Professional Documents
Culture Documents
}
where
k
= P
_
T k | X
[1:k]
_
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 9 / 22
The Classical Change Detection Problem (CDP)
A
Before Change X
(i )
k
f
0
A
After Change X
(i )
k
f
1
The Setting: Discrete time, k 0. The
change occurs at a random time T 0.
Need to detect the change
We have a prior on T, and the
observations X
(i )
k
, 1 i N, k 1
Conditionally i.i.d. in time and space
The Problem: Determine a stopping time
, such that
min E
_
( T)
+
_
s.t. P( < T)
The Solution: The optimal is given by
[Shiryaev 63]
= inf {k :
k
}
where
k
= P
_
T k | X
[1:k]
_
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 9 / 22
The Classical Change Detection Problem (CDP)
A
Before Change X
(i )
k
f
0
A
After Change X
(i )
k
f
1
The Setting: Discrete time, k 0. The
change occurs at a random time T 0.
Need to detect the change
We have a prior on T, and the
observations X
(i )
k
, 1 i N, k 1
Conditionally i.i.d. in time and space
The Problem: Determine a stopping time
, such that
min E
_
( T)
+
_
s.t. P( < T)
The Solution: The optimal is given by
[Shiryaev 63]
= inf {k :
k
}
where
k
= P
_
T k | X
[1:k]
_
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 9 / 22
Limitations of the Classical CDP
Recent motivation: event detection in wireless sensor networks
The classical model does not account for many practical aspects
Sleep-wake scheduling of sensors (cost for making observations)
Optimal detection with sleep-wake scheduling: [Infocom 2008]
Transient change: The event occurs and then disappears in a nite
time
Quickest transient change detection: [IWAP 2010]
Networked sensors
Random access packet network between the sensors and the fusion
centre [SECON 2007, ACM TOSN 2010]
Large extent network: spatial variation in the way the change (event)
aects the sensors
Distributed detection and localisation in large wireless sensor networks:
[Allerton 2009]
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 10 / 22
Limitations of the Classical CDP
Recent motivation: event detection in wireless sensor networks
The classical model does not account for many practical aspects
Sleep-wake scheduling of sensors (cost for making observations)
Optimal detection with sleep-wake scheduling: [Infocom 2008]
Transient change: The event occurs and then disappears in a nite
time
Quickest transient change detection: [IWAP 2010]
Networked sensors
Random access packet network between the sensors and the fusion
centre [SECON 2007, ACM TOSN 2010]
Large extent network: spatial variation in the way the change (event)
aects the sensors
Distributed detection and localisation in large wireless sensor networks:
[Allerton 2009]
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 10 / 22
Limitations of the Classical CDP
Recent motivation: event detection in wireless sensor networks
The classical model does not account for many practical aspects
Sleep-wake scheduling of sensors (cost for making observations)
Optimal detection with sleep-wake scheduling: [Infocom 2008]
Transient change: The event occurs and then disappears in a nite
time
Quickest transient change detection: [IWAP 2010]
Networked sensors
Random access packet network between the sensors and the fusion
centre [SECON 2007, ACM TOSN 2010]
Large extent network: spatial variation in the way the change (event)
aects the sensors
Distributed detection and localisation in large wireless sensor networks:
[Allerton 2009]
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 10 / 22
Limitations of the Classical CDP
Recent motivation: event detection in wireless sensor networks
The classical model does not account for many practical aspects
Sleep-wake scheduling of sensors (cost for making observations)
Optimal detection with sleep-wake scheduling: [Infocom 2008]
Transient change: The event occurs and then disappears in a nite
time
Quickest transient change detection: [IWAP 2010]
Networked sensors
Random access packet network between the sensors and the fusion
centre [SECON 2007, ACM TOSN 2010]
Large extent network: spatial variation in the way the change (event)
aects the sensors
Distributed detection and localisation in large wireless sensor networks:
[Allerton 2009]
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 10 / 22
Limitations of the Classical CDP
Recent motivation: event detection in wireless sensor networks
The classical model does not account for many practical aspects
Sleep-wake scheduling of sensors (cost for making observations)
Optimal detection with sleep-wake scheduling: [Infocom 2008]
Transient change: The event occurs and then disappears in a nite
time
Quickest transient change detection: [IWAP 2010]
Networked sensors
Random access packet network between the sensors and the fusion
centre [SECON 2007, ACM TOSN 2010]
Large extent network: spatial variation in the way the change (event)
aects the sensors
Distributed detection and localisation in large wireless sensor networks:
[Allerton 2009]
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 10 / 22
Collocated Nodes: Change Detection over a Network
Detection Over a Random Access Network
Nodes make periodic observations
(e.g., temperature; gas concentration
(e.g., LPG))
Measurements (or local decisions) are
sent to the fusion centre...
... over a random access network
Fusion
Centre
X
k
(1)
X
k
(2)
X
k
(N)
.
.
.
.
MAC
svc.
rate
1
2
N
sampling
.
.
.
.
Transmitter buffer
1
2
N
Maker
Decision
fusion centre
Sequencer
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 11 / 22
The Sequential Detection Problem with Network Delays
The sequencer at the fusion centre releases complete batches of
samples to the detection algorithm
K
~
t
K
~
U
T
~
U
~
t t t t t t t t
1 2 3 4 5 6 7 0
U U U U U
1 2 3 4 5
T
change time
detection time
without n/w delay
detection time
with n/w delay
Minimise detection delay s.t. a false alarm probability bound, i.e.,
min E
_
(
U T)I
{
TT}
_
subject to
Pr
_
T < T
_
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 12 / 22
Collocated Nodes: Change Detection over a Network
Decision and Network Delays Decouple
Theorem
If the sampling is periodic at rate r and the batch sojourn time process
D
b
, b 1, is stationary with mean d(r ), then
min
E
_
(
U T)I
{
TT}
_
= l (r ) (1 ) + d(r ) (1 ) +
1
r
min
E
_
K K
_
+
where l (r ) is the delay due to (coarse) sampling, and K is the index of the
rst sample after the change.
A cross-layer result: combines network performance & statistical
performance for the rst time
The problem min
E
_
K K
_
+
is the classical one
The threshold will depend on the sampling rate r
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 13 / 22
Collocated Nodes: Change Detection over a Network
Decision and Network Delays Decouple
Theorem
If the sampling is periodic at rate r and the batch sojourn time process
D
b
, b 1, is stationary with mean d(r ), then
min
E
_
(
U T)I
{
TT}
_
= l (r ) (1 ) + d(r ) (1 ) +
1
r
min
E
_
K K
_
+
where l (r ) is the delay due to (coarse) sampling, and K is the index of the
rst sample after the change.
A cross-layer result: combines network performance & statistical
performance for the rst time
The problem min
E
_
K K
_
+
is the classical one
The threshold will depend on the sampling rate r
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 13 / 22
Collocated Nodes: Change Detection over a Network
Decision and Network Delays Decouple
Theorem
If the sampling is periodic at rate r and the batch sojourn time process
D
b
, b 1, is stationary with mean d(r ), then
min
E
_
(
U T)I
{
TT}
_
= l (r ) (1 ) + d(r ) (1 ) +
1
r
min
E
_
K K
_
+
where l (r ) is the delay due to (coarse) sampling, and K is the index of the
rst sample after the change.
A cross-layer result: combines network performance & statistical
performance for the rst time
The problem min
E
_
K K
_
+
is the classical one
The threshold will depend on the sampling rate r
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 13 / 22
Collocated Nodes: Change Detection over a Network
Optimal Sampling Rate r : Tradeo Between Network and Decision Delays
Example: = 0, p = 0.0005, = 0.01, = 0.3636, N = 10
Stability requirement N r < , the success rate
Mean detection delay vs. sampling rate r
0.01 0.015 0.02 0.025 0.03 0.035
80
100
120
140
160
sampling rate (samples per slot)
m
e
a
n
d
e
t
e
c
t
i
o
n
d
e
l
a
y
(
b
a
s
i
c
s
l
o
t
s
)
The optimal sampling rate is r = 0.029 (sampling interval = 34 slots)
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 14 / 22
Collocated Nodes: Optimal Node Energy Management
Detect an event at a random time, T
n sensors in the observation eld
Sensor i makes the measurements
X
(i )
k
, k 1
How many sensors should we turn on?
Tradeos: detection delay, false alarm
rate, battery life
Fusion
Centre
observation distribution F observation distribution F
t t t
5 6 7
t t t t t
1 2 3 4 0
T
sensor observations
1 0
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 15 / 22
Collocated Nodes: Optimal Node Energy Management
Detect an event at a random time, T
n sensors in the observation eld
Sensor i makes the measurements
X
(i )
k
, k 1
How many sensors should we turn on?
Tradeos: detection delay, false alarm
rate, battery life
Fusion
Centre
observation distribution F observation distribution F
t t t
5 6 7
t t t t t
1 2 3 4 0
T
sensor observations
1 0
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 15 / 22
Collocated Nodes: Optimal Node Energy Management
Detect an event at a random time, T
n sensors in the observation eld
Sensor i makes the measurements
X
(i )
k
, k 1
How many sensors should we turn on?
Tradeos: detection delay, false alarm
rate, battery life
Fusion
Centre
observation distribution F observation distribution F
t t t
5 6 7
t t t t t
1 2 3 4 0
T
sensor observations
1 0
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 15 / 22
Collocated Nodes: Optimal Node Energy Management
Dynamic Control of Mote Sleep-Wake Cycling
M = m
k
M = m
k+1
M = m
k+2
k k+1 k+2
D D D
k k+1 k+2
Decisions:
stop and declare change or
continue sampling
fixed number of observations
M
k
M
k+1
M
k+2
k k+1 k+2
D D D
k k+1 k+2
Decisions:
stop and declare change or
determine how many sensors
to activate
number of observations
is dynamically controlled
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 16 / 22
Collocated Nodes: Optimal Node Energy Management
The Decision Rule
At sample time k, the observation vector is X
(M
k
)
k
the vector of observations from M
k
sensors
Compute
k
= P
_
T k
X
(M
1
)
1
, , X
(M
k
)
k
_
k
= posterior probability of the event having occurred at or before k
An example
n = 10 sensors
False alarm cost = 100.0
Sensor activation cost = 0.5
Delay cost = 1 per sample
delay
f
0
N(0, 1), f
1
N(1, 1)
Optimal algorithm
Stop if
k
0.9
Else use M
k
as in the plot
0
1
2
3
4
5
0 0.2 0.4 0.6 0.8 1
O
p
t
i
m
u
m
n
o
.
o
f
s
e
n
s
o
r
s
i
n
t
h
e
w
a
k
e
s
t
a
t
e
,
M
*
Posterior probability,
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 17 / 22
Collocated Nodes: Optimal Node Energy Management
The Decision Rule
At sample time k, the observation vector is X
(M
k
)
k
the vector of observations from M
k
sensors
Compute
k
= P
_
T k
X
(M
1
)
1
, , X
(M
k
)
k
_
k
= posterior probability of the event having occurred at or before k
An example
n = 10 sensors
False alarm cost = 100.0
Sensor activation cost = 0.5
Delay cost = 1 per sample
delay
f
0
N(0, 1), f
1
N(1, 1)
Optimal algorithm
Stop if
k
0.9
Else use M
k
as in the plot
0
1
2
3
4
5
0 0.2 0.4 0.6 0.8 1
O
p
t
i
m
u
m
n
o
.
o
f
s
e
n
s
o
r
s
i
n
t
h
e
w
a
k
e
s
t
a
t
e
,
M
*
Posterior probability,
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 17 / 22
Collocated Nodes: Optimal Node Energy Management
The Decision Rule
At sample time k, the observation vector is X
(M
k
)
k
the vector of observations from M
k
sensors
Compute
k
= P
_
T k
X
(M
1
)
1
, , X
(M
k
)
k
_
k
= posterior probability of the event having occurred at or before k
An example
n = 10 sensors
False alarm cost = 100.0
Sensor activation cost = 0.5
Delay cost = 1 per sample
delay
f
0
N(0, 1), f
1
N(1, 1)
Optimal algorithm
Stop if
k
0.9
Else use M
k
as in the plot
0
1
2
3
4
5
0 0.2 0.4 0.6 0.8 1
O
p
t
i
m
u
m
n
o
.
o
f
s
e
n
s
o
r
s
i
n
t
h
e
w
a
k
e
s
t
a
t
e
,
M
*
Posterior probability,
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 17 / 22
Collocated Nodes: Optimal Node Energy Management
Comparison with a Fixed Number of Active Sensors
10
-8
10
-6
10
-4
10
-2
10
0
40 80 120 160
P
o
s
t
e
r
i
o
r
p
r
o
b
a
b
i
l
i
t
y
,
k
Time slot, k
Threshold, = 0.9
Optimum M
M = 10
M = 3
0
20
40
60
80
100
0 0.2 0.4 0.6 0.8 1
T
o
t
a
l
c
o
s
t
,
J
(
)
Posterior probability,
Optimum M
M = 10
M = 3
M = 2
M = 1
The optimum rule initially keeps just one sensor active
Rapidly increases the number of active sensors when the event occurs
Plots on the right show the average cost comparison with various
xed number of sensors
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 18 / 22
Distributed Detection: Event Detection in a Large Network
Before the Event
A
X
(i )
k
f
0
.
After the Event
Event
A
r
s
X
(i )
k
_
f
1
, if i N
e
(r
s
),
f
0
, if i / N
e
(r
s
).
Problem: An event happens at an unknown time T and at an unknown
location
e
. Detect and locate the event
Objective: Raise an alarm with a small detection delay
subject to Time to False Alarm (TFA) and
Time to False Isolation (TFI) .
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 19 / 22
Distributed Detection: Event Detection in a Large Network
Before the Event
A
X
(i )
k
f
0
.
After the Event
Event
A
r
s
X
(i )
k
_
f
1
, if i N
e
(r
s
),
f
0
, if i / N
e
(r
s
).
Problem: An event happens at an unknown time T and at an unknown
location
e
. Detect and locate the event
Objective: Raise an alarm with a small detection delay
subject to Time to False Alarm (TFA) and
Time to False Isolation (TFI) .
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 19 / 22
Distributed Detection: Event Detection in a Large Network
Boolean Model - Performance of Distributed Algorithms
We deployed 10 nodes in the ROI
in a deterministic manner such
that we get N = 9 regions and
N
1
= {1, 2, 4}, N
2
= {2, 4, 5},
N
3
= {2, 3, 5}, N
4
= {4, 6, 7},
N
5
= {4, 5, 7}, N
6
= {5, 7, 8},
N
7
= {6, 7, 9}, N
8
= {7, 9, 10},
and N
9
= {7, 8, 10}.
0
5
10
15
20
25
30
10
0
10
1
10
2
10
3
10
4
10
5
10
6
S
A
D
D
(
s
a
m
p
l
e
s
)
TFA (samples)
MAX
HALL
ALL
Nikiforov
Nikiforov (1995) provides the optimal centralised algorithm
Distributed algorithms are based on fusion of local CUSUM statistics
We also provide an approach for a power-law signal decay model
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 20 / 22
Outline of Talk
1
Wireless Sensor Networks (WSNs): Quick Overview
2
Optimum Event Detection
3
Other Algorithm Design Work
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 21 / 22
Other WSN Algorithm Design Work
WSN node placement
Minimum cost
Performance objectives (delay and network reliability)
Node localisation without GPS
Hop count based localisation
Theory based on random graph models
Geographical forwarding of alarm packets in WSN
Optimal forwarding when nodes sleep-wake cycle
Algorithms for hole avoidance
Anurag Kumar (ECE Dept., IISc) Wireless Ad Hoc Sensor Networks 14 November, 2010 22 / 22