1. 1 Mobility-assisted Spatiotemporal Detection in Wireless Sensor Networks Guoliang Xing 1 ; JianpingWang 1 ; Ke Shen 3 ; Qingfeng Huang 2 ; Xiaohua Jia 1 ; Hing Cheung So 1 1 City University of Hong Kong 2 Palo Alto Research Center (PARC) Inc. 3 Michigan State University

2. 2 Outline Motivation Problem formulation Optimal movement scheduling Simulations Conclusion

3. 3 Mission-critical Target Detection Stringent Spatiotemporal QoS requirements –High detection probability of any target, e.g., 90% –Low false alarm rate, e.g., 5% –Bounded detection delay, e.g., 20s Network and environmental dynamics –Death of nodes (battery depletions, attacks…) –Changing noise levels and target profiles

4. 4 State of the Art Over-provisioning of sensing capability –Careful advance network planning –Dense node deployment –Incremental redeployments High (re)-deployment cost in order to deal with network and environmental dynamics

5. 5 Mobility-assisted Target Detection Mobile sensors collaborate with static sensors in target detection –Achieve higher signal-to-noise ratios by moving closer to possible targets –Reconfigure sensor coverage dynamically target

6. 6 Mobile Sensor Platforms Limitations –Low movement speed (0.1~1 m/s) –High power consumption (~60 W for PackBot) PackBot @ iRobot.com Robomote @ USC Koala @ NASA

7. 7 Overview of Our Approach Data-fusion-based detection model for collaboration between mobile and static sensors Optimal sensor movement scheduling algorithm –Minimizes the moving distance of sensors –Meets spatiotemporal QoS requirements: high detection probability, low false alarm rate, and bounded detection delay Simulations based on real data traces of target detection

8. 8 Outline Motivation Problem formulation Optimal movement scheduling Simulations Conclusion

9. 9 Signal and Noise Models Target's acoustic energy decays quadratically with distance Noise energy follows the Normal distribution Sensor reading = decayed target energy + noise energy Plotted based on real acoustic sensor data traces in military vehicle detection

10. 10 Fusion-based Detection Model All readings in a cluster are summed and compared with a threshold η sensor reading distribution detection threshold noise energy distribution energy false alarm rate detection probability sensor reading distribution False alarm rate: P F = 1-X n (n· η ) Detection prob.: P D = 1 – X n (n·η - Σ W(d i )) X n – CDF of Chi-squre distribution W(d i ) – Energy measurement of sensor d i from target

11. 11 A Two-phase Detection Scheme First phase – static detection –All sensors send readings to cluster head –Cluster head makes a detection decision, if positive, starts the 2 nd phase Second phase – movement scheduling –Mobile sensors move toward the possible target according to a movement schedule –Cluster head makes the final detection decision First phase – static detection –All sensors send readings to cluster head –Cluster head makes a detection decision, if positive, starts the 2 nd phase Second phase – movement scheduling –Mobile sensors move toward the possible target according to a movement schedule –Cluster head makes the final detection decision

12. 12 Problem Formulation Find two detection thresholds and a movement schedule –Minimizes the expected moving distance of sensors –Detection prob. ≥ α, false alarm rate ≤ β, detection delay ≤ T target Example movement schedule ( sensors are assumed to move at steps): M1: t 0 - one step, t 3 - two steps … M2: t 1 - one step, t 2 - one step… M3: t 1 - two steps, t 2 - one step… M1 M2 M3

13. 13 Y – target appearance probability S – movement schedule |S| – total number of steps Steps(η 1, η 2, S) – expected total num of steps Find η 1, η 2, and schedule S to minimize Constraints: –P D1 ·P D2 ≥ α –P F1 ·P F2 ≤ β –Moving distance of any sensor in schedule S ≤ T x speed Problem Formulation Contd. Steps(η 1, η 2, S) = [ Y·P D1 + (1-Y)·P F1 ) ] x |S| the probability that sensors move

14. 14 Outline Motivation Problem formulation Optimal movement scheduling Simulations Conclusion

15. 15 Structure of Optimal Solution For two schedules S and S', if |S| = |S'| and E(S) ≥ E(S'), we can find η 1, η 2, η 1 ', η 2 ', such that Steps(η 1, η 2, S) ≤ Steps(η 1 ', η 2 ', S') –E(S) is the total energy measured by sensors Implications –Detection thresholds can be found if a schedule is given –Optimal schedule maximizes the sum of energy readings

16. 16 Examples of Optimal Schedules Assume that all sensors can move one step every second, and detection delay is T seconds Case 1: only one step allowed –Opt schedule: move B/C one step at time zero Case 2: two steps allowed –Schedule I: move B and C one step at time zero –Schedule II: move A two steps at time zero A C B All move combinations must be considered to find the optimal schedule! sample T-2 second sample T-1 second

17. 17 If a sensor moves in the 2 nd phase, it moves continuously before a stop –Num of move combinations is limited Dynamic programming algorithm –E(j,h) : total energy measured by sensor 1…j with total num of h steps –e(h j ) : total energy measured by sensor j Finding the Optimal Schedule E(j,h) = max { E(j-1,h-h j ) + e(h j ) }

18. 18 Simulations Public dataset of detecting military vehicles [Duarte04] Target and noise energy models are estimated from training data set Sensors are randomly deployed in a 50×50m 2 field

19. 19 Performance Results Total 6 sensors are deployed MD-random1: randomly choose one sensor and for next step MD-random2: randomly choose one sensor and moves to the target False alarm rate (%) Detection Prob. (%) Requested Detection Prob. (%) Average Moving Steps Mobility improves detection prob. by 20~40%!

20. 20 Conclusion Proposed a two-phase target detection model based on data fusion Developed an optimal sensor movement scheduling algorithm –Minimizes the expected moving distance of sensors –Meets spatiotemporal QoS requirements Conducted simulations based on real data traces of target detection