Localized Fault-Tolerant Event Boundary Detection in Sensor Networks

1. Localized Fault-Tolerant Event Boundary Detection in Sensor Networks Min Ding, Dechang ChenΨ, Kai Xing and Xiuzhen Cheng Computer Science, The George Washington UniversityΨUniformed Services University of the Health Sciences IEEE INFOCOM 2005 Speaker: Shin-Wei Ho

2. Outline • Introduction • Notations and Network Model • Localized Faulty Sensor Detection • Localized Event Boundary Detection • Simulation • Conclusion

3. Introduction • In some application, the detection of the event boundary may become more important than the detection of the entire event region. • Forest fire • Chemical spills

4. Introduction(cont.) • Individual sensor reading is not reliable. • faulty sensor measurements • Filtering out faulty readings and transmitting only the boundary information to the base station can save energy.

5. Introduction(cont.) • Sensor readings may be faulty • Hardware crash • Security attack • Environment disturbance • A typical event region detection may fail due to the lack of meaningful sensor readings. • 0/1 decision predicate.

6. Introduction(cont.) • Both faulty sensors and sensors in an event region can generate “abnormal readings”. • Event • Most neighboring sensors observe the same phenomenon. • Failure • Reading is geographically independent.

7. Notations and Network Model • N sensors • bxb squared field in the two dimensional Euclidean plane • S: the set of all sensors • R: the radio range of sensors • xi: the reading of the sensor Si

8. Localized Faulty Sensor Detection • N(Si): a closed neighborhood of the sensor Si. • The difference of Si • medi should not be replace by the mean • The sample mean can not represent well the “center” of a sample

9. Localized Faulty Sensor Detection(cont.) • Consider another bounded closed set N∗(Si) ⊂ R2 that contains Siand additional n − 1 sensors. • This set N∗(Si) also represents a neighborhood of Si.

10. Localized Faulty Sensor Detection(cont.)

11. Localized Faulty Sensor Detection(cont.) • the n sensors in N∗(Si) • S1, ···, Si, ···, Sn • If diis extreme in D = {d1, ··· , di, ··· , dn}, Siwill be treated as a faulty sensor.

12. Localized Faulty Sensor Detection(cont.) • The sample mean and sample standard deviation of the set D :

13. Localized Faulty Sensor Detection(cont.) • Standardize the dataset D to obtain {y1, ··· , yi, ··· , yn}, where

14. Localized Faulty Sensor Detection(cont.) • DECISION: If |yi| ≥ θ, treat Si as a faulty sensor. Hereθ(> 1) is a preselected number. • C1 denote the set of sensors with |yi| ≥ θ.

15. Localized Faulty Sensor Detection(cont.) • ALGORITHM 1 • 1) Construct {N} and {N∗}. For each sensor Si, perform the following steps. • 2) Use {N(Si)} and equation (1) to compute di for sensor Si. • 3) Use {N∗(Si)} and equation (2) to compute yi for sensor Si. • 4) If |yi| ≥ θ, where θ > 1 is predetermined, assign Si to C1.

16. Localized Faulty Sensor Detection(cont.)

17. Localized Event Boundary Detection • A sensor Si ∈ C1 may be close to the event boundary but is not faulty. • When Si close to the boundary can not be detected, we should select a special neighborhood NN(Si) • such that di, compared with d values from surrounding neighborhoods, is as extreme as possible.

18. Localized Event Boundary Detection(cont.) • There are many options in doing this. • Random Bisection • Random Trisection • Consider an Sifrom the set S − C1.

19. Localized Event Boundary Detection-- Random Bisection

20. Localized Event Boundary Detection-- Random Trisection

21. Localized Event Boundary Detection(cont.) • Let C2 denote the set of all the sensors with |yi| ≥ θ.

22. Localized Event Boundary Detection(cont.) • The set C1 is expected to contain faulty sensors. • also contains some sensors near the boundary that are not faulty • The set C2 is expected to contain sensors close to the event boundary. • also contains some sensors that are not close to the boundary

23. Localized Event Boundary Detection(cont.) • Find C3 • For a sensor Si ∈ C1∪C2, draw a closed disk D(Si; c) with radius c centered at Si. • Si is expected to be close to the boundary if D(Si; c) contains at least one sensor from C2 that is different from Si.

24. Localized Event Boundary Detection(cont.) • ALGORITHM 2 • 1) Construct {N} and {N∗}. Apply Algorithm 1 to produce the set C1 • 2) For each sensor Si ∈ S−C1, perform the following steps. • Obtain NN(Si) and update difrom step 1) to the new di from NN(Si), keeping unchanged all the other d values from N∗(Si) obtained in step 1). • recomputting yi. If |yi| ≥θ, assign Sito set C2 otherwise, treat Sias a normal sensor. • 3) Obtain C3

25. Localized Event Boundary Detection(cont.)

26. Simulation • MATLAB • 1024 nodes • 32x32 units with one sensor randomly placed within each unit grid. • Normal sensor readings are drawn from N(μ1, σ12), μ1=10, σ1=1 • Event sensor readings are drawn from N(μ2, σ22), μ2=30, σ2=1 • N = N∗, and N(Si) contains all one-hop neighbors of Si.

27. Simulation-- Faulty sensor detection accuracy vs. p

28. Simulation-- False alarm rate in faulty sensor detection vs. p

29. Simulation-- Degree of fitting vs. network density when p = 0

30. Simulation-- False detection rate vs. network density when p = 0

31. Simulation-- Degree of fitting vs. p when density = 30

32. Simulation-- False detection rate vs. p when density = 30

33. Conclusion • The ideas in detecting faulty sensors and event boundaries can be extended to multi-modality sensor networks. • Future Works • An adaptive threshold that better fits the application context

34. Thank you !