Skip to main content

Full text of "DTIC ADA592356: Information Fusion and Cognitive Processing"

See other formats


Information Fusion and Cognitive 

Processing 

Dr. Rabinder N. Madan, 

Mathematical, Computer and Information Sciences 

Office of Naval Research 


September 22, 2010 


703-696-4217 


12/9/2011 


1 







Report Documentation Page 


Form Approved 
OMB No. 0704-0188 


Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and 
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information, 
including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington 
VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it 
does not display a currently valid OMB control number. 


1. REPORT DATE 

SEP 2010 


2. REPORT TYPE 

N/A 


3. DATES COVERED 


5a. CONTRACT NUMBER 


5b. GRANT NUMBER 


5c. PROGRAM ELEMENT NUMBER 


5d. PROJECT NUMBER 


5e. TASK NUMBER 


5f. WORK UNIT NUMBER 


4. TITLE AND SUBTITLE 

Information Fusion and Cognitive Processing 


6. AUTHOR(S) 


7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION 

Mathematical, Computer and Information Sciences Office of Naval report number 

Research Arlington, VA, USA 

9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS (ES) 10. SPONSOR/MONITOR’S ACRONYM(S) 

11. SPONSOR/MONITOR’S REPORT 
NUMBER(S) 

12. DISTRIBUTION/AVAILABILITY STATEMENT 

Approved for public release, distribution unlimited 

13. SUPPLEMENTARY NOTES 

See also ADA560467. Indo-US Science and Technology Round Table Meeting (4th Annual) - Power Energy 
and Cognitive Science Held in Bangalore, India on September 21-23, 2010. U.S. Government or Federal 
Purpose Rights License 

14. ABSTRACT 

In sensor fusion one expects that solutions from individual sensors when combined will lead to a solution 
that outperforms any one of the individual solutions. Though it is known to be true (Condorcet 1786, 
Democracy Models) for similar sensors, the story is still unfolding in reality and for dissimilar sensors. In 
the real world there is a myriad of sensors that are performing similar tasks. They were designed to 
operate individually and their integration is a current after thought and a compulsion, either to arrive at a 
better solution or at least a nonconflicting solution. This talk will review the seminal contributions of 
selected researchers and my involvement in the emergence of the field of data, sensor and information 
fusion and cognitive processes in sensing. 


15. SUBJECT TERMS 


16. SECURITY CLASSIFICATION OF: 


a. REPORT 

unclassified 


b. ABSTRACT 

unclassified 


c. THIS PAGE 

unclassified 


17. LIMITATION OF 

18. NUMBER 

ABSTRACT 

OF PAGES 

SAR 

26 


19a. NAME OF 
RESPONSIBLE PERSON 


Standard Form 298 (Rev. 8-98) 

Prescribed by ANSI Std Z39-18 






My Brush with Science 


ONR 

People you may know 
Science and Information Fusion 
Cognitive Radar and Sensor Fusion 



Scientists in the Program 


• Moeness Amin 

• Y. Bar-Shalom 

• Leon Chua 

• Petar Djuric 

• Dan Fuhrman 

• S. S. Iyengar 

• Thomas Kailath 

• Rudy Kalman 

• R. L. Kashyap 

• Qilian Liang 

• S. K. Mitra 

• Arye Nehorai 

• Athina Petropulu 

• Unnikrishna Pillai 

• Vincent Poor 

• P. P. Vaidyanathan 

• N. Vishwanadham 

• Xiadong Wang 

• Peter Willet 



Information in Sensor Fusion 


Combining information from two or more 
sensors 

Combining information from different modes of 
single sensor 

Fuse information from different algorithms 


Motivations 


Limits of performance for single sensors 
can be pushed only in small increments 

Multisensor integration MAY LEAD TO 
improved detection and identification with 
significantly lower false alarms 

More data - potential for improved 
information from measurements 


Multisensor Integration 

Sensors designed to operate independently: 
integration is an “after thought” 

Fusion approaches have been system and function 
specific 

Conventional techniques, models assume identical 
sensor statistics, equal thresholds, high SNRs, and 
uncorrelated sensor noise fields - example Radar 
PDI 

Issues remain in integrating dissimilar sensors 
collecting data asynchronously and communicating 
to central processor with different time delays 


18 th Century Information Fusion 


■ 1786: Condorcet - Democracy Models: 

Each individual has probability p of making correct 
decision: 

What is the probability of democracy making the correct 
decision ? 

Democracy Model 

p: individual probability of making correct decision; 

n: number of members of democracy 

P_n: probability of democracy making correct decision? 

If p > Vi then P_n > p > Vi P_n approaches 1 as n grows 

p < Vi then P_n < p < Vi P_n approaches 0 as n grows 

p= 1/2 then P_n = p= 1/2 

Informally, democracy will do well 
if p >1/2 and will do bad if p<l/2 



Information Fusion in Twentieth Century 

1956, Reliability: Von Neumann showed how to build 
a reliable system using unreliable components 
under independent failures. 

1962, Pattern Recognition: Chow showed optimal 
threshold fuser for multiple independent classifiers. 

1969, Forecasting: Bates and Granger, “better” 
forecasts can be made by combining different 
forecast methods rather than picking one of them 

Importance of “fusing” rather than picking the “best” 
has been demonstrated in a number of disparate 
disciplines 


Information Fusion in Late Twentieth Century 


Advances in Computing and Complex Engineering systems 
posed new challenges: 

- Distributed Detection: Bayesian methods for object detection using 
measurements from different detectors 

- Sensor Fusion: Multiple sensors became essential to many 
engineering systems - fusion is part of the problem specification 

- Mixture of Experts: Function and regression estimation can benefit by 
combining multiple estimators 

- Multiple Classifiers: There is no single best classifier but “combined” 
one is better than its components 

Information Fusion began taking roots as a discipline unto 
itself: 

Office of Naval Research sponsored first workshop on Information 
Fusion in 1996, jointly with National Science Foundation and 
Department of Energy 


First Workshop on Information Fusion 

Office of Naval Research was the lead sponsor, together with 
National Science Foundation and Department of Energy 


Brought together scientists from: 
Engineering, Computer 
Science, Mathematics, 
Econometrics, 
Bioinformatics, Statistics 


PROCEEDINGS 


WORKSHOP ON FOUNDATIONS OF 

Information / Decision Fusion 

WITH APPLICATIONS TO ENGINEERING PROBLEMS 


This workshop launched the field 
of Information Fusion 


Editors 

Nageswara Rao 
Vladimir Protopope&cu 
Jacob Barhen 
Gu na Seetharaman 



August 7-3, 
Washington, 


Sponsors 


DOE 


ONR 


NSF 













Information Fusion Area Today 

Integral part of newer disciplines including: 

• Distributed Sensor Networks 

• Cyber Data Mining 

• Cognitive sensor Fusion 

Dedicated International Conferences: 

1. International Conference on Information Fusion (13th in Edinburgh, 
2010 ) 

2. International Conference on Multisensor Fusion and Integration 
(Salt Lake City, UT, 2010) 

Journals: 

Information Fusion (2000) 

International Journal of Distributed Sensor Networks (2005) 

Journal of Advances in Information Fusion (2006) 


Information Fusion area - last 

decade or two? 

Rich Information Sources 

- Sophisticated sensors - visual, hyperspectral, radiation, chemical, 
biological, and others 

- Information sensors - web crawlers, information servers, 
sophisticated databases 

Advances in Computation 

- Fusion problems have become complex 

- Powerful computer hardware and algorithms have been developed 

Advances in Networking 

- Made access to computing and data resource easier 

- Wireless networking made ad hoc deployments possible 

- High-performance networks made it possible to support large data 
transport and remote control possible 


NEW in September 2010 


IEEE 2010 INTERNATIONAL CONFERENCE ON MULTISENSOR 
FUSION AND INTEGRATION FOR INTELLIGENT SYSTEMS (IEEE MFI 
2010), SEPTEMBER 5-7, 2010 

The theme of IEEE MFI 2010 was Cognitive Sensor Fusion 


Here the goal of multi-sensor fusion systems is to achieve human-like 
performance in terms of perception, knowledge extraction, and situation 
assessment, exploiting symbolic and/or dynamical systems approaches. 



Cyber-Physical Trade-Offs in 
Distributed Detection Networks 


Naoeswara S. Rao . David K.Y. Yau, J .C. Chin, Chirs Y.T Ma 

Oak Ridge National Laboratory Purdue University 

Rabinder N. Madan 
Office of Naval Research 

2010 IEEE Conference on Multisensor Fusion and I ntegration 
September 5-7, 2010, Salt Lake City, UT 


Research Sponsored by 
Applied Mathematics Program 
Office of Advanced Scientific Computing Research 
U.S. Department of Energy 




IEEE MFI 2010: Cognitive Sensor Fusion 


i l Cl 


BEST PAPER A WARD 


fr Cyber-Physical Trade-Offs in Distributed Detection Networks 


Nagi Rao, Jren-Chit Chin, David Yau, Chris Y. T. Ma 


and Rabinder Madan 


Presented 
September 6 t 2010 




Edward Grant, Program>0nair, MFI 2010 


Thomas C- Henderson, General Chair, MFI 2010 














Motivation 

Detection of Low-level Radiation Sources 


Task: 

• Detect the sources based on 
sensor measurements 


Several underlying math problems 
related to detection networks are open. 
Our work 

-addresses network-based detection 
-provides answers using statistical 
estimation and packing numbers 



Difficulty of Detecting Low-level Radiation Sources 

The radiation levels are only slightly above the background levels 
and may appear to be "normal" background variations 

• Varied Background : Depends on local natural and man-made 
sources and may vary from area to area 

• Probabilistic Measurements : Radiation measurements are 
inherently random due to underlying physical process - 
gamma radiation measurements follow Poisson Process 

Several solutions are based on thresholding sensor measurements 


Well-Studied Problem : Has been studied for decades using single or 
co-located sensors: analytical, experimental and 

- sensor networks offer "newer" solutions but also questions 

Open Mathematical Question : Q1 

I s there a mathematical quantification for a network of sensors to 
achieve better performance than single-sensor detectors? 






Detection of Sources amidst Background Noise 

A Traditional Method for Detection: 

1. SPRT to infer detection from measurements at sensors; 

2. Fuse the Boolean decisions at fusion center. 



Specific Question : Are there methods that perform better? 

Generic Question : Are there classes of detection problems that benefit 
from “fusing" measurements in place of decisions ? 

Our Results : Answer is yes to both questions under 

1. Lipschitz smoothness conditions - limited shielding conditions 

2. Vapnik-Chenvenenkis conditions - discrete intensity drops 

3. * SPRT-sequential probability ratio test 




















Detection Using Localization fl * 

s 


Proposed Method for Detection : 

1. Estimate the source parameters using measurements - A s ;(x S9 y s ) 

2. Utilize likelihood ratio test □ ~ at the fusion center 

^3 



^ 0,1 9 P h 0 9 A 5 ^;^ 



m.j < 



p p 
r o,v r \. 


0’ 



where a _ p ( a £ 

1 S ^5 9 A/ S’> 


y s > x i‘. 



sensor 






















Explanation of Results 


A fixed-threshold SPRT detection method optimizes the detection 
performance within a certain neighborhood of state-space. 

- characterized by sets S , S 

r L t h 


Localization facilitates the adaptation of the threshold to estimated 
neighborhood of the state-space albeit with a certain error probability. 

Our method achieves a trading-off between 

•error probability of the localization method in estimating the 
neighborhood and needed SPRT thresholds; and 
•probability of "uncovered" regions of fixed-threshold SPRT 
By suitable trade-off one can exceed the performance of the latter. 



Summary of Results 

I mproved detection using measurements at fusion center compared to 
existing decision fusion methods, using robust localization, under: 

General non-smooth conditions: 

Separability of probability ratios 

- complex analysis and less intuitive conditions 

+ valid under complex shielding of radiation sources 
Smoothness conditions: 

Lipschitz separable probability ratios; and 
Lipschitz source intensity 

+ intuitive conditions: "bigger" parameter space is better 

- valid typically under open-space environments 

First mathematical proofs for this class of problems to show: 

i) a network of sensors performs better than single or co-located sensors 

ii) measurement "fusion" performs better than detection fusion 
Performance improvement is characterized by the |packing number j 

single scalar 



Summary 


We proposed source detection method using a network of sensors: 

- utilizes localization followed by SPRT 

- out-performs : under both smooth and non-smooth conditions 

any single SPRT method; 
majority and other fusion methods 
•For radiation point source detection: 

- performs better than existing decision fusion methods 
•Shows cyber-physical trade-off: better detection at higher cyber 
cost 

•lowest cost: single sensor with SPRT 
•intermediate cost: SPRT at sensors and Boolean fusion 
•highest cyber cost: send measurements and localization-based 
fusion 



Cognitive Radar 



The University of Arizona is developing: 

- Robust Bayesian channel models for target recognition and surveillance 

- Algorithms for waveform optimization derived from the Bayesian models 

Cognitive Radar: 

- An approach to radar that closes the loop between exploitation (signal processing) 
and control/opfimi7ation ofthe measurements 

- Interrogate the radar environment through smart control of the interrogation 
.p rope rties (I: e' v " beamsteefinig ;" PRF; pul'Se" shape)' "=>" opt] ml zeavai fa bl e 1 1 rrie/en ergy 

- Adaptive measurement control 

* Where to go, what to transmit, where 
to aim, who to cooperate with, ... 




Inte 110gate/ELbirtmiaie the Channel 


Compute Waveforms 
& ManeuveE 




S igtial Prcxes sing; 
Update Hypothesis 
Ens enisle 


< 

:< 






i 




Feed Understanding Back 
to Transmitter 


Priorities and 
Constraints 


Additional 

Knowledge 

Sources 





Fixed search pattern: 
Includes areas where 
target unlikely 



Adaptive search pattern: 
Focus radar on important, 
but uncertain, areas 


Use past observations to enhance future measurements 


THE UNIVERSITY OF ARIZONA, 

















































Enor Rale 00 Transmissions) 


Sample Results: Two Applications 


Target ID 




* Characterize targets by transfer function 
and compute variance over the classes: 




2-1 


M 

2-1 


Search & Track 


Target Parameters 


6 



Describe target parameter space as a grid of 


00 

H 

0.7 
06 
06 ■ 

04 

03 

02 

0.1 


- A measure of entropy vs. frequency 
parameterized by target classes 


wf CIUHef 
- - -■ MI wto Clutter 
SNR wYCkifler 
Itt rfe band 



Gain due to 
adaptive, 
optimized 
waveforms 


target probabilities 

Convertthe probabilistic representation into 
the best beam steering location 

Probability of 
detecting a weak 
target moving 
through search 
zone; 

(500-target test) 


'THU m -. CNR im 
* fMt /kM, i'XR icus 




-40 -36 -30 -26 -20 -15 

Energy per Trarrernisekx’i 



Adaptive Control 
makes more 
efficient use of 
time/eneravi 


THE UNIVERSITY OF ARIZONA* 














































































Ongoing Work 

At UA: 

- Computationally efficient and robust probability update procedures 

• For example, how do we perform stable updates of the Bayesian probability map when 
interference/clutter have unknown pdf? 

- Probability updating procedures for multiple platforms (mapping of target parameters 
to range/Doppler/angle is unique for each platform) 

- Adaptive PRF selection for range-Doppler ambiguity mitigation 

- Practical classification algorithms and waveform constraints (e.g., constant modulus) 

Elsewhere: 

- Dr. Dan Fuhrmann (now at Michigan Tech University) 

• “Active-testing surveillance systems, or playing 20 questions with a radar”; 2003 ASAP 
workshop 

- Proposes radar measurement optimization via probabilistic representation (closed loop system!) 

- Dr. Simon Haykin, McMaster University 

• “Cognitive Radar: A Way of the Future”, IEEE Signal Processing Magazine, Jan. 2006. 

- Summarizes philosophy of cognitive radar and feedback from receiver to transmitter 

• Cognitive Tracking Radar; Cognitive Dynamic Systems 

( http://soma.mcmaster.ca/havkin.php ) 

- Dr. Arye Nehorai, Washington University-St. Louis 

• Adaptive waveform parameters and radar flight path (Asilomar 2008) 

• "Adaptive polarized waveform design for target tracking based on sequential Bayesian 
inference," IEEE Trans, on Signal Processing, Mar. 2008. 



Signal Processing for 
Networked Sensing 

A majority of performance improvements in 
Sensors, Networking and Communication 
connectivity are expected to come from 
conceptualization of new systems, and through 
innovations in Signal Processing. 

Leverage developments in Signal Processing 
techniques to bring about improvements in 
sensing, target resolution, small target 
detectability, multi-target tracking, and 
address hard problems in sensor data fusion, 
track fusion and communications.