1 / 24

Collaborative Target Detection in WirelessSensor Networks with Reactive Mobility

Rui Tan 1 Guoliang Xing1 Jianping Wang1

Hing Cheung So2

1Department of Computer ScienceCity University of Hong Kong

2Department of Electronic EngineeringCity University of Hong Kong

2 / 24

Outline

1. Motivation

2. Preliminaries

3. Problem Formulation

4. Near-optimal Solution

5. Performance Evaluation

3 / 24

Challenges for Mission-critical SensingApplications

• Stringent QoS requirements• Target detection/tracking, security surveillance• High detection probability• Low false alarm rate• Bounded detection delay

• Unpredictable network dynamics• Coverage holes caused by death of nodes

• Changing physical environments• Different spatial distribution of events

4 / 24

Exploit Mobility in Target Detection

• Sense better signal by moving sensors closer to targets

• Adapt to the changes of network condition and physicalenvironments

Example: fire detection

5 / 24

Mobile Sensor Platforms

Robomote @ USC Koala @ NASA GRC PackBot @ iRobot.com

Challenges• Low movement speed (0.1 ∼ 2m/s)

• Increase detection latency

• High manufacturing cost• A small number of mobile sensors available

• High energy consumption• Locomotion consumes much higher power than wireless

communication

6 / 24

Overview of Our Approach

• Data-fusion based target detection• Explore the collaboration between mobile and static

sensors

• Near-optimal sensor movement scheduling algorithm• Reduce moving distance of sensors• Satisfy QoS requirements:

• Low false alarm rate• High detection probability• Bounded detection delay

• Performance evaluation using real data traces

7 / 24

Outline

1. Motivation

2. Preliminaries

3. Problem Formulation

4. Near-optimal Solution

5. Performance Evaluation

8 / 24

Signl Energy Model and Noise Model

00.5

11.5

22.5

33.5

0 0.2 0.4 0.6 0.8 1

Sig

nalE

nerg

y(×10−

2)

1/distance2 (×10−2)

0

2

4

6

8

10

-4 -3 -2 -1 0 1 2 3 4

Occ

urre

nce

Noise energy (×10−4)

• Plotted using real data traces from DARPA SensIT experiments

e(x) =initial target energy

x2 noise ∼ N(µ, σ2)

Measurement = e(x) + noise

9 / 24

Single-sensor Detection Model

• Local decision of sensor i

=

{

1 if ei ≥ λ0 if ei < λ

• The false alarm rate ofsensor i

P iF = Q

(

λ − µ

σ

)

• The detection probability

P iD = Q

(

λ − µ − e(xi)

σ

)

CCDF: Q(x) = 1 −

R x−∞

φ(t)dt

eiλµ

noise

µ + e(xi)

measurement

• closer to the target, higher PD

10 / 24

Decision Fusion Model

• System detection decision

Majority Rule:{

1 if more than n/2 sensors decide 10 otherwise

• The system false alarm rate

PF = Q

n2 −

∑ni=1 P i

F√

∑ni=1 P i

F +∑n

i=1(PiF )2

• The system detection probability

PD = Q

n2 −

∑ni=1 P i

D√

∑ni=1 P i

D +∑n

i=1(PiD)2

11 / 24

Outline

1. Motivation

2. Preliminaries

3. Problem Formulation

4. Near-optimal Solution

5. Performance Evaluation

12 / 24

Target Detection with Mobile Sensors

• Long distance movement can• quickly deplete the battery of a mobile node• disrupt the network topology

• Problem formulation: minimize the moving distance ofsensors subject to

• PF ≤ α, e.g., 5%• PD ≥ β, e.g., 95%• Average detection delay ≤ D, e.g., 15s

13 / 24

A Two-phase Detection Approach

target

e1 <λ2

e1+e2 >λ2

terminate!e1 e2

• 1st phase: each sensor makes local decision by e0 ≷ λ1

• If the system decision is 1, the 2nd phase is initiated

• 2nd phase: mobile sensors move and periodically sense• A sensor terminates the detection and decides 1 if

e1 + e2 + · · · + ej ≥ λ2

• Make final detection decision

14 / 24

Advantages of Reactive Mobility

• Sensors move reacting to positive decision in the 1st phase

• Avoid unnecessary movement by consensus check in the1st phase

• Reduce the probability of movement when the target isabsent

• Terminate moving once enough signal energy is obtained• If a loud target appears, mobile sensors can terminate

movement quickly

15 / 24

Problem Formulation

Objective: Find the two detection thresholds λ1, λ2 and amovement schedule to minimize the expectedmoving distance:

Pa · PD1 · L1 + (1 − Pa) · PF1 · L0

correct detection false alarm

• Pa: the probability that a target appears• L0(L1): the expected moving distance when

the target is absent (present)

Constraints:• PF1 · PF2 ≤ α• PD1 · PD2 ≥ β

16 / 24

Outline

1. Motivation

2. Preliminaries

3. Problem Formulation

4. Near-optimal Solution

5. Performance Evaluation

17 / 24

The Structure of Optimal Solution

• Theorem 1: Total moving distance decreases with the systemdetection probability in the 2nd phase, i.e., PD2

• Linear approximation using the 1st order Taylor expansion

Q−1(PD2) =n2 −

∑ni=1 P i

D2√

∑ni=1 P i

D2 −∑n

i=1(PiD2)

2

≃ − 2√n

n∑

i=1

P iD2 + constant

PD2 increases with∑n

i=1 P iD2 with high probability

• Simplified problem formulation

• Maximize∑n

i=1 P iD2 subject to the constraints:

PF1 · PF2 ≤ α PD1 · PD2 ≥ β

18 / 24

The Structure of Optimal Solution (Cont.)

• Combination of sensor movement is exponential• Finding maximized

∑ni=1 P i

D2 is exponential

• Theorem 2: In the optimal solution, each mobile sensormove in parallel and consecutively

• Implication•

∑ni=1 P i

D2 can be maximized by DynamicProgramming

19 / 24

Dynamic Programming: An Example

• Two sensors: A and B

• Budget: two sensor moves

• Suppose:

PAD(0) = 0.40, PA

D(1) = 0.50, PAD(2) = 0.60

PBD (0) = 0.46, PB

D (1) = 0.60, PBD (2) = 0.67

PAD(2) + PB

D(0) = 1.06A

B

PAD(0) + PB

D(2) = 1.07A

B

PAD(1) + PB

D(1) = 1.10A

B

This procedure can be implemented via Dynamic Programming

20 / 24

Outline

1. Motivation

2. Preliminaries

3. Problem Formulation

4. Near-optimal Solution

5. Performance Evaluation

21 / 24

Simulation Settings

• Data: public dataset ofDARPA SensIT experiment

• Targets: AmphibiousAssault Vehicles (AAVs)

• Sensors are randomlydeployed in a 50m×50mfield

75m

?

6

50m

?

6

50m� -

N

N

N

N

cc

c

cc

sCCO

s6

sXXy

sAAK

sCCO

bfixed sensorrmobile sensor

N surveillance spot

��

��

c c

� ���

22 / 24

Impact of The Number of Mobile Sensors

• Total 12sensors

• 10% to 35%performanceimprovementby 6 mobilesensors 30

40

50

60

70

80

90

100

0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

De

tect

ion

prob

abi

lity

(%)

False alarm rate (%)

static1/2 mobileall mobile

23 / 24

Conclusions

• Propose a two-phase detection approach• Reactive mobility• Collaboration between static and mobile sensors

• Develop a near-optimal movement scheduling algorithm

• Provide insights into detection system design• Efficient movement schedule of a small number of mobile

sensors significantly boost the detection performance

24 / 24

Thanks!