distributed localization for wireless distributed networks in indoor environments
DESCRIPTION
Masters thesis defense slidesTRANSCRIPT
Distributed Localization for Wireless DistributedNetworks in Indoor Environments
Hermie P. Mendoza
Wireless @ VTVirginia Polytechnic and State University
June 28, 2011
Masters Thesis Defense Presentation
Agenda
1 Preliminaries of PL and WDC
2 Fingerprint-based PL
3 WDC-based Fingerprinting System
4 Algorithm Performance and Results
5 PL Demo
6 Conclusion and Future Work
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 2 / 46
Preliminaries
Preliminaries Overview
Location-Awareness in Ubiquitous Computing
Position Location Fundamentals
Wireless Distributed Computing (WDC) Fundamentals
Why Position Location and WDC?
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 3 / 46
Preliminaries Position Location
Location Awareness in Ubiquitous Computing
Figure: User accessing location-based service on a smartphone.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 4 / 46
Preliminaries Position Location
The Principles of Positioning I
Positioning Problem: Reasonably localize an object within aglobal or local frame of reference.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 5 / 46
Preliminaries Position Location
The Principles of Positioning II
Figure: Summary of Position Location
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 6 / 46
Preliminaries Position Location
The Principles of Positioning III
Figure: Summary of Position Location
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 7 / 46
Preliminaries Wireless Distributed Computing
What is WDC?
New paradigm emphasing distributed information services!
Figure: Information service shift from centralized to de-centralized computation.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 8 / 46
Preliminaries Benefits
Benefits of WDC
Potential Benefits Results
1 Lower energy and powerconsumption per node
2 Efficient load balancingacross collaborating nodes
3 Harnesses availablenetwork resources
4 Robust, secure, & faulttolerant execution
5 Simplifies radio’s formfactor
Extends total networklifetime
Better resource demandand supply matching
Meets computationallatency requirements ofcomplex processing tasks
Attain stringent QoSrequirements
Economic cost savings
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 9 / 46
Preliminaries Benefits
Benefits of WDC
Potential Benefits Results
1 Lower energy and powerconsumption per node
2 Efficient load balancingacross collaborating nodes
3 Harnesses availablenetwork resources
4 Robust, secure, & faulttolerant execution
5 Simplifies radio’s formfactor
Extends total networklifetime
Better resource demandand supply matching
Meets computationallatency requirements ofcomplex processing tasks
Attain stringent QoSrequirements
Economic cost savings
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 9 / 46
Preliminaries Benefits
Benefits of WDC
Potential Benefits Results
1 Lower energy and powerconsumption per node
2 Efficient load balancingacross collaborating nodes
3 Harnesses availablenetwork resources
4 Robust, secure, & faulttolerant execution
5 Simplifies radio’s formfactor
Extends total networklifetime
Better resource demandand supply matching
Meets computationallatency requirements ofcomplex processing tasks
Attain stringent QoSrequirements
Economic cost savings
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 9 / 46
Preliminaries Benefits
Benefits of WDC
Potential Benefits Results
1 Lower energy and powerconsumption per node
2 Efficient load balancingacross collaborating nodes
3 Harnesses availablenetwork resources
4 Robust, secure, & faulttolerant execution
5 Simplifies radio’s formfactor
Extends total networklifetime
Better resource demandand supply matching
Meets computationallatency requirements ofcomplex processing tasks
Attain stringent QoSrequirements
Economic cost savings
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 9 / 46
Preliminaries Benefits
Benefits of WDC
Potential Benefits Results
1 Lower energy and powerconsumption per node
2 Efficient load balancingacross collaborating nodes
3 Harnesses availablenetwork resources
4 Robust, secure, & faulttolerant execution
5 Simplifies radio’s formfactor
Extends total networklifetime
Better resource demandand supply matching
Meets computationallatency requirements ofcomplex processing tasks
Attain stringent QoSrequirements
Economic cost savings
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 9 / 46
Preliminaries Location Awareness for WDC Paradigms
Location Awareness for WDC Paradigm
Improve overall wirelesscommunication system
Needed to achieveinteroperability
Figure: Cognitive radio sensingenvironment
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 10 / 46
Preliminaries PL and WDC
Motivations I
Localization is generally accomplished in a centralized manner at theexpense of a single network node’s resources. Can the problem ofpositioning be solved in a distributed manner or parallelized?
Figure: Resource constrained mobile phone.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 11 / 46
Preliminaries PL and WDC
Motivations II
(a) Point inside the mall (b) Point inside an airport
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 12 / 46
Preliminaries Min Makespan
Min Makespan Problem I
Goal
Minimize the time taken to compute the individual localizationcalculations.
Problem Formulation
Given a set of J of m jobs and a set of N of n nodes, theprocessing time for a job j ∈ J on node i ∈ N is pij ∈ Z
+. Thenwe must find an assignment of the jobs J to the nodes N suchthat the makespan, or the completion time, is minimized.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 13 / 46
Preliminaries Min Makespan
Min Makespan Problem II
Integer programming formulation
minimize t
subject to∑i∈N
xij = 1, j ∈ J
∑j∈J
xijpij ≤ t, i ∈ N
xij ∈ {0, 1} , i ∈ N, j ∈ J
(1)
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 14 / 46
Fingerprinting High Level Overview
Fingerprint Overview
Problem Formulation
The Fingerprint
Fingerprinting Algorithms
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 15 / 46
Fingerprinting Problem Formulation
Fingerprint Problem Statement
Problem Statement
Using only RSS observations of an arbitrary transmitter, locate andestimate its position in a distributed manner.
Goal
Distributed algorithms must be flexible and applicable for variousfingerprint-based positioning systems.
Computational nodes must form a WDCN.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 16 / 46
Fingerprinting The Fingerprint
The Fingerprint I
Figure: Fingerprinting Concept
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 17 / 46
Fingerprinting The Fingerprint
The Fingerprint II
Mathematical Interpretation
(xi , yi ) = [FP1,FP2, . . . ,FPn] (2)
for fingerprint location i , using n sensor nodes.
Alternative Interpretation
f = (xi , yi ) = [FP1,FP2, . . . ,FPn] (3)
for fingerprint location i , using n sensor nodes.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 18 / 46
Fingerprinting Fingerprinting Algorithms
Fingerprinting Algorithms
Euclideandistance
Bayesianmodeling
NeuralNetworks
Deterministic positioning method
L2 =
√√√√ n∑i=1
∣∣FPi − FP′i
∣∣2 (4)
(x , y) = minFPi
√√√√ n∑i=1
∣∣FPi − FP′i
∣∣2 (5)
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 19 / 46
Fingerprinting Fingerprinting Algorithms
Fingerprinting Algorithms
Euclideandistance
Bayesianmodeling
NeuralNetworks
Probabilistic positioning method
P( l | f ) = P ( f | l)P(l)P(f )
, P(f ) �= 0 (4)
P( f | l) =n∏
j=1
P( fj | l) (5)
P ( lt | lt−1) =
∫P(lt | l ′t−1
)P(l
′t−1) dl
′t−1 (6)
(x , y) = maxl
P ( f | l)P(l) (7)
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 19 / 46
Fingerprinting Fingerprinting Algorithms
Fingerprinting Algorithms
Euclideandistance
Bayesianmodeling
NeuralNetworks
Pattern Recognition
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 19 / 46
Fingerprinting Fingerprinting Algorithms
Distributed Target Localization I
Distributed Localization Approaches
Transfering computationally complex operations to a single node withgreater capabilities.
Parallelizing the position location calculations.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 20 / 46
Fingerprinting Fingerprinting Algorithms
Distributed Target Localization II
Figure: Partitioning a service area for a WDCN.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 21 / 46
Fingerprinting Fingerprinting Algorithms
Notations
f number of fingerprint locations
p number of partitions
(x , y) estimated position of user
gi vector of probabilities calculated by node i
FPi tuple of RSS at fingerprint location i
FP i vector of distances calculated by node i
FPi RSS received at sensor i
FP′i RSS database entry of sensor i
pi AOR or partition assigned to a node i
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 22 / 46
Fingerprinting Fingerprinting Algorithms
Distributed Euclidean Distance Algorithm (DEDA) I
Centralized Approach
(x , y) = minFPi
√√√√ f∑i=1
∣∣FPi − FP′i
∣∣2 (4)
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 23 / 46
Fingerprinting Fingerprinting Algorithms
Distributed Euclidean Distance Algorithm (DEDA) II
Distributed Approach
Initialize FPi = 0.while pi is assigned and received, do
for all FP′j ∈ fj , do
FP i ←√∑j
k=1
∣∣FPk − FP′k
∣∣2end for
end while(xj , yj )← minFPi
∈ fj .return (xj , yj)
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 24 / 46
Fingerprinting Fingerprinting Algorithms
Distributed Bayesian Model Algorithm (DBMA) I
Centralized ApproachSEE
P( l |FPi ) =P (FPi | l)P(l)
P(FPi), P(FPi ) �= 0 (5)
ACT
P ( lt | lt−1) =
∫P(lt | l ′t−1
)P(l
′t−1) dl
′t−1 (6)
where lt is the current location and lt−1 is the previous location.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 25 / 46
Fingerprinting Fingerprinting Algorithms
Distributed Bayesian Model Algorithm (DBMA) II
Distributed Approach
Initialize gi = 0.while pi is assigned and received, do
for all FP′j ∈ fj , do
gi(j)← P( j | {FP1,FP2, . . . ,FPn})end for
end whilereturn gi
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 26 / 46
Fingerprinting Fingerprinting Algorithms
Distributed Neural Networks (DNN) I
Types of Neural Networks
Multilayer Perceptron
Generalized Regression
Both will require a supervised learning to train the network.
Figure: Artificial Neural Network (ANN) Architecture
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 27 / 46
Fingerprinting Fingerprinting Algorithms
Distributed Neural Networks (DNN) II
Figure: WDCN with neural networks
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 28 / 46
WDC-based Fingerprinting System Overview
System Overview
Experimental Setup
Hardware and Software
The Radio Map
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 29 / 46
WDC-based Fingerprinting System Experimental Setup
Experimental Setup
Figure: System block diagram
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 30 / 46
WDC-based Fingerprinting System Hardware and Software
Hardware
Figure: USRP2 with custom WBX daughterboard
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 31 / 46
WDC-based Fingerprinting System Hardware and Software
Software
WDCN communications - GNU Radio
Fingerprint position processing - Python
Web-based user interface - PHP
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 32 / 46
ICTAS
ORIGIN
WDC-based Fingerprinting System The Radio Map
The Radio Map
0
Radio Map for 1st Floor ICTAS
10
-5
-15
-10
N22
N21
N20
-20
RSS
(dB)
N19
N18
N17
N16
-30
-25 N15
N14
N13
N12
-35
N11
-40
0 5 10 15 20 25 30 35 40 45
Position Number
Figure: Radio Map with 45 Positions
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 34 / 46
Algorithm Performance and Results
Algorithm Performance and Results
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 35 / 46
Algorithm Performance and Results Algorithm Evaluation
Algorithm Evaluation I
Comparison of Distributed Localization Algorithms
0 20 40 60 80 100 120 140 160 180 200 2203
4
5
X−direction (ft.)
Y−
dire
ctio
n (f
t.)
Distributed Neural Network − MLP
Actual PathEstimated Path
0 20 40 60 80 100 120 140 160 180 200 2203
4
5
X−direction (ft.)
Y−
dire
ctio
n (f
t.)
Distributed Neural Network − GR
Actual PathEstimated Path
0 20 40 60 80 100 120 140 160 180 200 2203
4
5
X−direction (ft.)
Y−
dire
ctio
n (f
t.)
Distributed Markov
Actual PathEstimated Path
0 20 40 60 80 100 120 140 160 180 200 2203
4
5
X−direction (ft.)
Y−
dire
ctio
n (f
t.)Distributed Euclidean
Actual PathEstimated Path
Figure: Comparison of solutions of distributed localization algorithms
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 36 / 46
Algorithm Performance and Results Algorithm Evaluation
Algorithm Evaluation II
50 100 150 20010
20
30
40
50
60
70
80
90
100
Error Radius (ft)
Per
cent
age
(%)
DEDADBMAGRNNMLPNN
Figure: Performance comparision of distributed localization algorithms
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 37 / 46
Algorithm Performance and Results Error Statistics
Error Statistics
Error statistics of distributed localization algorithmsAlgorithm Minimum Error Mean Error Max Error
DEDA 0 ft. 10.81 ft. 55 ft.
DBMA 0 ft. 35.33 ft. 220 ft.
GRNN 0 ft. 14.95 ft. 95 ft.
MLP 0 ft. 16.90 ft. 155 ft.
Average 0 ft. 19.50 ft. 131.25 ft.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 38 / 46
Position Location Demo Overview
Overview
Functional Workflow of WDC process
Video of Demo
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 39 / 46
Position Location Demo Functional Workflow
Task dissemination and retrieval I
Figure: Phase I: Task dissemination
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 40 / 46
Position Location Demo Functional Workflow
Task dissemination and retrieval II
Figure: Phase II: Task retrieval
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 41 / 46
Position Location Demo Demo
Fingerprinting Position System
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 42 / 46
Position Location Demo Computational Complexity
Computational Complexity of Online Phase
Single node
Algorithm Computation Searching Sorting
EDA O(n) N/A O(n log n)BMA O(n) O (n (log u + 1)) O(n log n)
WDC slave node
Algorithm Computation Searching Sorting
DEDA O(n/4) N/A O(n/4 log n/4)DBMA O(n/4) O (n/4 (log u + 1)) O(n/4 log n/4)
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 43 / 46
Concluding Remarks Conclusions
Conclusions
Successful location estimates are highly dependent on quality anduniqueness of RF fingerprints.
Increasing spatial granularity of fingerprint positions does notnecessarily improve performance of position estimation.
Distributed PL is beneficial for large service areas with largedatabases.
De-centralized computations removes single-point of failure andsecurity intrusions.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 44 / 46
Concluding Remarks Future Work
Future Work
Examine optimization techinque of multisplitting for conventional PLtechniques.
Expand distributed sensor system to all CORNET nodes and createmobile WDCN.
Implement demo with new UHD driver for USRP2.
Implement neural network for WDCN.
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 45 / 46
Concluding Remarks Future Work
Questions
Hermie P. Mendoza (VA Tech) Distributed Localization for WDN June 28, 2011 46 / 46