Plume Source Position Estimation Using Sensor Networks

1
Plume Source Position Estimation Using Sensor
Networks

• Michalis P. Michaelides Christos G.
  Panayiotou
• Dept. of Electrical and Computer Engineering
• University of Cyprus
• E-NEXT WG1 Meeting
• September 29, 2005

2
Motivation
3
Motivation

• Diversity and quantity of chemicals released into
  the environment has risen dramatically in recent
  years.
• Legacy of land and groundwater contaminated by
  human activities affects the ecosystem, human
  health and quality of life.
• Need for reliable, cost-effective monitoring of
  contaminating compounds in water, soil and
  sediments

4
Motivation cont.

• One of the most dangerous terrorist attacks is
  the release of toxic plumes.
• Internationally this is a hot research topic
  following the September 11, 2001 in New York
  terrorist attacks.
• A contaminant source can also occur as a result
  of an accident at a ship or factory.
• In both cases people and the proper authorities
  need the necessary information within minutes
  after the event to deal with the crisis

5
Presentation Overview

• Introduction
• Related work
• Simulation model
• Simulation results
• Conclusion
• Future Work

6
Plume propagation

• Once released at its source odor is carried by
  the wind to form a plume.
• As the plume travels further away it becomes on
  average more dilute due to molecular diffusion.
• Dominant cause of diffusion is turbulence.
• Characteristics of odor plume depend on physical
  environment.

7
Output of odor plume

• Graphical interpretation of the output of an odor
  plume with moderate turbulence.
• Sensor was stationary at 10 cm downstream of the
  odor source and at the geometric center of the
  plume.

8
Time averaged vs. Instantaneous Plume

• Smooth
• Time-invariant
• Gaussian plume model with peak near the source.
• Odor concentration gradient along wind direction
  is negligibly small.
• Time averaged gradient points towards the source
  only close to the source.

• Discontinuous
• Time-varying
• Detected a lot more often than Gaussian model
  predicts.
• Instantaneous concentrations available at
  significant distances from source.
• Direction of instantaneous gradient does not
  always point to odor source.

9
Related work in plume tracking using unmanned
vehicles

• Bio-mimetic robotic plume-tracing algorithms
  based on olfactory sensing (Homing, Foraging,
  Mate seeking)
• Basic steps in robotic plume-tracing
• Sensing the chemical and sensing or estimating
  fluid velocity.
• Generating sequence of searcher speed and heading
  commands such that the motion of the vehicle is
  likely to locate the odor source.
• J. Farrell et al. uses an autonomous vehicle
  operating in the fluid flow capable of detecting
  above threshold chemical concentration and
  sensing fluid flow velocity.

10
Chemical Plume Tracing by J. Farrell
11
Vehicles vs. Sensor Networks

• Plume finding If the plume is within sensing
  radius of any of the sensors it is immediately
  discovered. (among people, around buildings,
  obstacles)
• Plume maintaining Contact with the plume is
  maintained throughout the sensor field- no
  reacquisition necessary. (time-averaging is
  possible)
• Assuming static sensors the position of the
  source needs to be remotely estimated using
  fusion techniques.
• Energy constraints
• Need efficient routing techniques to relay the
  information hop by hop to the sink.

• Plume finding Has to spend a good amount of time
  searching for the plume in reachable areas.
• Plume maintaining Has the problem of maintaining
  contact with the plume once found and reacquiring
  contact in case it is lost.
• Can move closer to source for better estimation
  until it finds source location.
• All necessary computation can be done on-board.
• Once source location is identified it returns to
  base to report.

12
Related work in sensor networks

• By summer 2005 Syracuse University researchers
  will have installed a dozen robotic sensors to
  form the largest underwater monitoring system in
  USA.
• In Europe SENSPOL Thematic Network focused
  European expertise on the problems associated
  with monitoring environmental pollutants in
  water, soil and sediments.
• Oak Ridge National Laboratory in USA are working
  to develop a SensorNet that will serve as a
  national system for comprehensive incident
  management that will rapidly respond to a
  chemical, biological or radiological event.
• Los Alamos National Laboratory are working in
  developing a DSN (Distributed Sensor Network)
  that will detect a motor vehicle carrying a RDD
  (Radiological Dispersion Device)
• CSIP (Collaborative Signal Information
  Processing) deals with the energy constrained
  dynamic sensor collaboration.

13
Sensor Network Plume Tracking
Contaminant Source
Sensor nodes
14
Simulation model

• N sensor nodes stationary, randomly placed in a
  rectangular field R with locations known ( xi ,
  yi ).
• Contaminant source ( xs , ys ) is somewhere
  inside R (1km x 1km).
• Propagation of contaminant transport is uniform
  in all directions.
• We assume additive Gaussian white noise.
• a2, c106 or simulation results.

Measurement of sensor i at time t
Concentration at source
Gaussian white noise
Radial distance of sensor i from the source
15
Least squares estimation

• Sensor nodes calculate the mean of M measurements
  and then send the computed mean to the sink.
• After the sink receives the information from all
  sensor nodes it employs the nonlinear least
  squares method to compute an estimate of the
  source location by minimizing function J.

16
Least squares start position

• LS max start start the minimization in the
  neighbourhood of the sensor node with the highest
  measurement.
• LS random start randomly pick 10 start
  positions in the sensor field.
• LS combo choose the method that minimizes the
  squared 2-norm of the residual.
• CPA Closest point approach
• The source position is the location of the sensor
  that measured the highest concentration.

17
Simulation results

• MATLAB simulation package
• 100 randomly placed sources for each experiment
  (K100)
• Effect of varying number of sensors, noise
  variance and number of measurement samples.

18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
(No Transcript)
22
(No Transcript)
23
(No Transcript)
24
(No Transcript)
25
Conclusions

• Our proposed sensor network estimates the plume
  source location in a constrained sensor field
  assuming a uniform propagation of the plume.
• The proposed Nonlinear Least Squares optimization
  achieves better estimation results than the CPA(
  Closest Point Approach ).
• Our results indicate that in situations of high
  noise variance it is necessary to increase the
  number of sensors or the number of measurements
  to achieve satisfactory results.

26
Sensor Network Plume Tracking with wind
Wind Direction
Contaminant Source
Sensor nodes
27
New model

• Only a few of the sensors are able to detect the
  plume based on wind direction and spread of the
  plume.
• When a sensor node is triggered by the presence
  of the plume it wakes-up, it takes a number of
  discrete measurements and calculates the mean.
• If the mean exceeds a predefined threshold T it
  communicates this value to the sink and continues
  measuring otherwise it goes back to sleep.
• At the sink as before the nonlinear least squares
  optimization is used to find the source position
  using all available measurements.

28
(No Transcript)
29
New estimator

• LSp Least squares estimator with initial
  concentration known.
• LSc Least squares estimator with initial
  concentration unknown.
• Use separable least squares techniques
• Further improvements
• LSu Unconstrained optimization
• LSw Constrained search based on wind direction

30
(No Transcript)
31
(No Transcript)
32
Threshold considerations

• Determines the number of sensors involved in the
  estimation.
• Needs to be large enough to minimize probability
  of false alarms.
• Needs to be small enough to ensure maximum
  detection probability.
• Needs to be appropriately chosen to minimize
  energy consumption while not compromising
  estimation accuracy.

33
(No Transcript)
34
Future Work

• Propagation model
• Gaussian model, wind, turbulence
• Noise
• Other models, e.g., lognormal or Chi-Square
• Estimation techniques
• Maximum Likelihood, Bayesian estimators
• Data fusion, aggregation
• Multiple sources
• Real-time implementation
• Berkeley modes test-bed