1. On Minimalism and Scale in Sensor Networks
Upamanyu Madhow
ECE Department
University of California, Santa Barbara
Funding Sources

2. Why sensor networks? Sensors have been around for a long time
Why are we so interested in sensor networks?
Role of networking
Can we usefully exploit data from lots of sensors?
A whole greater than the sum of its parts?
Why lots of sensors?
Large areas to be covered, small sensing range
Many big picture issues for theorists to ponder
How does scale improve information?
How do we communicate with a million sensors?
How can inter-sensor collaboration help?
How can we engineer trust in wireless sensor networks?

3. Big picture questions lead to new ideas
How to gather information from 100,000 sensors?
Imaging sensor nets
How do we scale surveillance and tracking?
Tracking with binary proximity sensors
Large-scale camera networks
Can sensors act like a distributed antenna array?
Practical algorithms for distributed transmit beamforming

4. Today’s focus: minimalism and scale
Scaling to a very large number of sensors
Minimalism in sensor capability and deployment requirements
Imaging sensor nets
Scaling in energy
Collaboration with minimal inter-sensor communication
Distributed transmit beamforming with feedback control
Scaling in sensor cost/functionality
Minimalism in sensor output: yes or no
Tracking with binary proximity sensors
Bharath Ananthasubramaniam, Munkyo Seo, Prof. Mark Rodwell
Raghu Mudumbai, Ben Wild, Prof. Joao Hespanha, Prof. Kannan Ramchandran
Nisheeth Srivastava, Raghu Mudumbai, Jaspreet Singh, Rajesh Kumar, Prof. Subhash Suri

5. Imaging Sensor Nets
collector: satellite
base station on UAV
vast numbers of low-complexity "dumb" pixels sensor + RF transducer + antenna.
Sensor field
Sensor field
Field of simple, low-power sensors dispersed across field of view Cast on ground from truck, plane, or satellite
Sensor as pixels (“dumb dust”)
Collector-driven data collection: Sensors electronically reflect, with data modulation, beacon from collector (“virtual radar”)
Sensor-driven data collection: Sensors transmit at will, sensor location and data estimated by collaboration among a network of collectors
Minimal functionality: no GPS, no inter-sensor networking
Lifetime of year on watch cell battery
Sophisticated collectors
Radar and image processing, multiuser data demodulation
Joint localization and data collection
Range varies from 100m to 100 km
Active versus “passive” sensors, collector characteristics

6. Prototype with stationary collector
Millimeter wave carrier frequencies
Narrow beam with moderate size collector antenna
Small sensor form factor
Key challenges
Low-power, low-cost sensor ICs: mm-wave in CMOS
Collector signal processing

7. Inducing a radar geometry
Beacon with location code
Active sensor reflects beacon
Collector
Sensor field with
active sensors and inactive sensors

8. Basic Link Diagram Sensor Collector DOWN-LINK UP-LINK
PRBS
DATA
Jointly detect
DATA and DELAY
Collector
Sensor
BPSK modulation
DOWN-LINK
UP-LINK
Freq. shift to filter out
ground return

9. Zeroth order Link Budget
75 GHz carrier
Collector with 1 meter diameter antenna, 100 mW transmit antenna
100 Kbps using QPSK/BPSK at BER of 10-9
300 m range for semi-passive sensor
Down-link
Up-link
100 km range for active sensor with 5 mW transmit power
Downlink
Uplink
(bottleneck)

10. Collector Imaging Algorithm
: location code
: beam pattern

11. Localization for large sensor density
Single Sensor Algorithm + SIC – works well
25 sensors
100 sensors

12. Brassboard concept validation

13. Ongoing research and open issues
Hardware
Sensor IC design
Collector system integration
Algorithms
Alleviating inter-sensor interference
Collaboration among collectors
Data compression and representation
Collector-driven data collection
Sensor-driven data collection
Integration of transceiver with sensing
Well-defined interfaces for mm wave motes

14. Distributed Transmit Beamforming
Distributed beamforming can increase range or cut power
Rec’d power = (A + A + …+A)2 = N2 A2 if phases line up
Rec’d power = N A2 if phases don’t line up (+ fading)
Can use low frequencies for better propagation
Large “antenna” using natural spatial distribution of nodes
Diversity
BUT: RF-level sync is hard!
Receiver 1 f j e 2
SNR
feedback
Sync can be achieved using RX feedback!

15. Feedback Control Mechanism
Initially the carrier phases are unknown
Each timeslot, the transmitters try a random phase correction
Keep the corrections that increase SNR, discard the others
Carrier phases become more and more aligned
Phase coherence achieved in time linear in number of nodes
Typical phase evolution
(10 nodes)
Concept experimentally verified
by Ben Wild (UC Berkeley)

16. Towards an analytical model
Empirical observation: convergence is highly predictable
Net effect of phase perturbations
y[n+1]
α.y[n] x1 x2
What can we say about the distributions of x1 and x2?
(without knowing all the individual transmitter phases)

17. Key idea: statistical approach
Received signal proportional to
Infinitely many possible i[n] for any given y[n]
Analogy with statistical physics
Given total energy i.e. temperature
What is the energy of each atom?
More interesting: how many atoms have a energy, E
Concept of Macrostates
Distribution of energy is fixed
Maxwell-Boltzmann distribution
Density ~ exp(-E/kT)

18. The “exp-cosine” distribution
Initially i[0] is uniform in (-π, π]
The phases i[n] get more and more clustered
Given , what is the distribution of i[n]?
“Typical” distribution closest in KL distance to uniform
The Conditional Limit Theorem
The “exp-cosine” distribution

19. Exp-cosine matches simulations
N = 500 transmitters
Can now predict trajectory, optimize convergence rate, and prove scalability

20. Accurate prediction of trajectory

21. Optimize phase perturbation
Optimize pdf for δi at each iteration
restrict to uniform pdf: δi~uniform[-δ0,+δ0]
Choose δ0 to maximize E(Δy[n]), given y[n]
200 transmitters
Fixed uniform distribution vs. uniform distribution optimized at each slot

22. Scalability and Convergence
Phase perturbation not optimized
Uniform over (-2o,2o)
Phase perturbation optimized
Scalable: Convergence is linear in the number of nodes N
(provably so for optimized phase perturbations)

23. Ongoing research and open issues
Detailed understanding of time variations
Trade off tracking versus misadjustment
Protocols leveraging distributed beamforming
Generalization to other distributed control tasks?

24. Tracking with binary sensors
Minimalistic model appropriate for microsensors
Sensor says target present or absent
Appropriate for large-scale deployments
How well can we track with a network of binary sensors?
Fundamental limits
Minimal path descriptions
Efficient geometric algorithms

25. The Geometry of Binary Sensing
Target Path
Localization patches
Sensor Outputs
Localization arcs

26. Results Attainable resolution ~ 1/(sensor density*sensing range)
Spatial low pass filtering
Can track only “lowpass” version of the path
OccamTrack algorithm
Minimal piecewise linear representations
Velocity estimates for “lowpass” version
Robustness to sensing range variation
Particle filter + Geometric clean-up
Multiple targets
Many possible explanations for snapshot of sensor readings
Temporal evolution can be exploited by particle filter

27. How a trajectory is localized
Max patch size at least 1/(density*range) for any deployment
Patch size of the order of 1/(density*range) for Poisson deployment
Resolution
theorems

28. Minimal representation: OccamTrack
Greedy algorithm for piecewise linear representation:
Draw lines stabbing localization arcs that are as long as possible

29. Spatial Lowpass Filtering
Cannot capture rapid variations
Can only reconstruct “lowpass” version of path
Justifies simple piecewise linear representation

30. Velocity estimation and minimal representation
Which path should we use to estimate (lowpass version of) velocity?
We can be off by a factor of two!
A simple formula: dv/v = dL/L
Proposition: If piecewise linear approx works well, then velocity estimate is accurate.

31. Simulation Results Weighted Centroid Output OccamTrack Output
(Kim et al, IPSN 2005)
OccamTrack Output
OccamTrack
Velocity Estimate

32. Fundamental Resolution Limits
Theoretical resolution attained by both regular and random deployment

33. Handling non-ideal sensing
Not detected
Experiment with acoustic sensor ?
Detected
Low pass filter
observations
A simple model
Acoustic sensors are unreliable & unpredictable
Localization patch = intersection of outer circles with complements of inner circles
Particle filtering algorithm provides robust performance
Geometric clean-up provides minimal description

34. Experiments with acoustic sensors
Non-ideal sensor
response
Particle Filter
OccamTrack
OccamTrack
with ideal sensing
Particle Filter +
Geometric

35. Multiple Targets: what does a snapshot tell us?
Significant ambiguity regarding how many targets and where they are
Can get a minimal explanation for sensor readings: greedy algorithm
But how do we piece together snapshots?

36. Particle filters work!

37. Simulation results with non-ideal sensing

38. Experiments with passive IR sensors

39. Ongoing research and open issues
Signal processing framework
Spatial lowpass filtering
“Multiuser interference” due to multiple targets
Multimodal sensing
Enhancement of binary model
Bayesian frameworks for combining sensor observations
Distributed realizations

40. A Big Picture What is needed for sensor networks to be a success?
Two or three successful big applications
What can theorists do?
Thought experiments based on application scenarios
Ask the big picture questions
Follow up with logical answers
Collaborate with builders and experimentalists
Example big picture questions
How best to use sensor correlations?
How to provide wire-like guarantees?
What are good architectures for nextgen surveillance and tracking?
What are good models for distributed sensing & actuation?