Dynamic Sensor Networks Project Review of Virginia Tech

1
Dynamic Sensor Networks ProjectReview of
Virginia Techs Activities

• Mark Jones
• Virginia Tech

2
Virginia Tech Subcontract

• DSN (Subcontract from USC/ISI)
• PIs Mark Jones
• Focus Network management and tasking
• GUI for specifying user queries and receiving
  results user layer
• GUI for managing a distributed sensor network
  maintenance layer
• User information box provides location, bearing,
  etc
• Distributed algorithms for assigning tasks to
  nodes
• Algorithms for determining optimal node
  placement/deployment
• Distributed algorithms for gathering sensor
  network status information

3
This ReviewSelected Activities

1. Update on the GUI and user information box
2. Update on distributed tasking
3. Outline of node placement/deployment algorithms
4. Plans for distributed algorithms for gathering
   sensor network status information

4
I. GUI and User Information Box

• Graphical User Interface
• User Allows specification of user queries on a
  map-based interface
• User Displays results of user queries on a
  map-based interface
• Technician Allows for monitoring and control of
  network
• Designed for portability to a variety of form
  factors
• Current release at http//www.ccm.ece.vt.edu/dsn
• User Information Box provides real-time data on
  the location, bearing, and inclination of the
  user within the sensor network
• GUI fielded successfully at Sitex00 GUI and User
  Information Box fielded successfully at Sitex01

5
GUI Characteristics

• Extensible architecture
• SenseIT integration via U-Maryland
• Rockwell nodes
• ns simulators
• Interface to User Information Box
• GPS
• Bearing
• Inclination
• Other info
• Interface to 3-D GIS system and other map
  information
• Working toward handhelds
• Quality Java will soon be available on iPAQ

6
Sensor Network Management Layer

• Displays of network status
• Battery level
• Node comm type
• Traffic on a link
• Node sensor activation
• True vs. GPS location (useful in experiments)
• Individual node range
• Link to UCLA-Sensorware coverage server to
  display sensor field coverage breach paths
• Planned (discussed later)
• Moving the computation of these quantities from a
  centralized, node-based scheme to a distributed,
  network area-based scheme

7
II. Distributed Tasking

• Problem Need to efficiently task the sensors in
  a region to avoid wasting power
  processing/sensing capability
• Task only a subset of the sensors in a query
  region
• Allow sensors/preprocessor/processor to be turned
  off when possible to save battery power
• Use processing/sensing resources on only those
  nodes required to allow for more queries to be
  serviced
• Routing gets the query to the region, it does not
  necessarily directly task the sensor nodes
• The local, distributed tasking algorithms should
• avoid complex negotiations or artificial
  assignment of regions
• allow for sensor nodes to leave and join the
  network on-the-fly
• Focus on provably good tunable methods to
  allow for the desired level of redundancy/accuracy
  

8
Tasking Based on Independent Sets

• Current algorithms elect an Application Query
  Server for a region
• The region is based on the range of the sensors
  and/or radios regions are determined on-the-fly
  every node is adjacent to one or more AQSs
• AQSs are selected with energy and other
  constraints in mind
• Periodically re-elected due to failure or energy
  level reduction
• AQS is responsible for accepting tasks and
  assigning them to nodes to best maintain sensor
  lifetime

• Election uses scalable, distributed, asynchronous
  algorithm EO(log(n)/loglog(n))
• Only local communication is required neighbors
  in the region

Example Neighborhood
9
Results from Simulation
Time to first breach

• Simulation
• Algorithms implemented in OpNet with simulations
  results up to 3600 nodes
• Assume that a low-power wakeup radio is available
• Model battery life changes in processor/radio
  state

Time to energy level drop
Scalable Election Time
10
The Next Steps

• The algorithms are designed to be robust to
  multiple types of failures
• Currently simulating failures to ensure
  robustness (correctness and graceful degradation
  of performance)
• Currently simulations defined neighborhood/region
  based on radio range
• Currently extending simulation to define based on
  sensor neighborhood when the sensor range is gtgt
  than the radio range (e.g., seismic detection of
  tanks)
• In the abstract, the algorithms are designed to
  allow for incorporation of mobile sensors into
  the tasking scheme
• Need to incorporate mobile sensor platforms into
  the simulations
• Need to examine the effect of several different
  tasking applications simultaneously running on
  the network
• Load-balancing
• Energy management
• Coordination of differing neighborhoods/regions

11
III. Sensor Node Placement/Deployment

• Problem How to place and deploy sensors for a
  given application in a specified geographic
  region.
• Requirements
• Take into account existing sensor field coverage
• Allow for appropriate level of redundancy (energy
  and faults)
• Take into account the type of signal processing
• Individual nodes
• Aspects of a particular collaborative algorithm
• Take into account differing types of sensors and
  node capabilities
• Allow for differing deployment methods
• Take into account effects of terrain on sensor
  range
• Take into account the target types
• E.g., small vehicles will travel on roads, tanks
  are likely to travel on roads, soldiers might
  travel on roads

12
Our Approach

• Formulate this as an unconstrained optimization
  problem
• Carefully construct the function so that it is
  amenable to fast optimization methods
• Make the function general enough to incorporate
  all of the requirements listed on the previous
  slide
• Use an optimization framework
• Fast local optimization schemes based on Newtons
  method
• Initial guesses based on
• Knowledge of the terrain/topology/application
  which gives an indication of how many nodes may
  be required in a given area
• Try to draw on networking theory which indicates
  how many nodes are required to form an effective
  communication network
• Use global optimization schemes to drive this at
  a high level because this is inherently a problem
  with many local minima (which brings many
  computational challenges along with it)

13
Defining the Function

• The function should be well-behaved
• Continuous and smooth (as realistically possible)
• Continuous gradient and Hessian (where possible)
• Avoid unnecessary local minima
• The function should be easy to compute
• Use simple approximations of more complex
  functions (i.e., approximate exp(x) using a
  polynomial)
• Be able to compute the gradient and Hessian
  exactly (provides faster evaluation and better
  optimization behavior)
• Keep in mind that this is simply an approximation
  of the physical environment of which we have
  limited information
• Our function is a combination of
• Node functions with interactions between nodes to
  represent desired level of overlap
• Precomputed background function which
  represents the desirability of having a node at
  location (x,y,z)

14
Current Status

• We have implemented version 1.0 of the algorithm
  and interfaced this to the GUI
• Straightforward optimization schemes
• Function evaluation not optimized
• It is, as expected, fairly slow, taking 1
  minute to provide a solution for 10s of nodes _at_
  29 Palms
• We are in the process of improving the algorithms
  which will provide dramatic improvements in speed
  and the number of nodes which can be handled
• After that, we will work on automatically
  building the background function and node
  function based on user and sensor network
  information

15
IV. Distributed Network Management

• The GUI currently provides information on the
  status of the network such as
• Energy level, sensor node type, link activity
• Coverage and breach path information
• All of this information is gathered on a per node
  basis, collected at the GUI user node, and then
  displayed
• Using this information, a network operator would
  try to
• Identify and diagnose problems
• Determine where to deploy new nodes
• Problem It is expensive to gather all of this
  information to a central location and the network
  operator most likely just needs to know where a
  problem exists
• Probably not a need for per-node information

16
Plans for Distributed Algorithms

• Assumptions
• we are working with some type of diffusion or
  geographic routing scheme as our underlying
  transport mechanism
• Will present the user with information on a
  region basis rather than a per-node basis
• Task
• Design algorithms which gather information in
  regions and combine for transmission to GUI
• Hierarchical approach
• Representing regions compactly
• What happens with voids

17
Recent Accomplishment Summary

1. Sensorsim
2. GPS-less ad hoc localization
3. Low-latency packet forwarding
4. Dynamic assignment of MAC addresses