

52. 52 Conclusions & Future Work A practical low overhead strategy for tracking dynamic boundary A practical low overhead strategy for tracking dynamic boundary DBTR does not require prior knowledge about the dynamics DBTR does not require prior knowledge about the dynamics Confidence band with LOC < 2% from estimates at a few selected locations Confidence band with LOC < 2% from estimates at a few selected locations Handle situations where the boundary changes very fast Handle situations where the boundary changes very fast Strategy for estimating boundary in presence of a deadline Strategy for estimating boundary in presence of a deadline Comparison with Parametric and other Approaches Comparison with Parametric and other Approaches Boundary tracking with sensors having local sensing capability. Boundary tracking with sensors having local sensing capability.

53. 53 Future Work: Building Sensor Test-bed Real-time Dynamic 2D sensor test-bed Real-time Dynamic 2D sensor test-bed Light sensors mapping solar insolation Light sensors mapping solar insolation  Intensity  Quality Region Estimation under cloud cover Region Estimation under cloud cover Incorporates emulated motes Incorporates emulated motes Application: Optimal power generation for air-borne PV powered surveillance Application: Optimal power generation for air-borne PV powered surveillance

54. 54Acknowledgment Subhasri Duttagupta Subhasri Duttagupta Prof. Purushottam Kulkarni Prof. Purushottam Kulkarni Prof. Kannan M. Moudgalya Prof. Kannan M. Moudgalya Prof. Parmesh Ramanathan Prof. Parmesh Ramanathan

55. 55 Loss of Coverage vs. Conf. Level Better Coverage with higher ON nodes δ Should be higher than the error variance

56. 56References [1] K. Moore, Y. Chen, and Z. Song, “Diffusion-based path planning in mobile actuator-sensor networks (mas-net): Some preliminary results,” in Intelligent Computing: Theory and Application II. SPIE Defense and Security Symposium, 2004. [2] M. F. Fingas and C. E. Brown, “Review of Oil Spill Remote Sensing,” in Eighth Int. Oil Spill Conference, SPILLCON, 2000. [3] R. Nowak, U. Mitra, and R. Willett, “Estimating inhomogeneous fields using wireless sensor networks,” IEEE Journal on Selected Areas in Communications, vol. 22, no. 6, pp. 999-1006, 2004. [4] K. Wang and P. Ramanathan, “Collaborative sensing using sensors of uncoordinated mobility,” in Intl. Conference on Distributed Computing in Sensor Systems, June 2005. W. H¨ardle, Applied Nonparametric Regression. Cambridge University Press, 1990. G. Werner-Allen et al., “Deploying Wireless sensor Network on an Active Volcano”, IEEE Internet Computing, March/April 2006.

57. 57References E. A. Nadaraya, “On estimating regression,” Theory Prob. Appl. 10, 186-90, 1964. T. Gasser and H. G. M¨uller, “Estimating regression functions and their derivatives by the kernel method,” Scandanavian Journal of Statistics, 11, 171-85, 1984. W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “An application-specific communication protocol for wireless microsensor networks,” IEEE Transactions on Wireless Communications, vol. 1, no. 4, pp. 660-670, Oct 2002.

58. 58 Thank You!

59. 59 Smooth Boundary CIs using non-parametic regression Uses Spatial Correlation PAs sent from CHs to parents using multi-hop