A framework for QoI-inspired analysis for sensor network deployment planning

1
A framework for QoI-inspired analysis for sensor
network deployment planning

• Sadaf Zahedi
• EE Department, UCLA
• Chatschik Bisdikian
• T. J. Watson Research Center, IBM US

2
Problem Statement
Objective Goal Evaluate, and ultimately optimize
the quality of information (QoI) of the sensor
networks, which support sensor-based
applications QoI Definition QoI is the
collective effect of the (accessible) knowledge,
derived from the sensor-collected data, to
determine the degree of accuracy Event Detection
is common in most of the sensorbased
applications such as surveillance and
intelligence gathering, detecting presence of
enemy weaponry, hostile activities (e.g.,
gunfire, explosion), and etc QoI attributes of
importance for event detection class of
applications are Detection probability (Pd), of
correctly detecting the occurrence of an
event False alarm probability (Pf), of declaring
the occurrence when it did not occur Error
probability (Pe), of making any kind of error in
decision
Observation field
Event S(t)
d1
d5
d2
d3
d4
S1
S5
S4
S3
S2
Sensor deployment field
Sensor-collected data
Is QoI good enough?
designer
3
Reference Detection System
Core fusion detection analysis engine
measurements
detection subsystem
signal
y
noise n(t)
Communications
M
Sensor subsystem
sensor subsystem
sensor subsystem (sampler)
Communications
L
Samples of the Projection of the event signal
s(t)
signal propagation
Anchored on the core analysis engine, a
system-level analysis framework can be developed
that contains the required system parameters, to
provide the knowledge of the signal projections
at the sensor locations.
event signature signal s(t)
4
QoI Analysis Framework/Toolkit Architecture
signal(s) s(t)
others
signal s(t)

• Planner
• provides deployment topology, QoI objectives,
  cost constraints, application domains, etc.
• Designer
• provides sample policies, and system models
  (libraries)

propagation/ attenuation model(s)
propagation/ attenuation model
topology, cost constraints
topology, cost constraints
measurement error model(s)
sampling policy
sampling policy(ies)
sampling policy
noise model(s)
Input Pre-processing
Detection Test Tools
core QoI analysis engine

• Good enough?
• What if scenarios
• .

integration (e.g., averaging)
optimization (e.g., select best deployment plan)
integration (e.g., averaging)
optimization (e.g., select best deployment plan)
Output Post-processing

• Deployment plan

5
Core QoI Analysis Engine

• Binary Hypothesis testing
• Hypothesis H1 ri sini i1,2,,N event
  occurrence
• Hypothesis H0 ri ni i1,2,,N no event
• The Likelihood Ratio Test (LRT)
• fRN Hi(rN) represents the pdf conditionen on Hi
• ?P0/P1 Bayesian threshold
• Decision Test
• where C is the noise covariance matrix CEnTn
• nn1,n2,.,nN

6
Fully Distributed Detection (LM)
vs. Centralized (L1)
Fully distributed detection (LM)
Local Decision 1
Fusion Subsystem 1
Noise
Make decision Based on the detection policy
(Q)
Sensor Subsystem 1
Communication
Local Decision 2
Fusion Subsystem 2
Sensor Subsystem 2
Communication
Local Decision M
Fusion Subsystem M
Centralized detection (L1)
Sensor Subsystem M
Communication
Make decision
Fusion Subsystem
local QoI metrics
system level QoI metrics
iterative calculation of QoI parameters
7
Performance Comparison
?2
?0.5
?1
S4
S3
d3
d4
?P0/P1
d1
d2
event
S1
S2
8
Conclusion

• QoI-based framework for analyzing rather
  non-homogenous systems
• For a finite number of sensors, transient
  signals, arbitrary sensor deployment, and
  different noise level at each sensor
• Framework facilitates
• Decoupling of analysis approach in three steps
  (input preprocessing, QoI core analysis, output
  post processing)
• Mix-and-match of different analysis, and modeling
  approaches
• Compared the centralized vs. distributed
  detection architectures with respect to QoI
• Influence of the priori knowledge on selection
  of the best detection policy for distributed
  schemes
• Future work
• Deployment algorithm which optimize both QoI and
  cost subject to constraints
• Extension of the noise models to models with
  spatio-temporal correlation
• Consider the measurement error models (i.e.,
  errors from the faulty sensors, )