UAV Navigation by Expert System for Contaminant Mapping

1
UAV Navigation by Expert System for Contaminant
Mapping

• George S. Young
• Yuki Kuroki,
• Sue Ellen Haupt

2
Goals

• Background
• Source and wx information needed for contaminant
  modeling
• Long et al.(2008) demonstrated the use of
  Gaussian puff to back-
• calculate the source characteristics via a
  Genetic Algorithm

• Constraints
• Number of sensors time to solution

• Mission
• Identify a total of 4 parameters (source
  strength, source location (x,y)
• and wind direction) describing the release
  using mobile sensors

3
System Components
4
Dispersion model
5
Hybrid Genetic Algorithm (GA)
Optimization with a GA
Nelder Meade Downhill Simplex
Fine-tune GA solution
6
GA Tuning

• What we did?
• Determine best combination
• of GA parameters

• 2. Concerns?
• Minimizing CPU time
• Increasing accuracy

Pseudo-Runtime popit
7
Experimental Setup

• Wind direction 270 degrees
• Random source location in upwind half of domain
• Single fixed sensor in downwind half of domain
• UAV takes off from upwind corner of domain
• Worst case position
• Launches on first detection by fixed sensor
• UAV speed is 4 times wind speed

8
Autonomous Aircraft

• Why use aircraft?
• Equipping the UAV with GPS concentration sensor
  
• Avoid the cost of a dense array of fixed sensors

• Why autonomous?
• AI required for rapid decision making
• Ensemble of manned aircraft would be too
  expensive

• Why virutal
• Test in a fully controlled environment
• Test UAV naviagtion algorthims without
  societal risk

9
Information Flow

• UAV AI needs observed modeled concentration
• fields to navigate

• GA needs UAV wind concentration observations
• to locate source

• Forward model needs wind and source locaton to
• predict concentration field

10
Expert System Design
Amount of data needed

• Plume

Puff
Difference

• How many passes
• through plume?
• How much separation
• in space?

• Why the difference?

• How many passes
• through puff?
• How much separation
• in time?

11
Plume Expert System

• Plume decision logic

700
-700
700
12
Puff Expert System

• Puff decision logic

(-7000,7000)
Mean wind direction
13
Flight Track Plume Example
14
Flight Track Puff Example
15
Testing Architecture

• Monte Carlo testing of UAV non-collaborative
  ensemble
• Pseudo-random initial population and sensor
  location

Identical twin experiment Create data
Noise Contaminate data

• Ensemble median to back calculate source and
  wind dir.
• Monte Carlo mean of ensemble median will be
  shown

16
Plume Results
Wind
Concentration
X
Y
17
Puff Results
Wind
Concentration
X
Y
18
Conclusions
Experimental Setup
Gaussian Puff UAV
Discussion
Gaussian plume UAV

• 2 flight legs
• 1 UAV
• UAV navigation
• by expert system
• GA optimization
• for source dir
• 1400m domain
• Results improve

• 6 flight legs
• 20 UAVs
• Median Solution
• 14km domain
• Greater tracking
• challenge
• Most UAVs
• succeed

• Idential twin
• 1 fixed sensor
• Single UAV
• or
• UAV ensemble
• No cooperation

19
Future Work
Goal Compensate for the tight time constraints
inherent in emergency management

• Cooperation between Multiple UAVs
• Improve Gaussian Puff Model Navigation
• Actual UAVs
• Field Test

20
Acknowledgements

• The second author was supported by Japan Ground
  Self Defense Forces during this study
• Thanks to J. Wyngaard, K. Long, A. Annunzio, A.
  Beyer-Lout, L. Rodriguez for insights and advice

21
Questions?