cooperative transmit power estimation under wireless fading murtaza zafer (ibm us), bongjun ko (ibm...
TRANSCRIPT
Cooperative Transmit Power Estimation under Wireless Fading
Murtaza Zafer (IBM US), Bongjun Ko (IBM US), Ivan W. Ho (Imperial College, UK) and Chatschik Bisdikian (IBM US)
Problem Synopsis
Node T is a wireless transmitter with
unknown Tx power P, and unknown
location (x,y)
Nodes {m1,…, mN} are monitors that
measure received power {pi}
Goal – given {pi} and {(xi,yi)} (monitor
locations), estimate unknown P (and
also unknown location (x,y))
m2
P (x,y)
m3
mN
pN
(xN,yN)p3
(x3,y3)p2
(x2,y2)
p1
(x1,y1)m1
T
Problem Synopsis
Sensor Networks – Event detection
– {m1,…, mN} are sensors, and T is the source point of an event
– Goal – detect important events, eg: bomb explosion, based on measured power
Wireless Ad-hoc Networks – physical layer monitoring
– {m1,…, mN} monitor a wireless network
– Goal – detect maximum transmit power violation; i.e. detect misbehaving/mis-configured nodes, signal jamming
m2
P (x,y)
m3
mN
pN
(xN,yN)p3
(x3,y3)p2
(x2,y2)
p1
(x1,y1)m1
T
Applications
“Blind” estimation – no prior knowledge (statistical or otherwise) of the location or
transmission power of T
Talk Overview
• Power propagation model – Lognormal fading
• Deterministic Case – geometrical insights• Single/two monitor scenario
• Multiple monitor scenario
• Stochastic Case• Maximum Likelihood (ML) estimate
• Asymptotic optimality of ML estimate
• Numerical Results
• Conclusion
Power Propagation model
Lognormal fading
Pi = received power at monitor i
di = distance between the transmitter and monitor i
α = attenuation factor, (α > 1) k = normalizing constantHi = lognormal random variable
iii WdkPP )ln()ln(ln
i
ii dkPHP r.v. lognormal;iW
i eH
Wi – unknown to the monitor – represents the aggregated effect of randomness in the environment; eg: multi-path fading
di
Pi
T
mi
P
Deterministic Case
dkPPr Power propagation model:
T 1
Monitor 1
P P1
d1
best estimate of transmit power:
P* ≥ P1
Single monitor measurement
(no fading/random noise in power measurements)
Deterministic Case
Monitor 2
Note: d1, d2 are unknown
Monitor 1
P
P1
P2d12
d1 d2
2
T
1
Simple Cooperation: P* ≥ max(P1, P2)
Q: Can we do better?
Locus of T, constant)(,/1
1
2
2
1 cP
P
d
d
Two monitor scenario
1
1 dPkP Eqn (1)
2
2 dPkP Eqn (2)
Equation of a circle
cyyxx
yyxx
22
22
21
21
)()(
)()(
Deterministic Case
Two monitor scenario
cos1cos)1(
1 222
121*
cc
dP
kP
P achieves lower bound,
/1
2
/1
1
12*
11
1
PP
d
kP
21
T
(x1, y1) (x2, y2)
P1
P2T
T
T
T
T
(x, y)
xθ
cyyxx
yyxx
22
22
21
21
)()(
)()( /1
1
2
P
Pcwhere,
center of circle
0,
)1(2
)1(2
122
c
dc
Deterministic Case
Multiple monitor scenario
;)( 1
/1
1
2
2
1 cP
P
d
d
;)( 2
/1
2
3
3
2 cP
P
d
d
)( 1
/1
1
1
N
N
N
N
N cP
P
d
d
• With multiple monitors – diversity in measurements
• System of equations with unknowns (x,y,P)
• We should be able to solve these equations to obtain exact P ?
Answer: Yes and No !!
Deterministic Case
1
2(xr, yr)
dr,1
dr,2
T(x, y)
3
4
d1
d2
Theorem: There is a unique solution (P*, x*, y*) except when the monitors are placed on an arc of a circle or a straight line that does not pass through the actual transmitter location.
Proof:
• A location (x, y) is a solution if and only if it satisfies d1/d2=c1, …, dN-1/dN = cN-1
• The actual location (xr, yr) is one solution; thus dr,1/dr,2=c1, …, dr,N-1/dr,N = cN-1
• There exists another solution at (x, y) if and only if, dr,1/dr,2 = d1/d2 , …; equivalently,
T
Deterministic Case
1
2(xr, yr)
dr,1
dr,2
T(x, y)
3
4
d1
d2
Observation:
Without transmit power information, and if monitors lie on an arc of a circle, even with infinite monitors and no fading, the transmission power (and transmitter location) cannot be uniquely determined.
T
Theorem: There is a unique solution (P*, x*, y*) except when the monitors are placed on an arc of a circle or a straight line that does not pass through the actual transmitter location.
Deterministic Case
Multiple monitor scenario
1 2
Corollary 1: Two monitors always has multiple solutions
Deterministic Case
Multiple monitor scenario
1 3
Corollary 1: Two monitors always has multiple solutions
Counter-intuitive Insight: For any regular polygon placement of monitors the transmission power cannot be uniquely determined !!
Corollary 2: Three monitors as a triangle always has multiple solutions
2
Conversely: For all non-circular placement of monitors, transmission power can be uniquely determined.
Talk Overview
• Power propagation model – Lognormal fading
• Deterministic Case – geometrical insights• Single/two monitor scenario
• Multiple monitor scenario
• Stochastic Case• ML estimate
• Asymptotic optimality of ML estimate
• Numerical Results
• Conclusion
Stochastic Case
m1
P (x,y)
m2 mN
pN
(xN,yN)p2
(x2,y2)p1
(x1,y1)
Let zi = ln(pi), Let Z = ln(P), and ),,( yxZ
ML estimate (Z*,x*,y*) is the value that maximizes the joint probability density function
);(maxarg*)*,*,(
zfyxZ
The joint probability density function
Maximum Likelihood Estimate
iii WdkPP )ln()ln(ln T
Power attenuation model
Stochastic Case
Theorem: The ML estimate for N monitor case is given as,
• (x*,y*) is the solution to the minimization above, where the objective function is sample
variance of {ln(pidiα)}
22 )()( yyxxd iii
22* *)(*)( yyxxd iii
distance between some location (x,y) and monitor i
distance between estimated Tx. location (x*,y*) and monitor i
• P* is proportional to the geometric mean of {pi(d*i)α}
Stochastic Case
What happens when N increases ?
more number of measurements of received power
increase in the spatial diversity of measurements
Does the transmission power estimate improve ?
Answer: Yes !! ; Estimator is asymptotically optimal
Stochastic Case
Asymptotic optimality as N increases
Random Monitor Placement
N monitors placed i.i.d. randomly in a bounded region Г
Each monitor makes an independent measurement of the received power
Random placement is such that it is not a distribution over an arc of a circle
Let PN* be the estimated transmit power using the results presented earlier
Theorem: As N increases the estimated transmit power converges to the actual power P almost surely,
Numerical Results
Synthetic data set
– N = 2 to 20 monitors placed uniformly at random in a disk of radius R = 40.
– Received power is generated by i.i.d. lognormal fading model for each monitor.
– Performance measured: averaged over estimation for 1000 transmitter locations.
Empirical data set
– Sensor network measurement data by N. Patwari.
– Total 44 wireless devices; each device transmits at -37.47 dBm; received powers are measured between all pairs of devices
– The data is statistically shown to fit well to the lognormal fading model = 2.3, and dB = 3.92.
– Randomly chosen N=3,4,…,10 monitors out of 44 devices.
Numerical Results
Performance metric
– The above metric measures the average mean-square dB error
Estimators
– MLE-Coop-fmin• ML estimate with fminsearch in MATLAB for location estimation
– MLE-Coop-grid• ML estimation with location estimation by dividing region into grid points
– MLE-ideal• ML estimate by assuming that the transmitter location is magically known
– MLE-Pair• ML estimate is obtained by considering only monitor pairs• Average taken over all the pair-wise estimates
])log10*log10[( 21010 PPEdBError K
Numerical Results
Synthetic data set
Empirical data set
(MLE-Coop-grid)
Conclusion
Blind estimation of transmission power
– Studied estimators for deterministic and stochastic signal propagation
– Utilized spatial diversity in measurements
– Obtained asymptotically optimal ML estimate
– Presented numerical results quantifying the performance
Geometrical insights
– Two-monitor estimation was equivalent to locating the transmitter on a certain unique circle
– If monitors are placed on a arc of a circle, the transmission power cannot be determined with full accuracy (even with infinite monitors)