Title: Sensor Networks, Aeroacoustics, and Signal Processing ICASSP 2004 Tutorial Brian M. Sadler Richard J. Kozick 17 May 2004
1Sensor Networks, Aeroacoustics,and Signal
ProcessingICASSP 2004 TutorialBrian M.
SadlerRichard J. Kozick17 May 2004
2Sensor Network Publication Trend
NSF Boost Phase
Source IEEE Xplore, sensor networks (IEEE only)
3Sensor Networks, Aeroacoustics,and Signal
ProcessingIntl. Conf. on Acoustics,
Sensor-Nets, and Signal Proc.Brian M.
SadlerRichard J. Kozick17 May 2004
4CaveatsSP SP-Comms Perspective, Finite
Citations, RMF Acknowledgements S.
Collier, M. Dong, P. Marshall, S. Misra, T.
Moore, R. Moses, T. Pham, N. Shroff, N. Srour, A.
Swami, R. Tobin, L. Tong, D. K. Wilson, Q. Zhao,
T. Zhou, etc!
rapidly moving field
5Outline
- Part 1 Overview of Sensor Networks
- Consider the rich interplay between sensing,
signal processing, and communications, with a
focus on energy preserving strategies. - Part 2 Aeroacoustic Sensor Networks
- Application of aeroacoustic sensing with
distributed nodes, including propagation effects,
and optimal signal processing, under
communication constraints.
6Sensor Networks, Aeroacoustics,and Signal
ProcessingICASSP 2004 TutorialPart I
Overview of Sensor NetworksBrian M.
SadlerRichard J. Kozick17 May 2004
7Modalities and Applications
Application Domains
- Point sources
- Detection, estimation, geolocation, tracking
moving sources - Imaging sampling a field
- Environment (e.g., temperature, atmosphere)
- Monitoring dedicated sensor / source groupings
(IEEE 802.15.4 / ZigBee) - Assembly lines, machines, hospital patients, home
intrusion - Logistics where is it?, what condition?
- Warehouse, dock, container, on-ship
- Mobility Control
- Robotics, UAVs
- Sensing Modalities
- acoustic, seismic
- vibration, tilt
- thermal, humidity, barometer
- NBC (nuke / bio / chemical)
- magnetic, RF
- light
- high bandwidth (video, IR)
- etc!
- Active sensing
- radar, RF tags
- A range of environments
- home, office, factory
- toxic, inhospitable, remote
- etc!
8Rich Multi-Disciplinary Interplay
- Ad hoc networking
- Sensing / physics / propagation
- Low power / adaptive hardware
- Controls, robotics, avionics
Types of constraints
- Energy
- battery vs continuous power supply
-
- Wireless communications
- 1 or multi-hop to fixed infrastructure vs no
fixed infrastructure - homogeneous vs non-homogeneous nodes (base
stations) - synchronization (beacons, message passing)
geolocation - degree of robustness
- highly variable RF propagation conditions
- and more
- random vs deterministic placement
- sensor density
-
9What is a Sensor Network?
- Postulate (something for everyone)
- Given any definition of a sensor network, there
exists a counter-example. - Extremely varied requirements, environments,
comms ranges and propagation conditions, and
power constraints. - Our focus
- Energy constrained, battery driven, robust radio
communications with little or no fixed
infrastructure - (other possible comms acoustic, laser, UV)
- DSP / MEMS / Nano Moores Law vs Shannon /
Maxwell - Digital Processing Power Requirements Drop by
Factor of 1.6/Year - Eb/No Required Remains Constant
- Maximum lifetime implies minimal communications
-
10Mobility and Overhead
Ad Hoc Mobile Network Aggregate 200 Mbps
Capability
- DoD ad hoc network experiment (mobile high
QoS) - Network overhead dominates
- Fixed overhead increasingly less efficient as
duty cycle decreases
512 byte packet, 32 mcps FEC 1/2 _at_ 4000 kbps
maximum burst
- Headers for each level
- Timing
- Status
- etc
From SUO SAS TIM, June 12 13 2001
- Does Not Include Initial Acquisition, Other Entry
Requests, TCP, Routing Table, and Related
Bandwidth Requirements
Chip-scale sensor
Chip-scale radio
Actual Application 1.8 Mbps Data ? 0.9
The future?
11Energy Themes
- Reduce communications to a minimum
- Idle listening duty cycling
- Reduction of protocol overhead
- Common channel access limits communications
performance - Medium access control (MAC) a critical element
- Coordinated signal processing
- Collaborative distributed signal processing vs
centralized - Optimality and performance under communications
constraints - Specialized low power hardware
- DSP, clocks, radios
-
12Outline
- Intro Energy Themes
- Architectures Connectivity
- Some Fundamental Limits
- Clocks Synchronization
- Hardware Trends
- Node Localization
- Medium Access Control Routing
- Conclusions
-
13Architectures
- flat
- cluster, hierarchical
- mobile collectors
- mobile nodes / robotics / UAVs
- k-hop to fixed infrastructure (k1)
- the likely dominant commercial paradigm
14Connectivity
- Connectivity multi-hop path exists between all
(or desired) nodes - Connectivity is a function of
- Radio channels, power assignment (control), node
locations (density), traffic matrix - Model
- n total nodes, obey Poisson distribution
- geometric path loss
- radius r connectivity
- What density to ensure connectivity?
- Does this scale with area for fixed density?
r
15Connectivity
- 1970s - 80s Magic number 6 (2 to 8
perhaps) - Postulate connecting with approx 6 neighbors
ensures connectivity with very high probability - Under Poisson model with fixed node density, as
area grows then there is a finite probability of
disconnection - Scaling
- Each node should be connected to O(log n) nearest
neighbors, so prob(connected) ? 1. Philips, et
al 1989 Xue Kumar 2004 - Implies a connectivity capacity tradeoff due to
increased multi-user interference - Relation with sensor coverage?
- e.g., Nyquist sampling, detection coverage
16Ad Hoc Network Capacity
- Define new notion of network capacity Gupta
Kumar 2000 - (aggregate transport capacity, bit-meters /
sec) - Comms between random i-j node pairs
(peer-to-peer, multi-hop, random planar network) - For n nodes, and W Hz shared channel, at best
throughput (bits/sec) for each node
scales as - Fundamental limit due to common access
- Splitting channel does not change things
- e.g., FDMA, base-stations
- P-to-P traffic model for sensor nets
- the right one?
- Assumptions
- Fully connected
- Geolocated nodes
- Global routes known
- Perfect slot timing scheduling
- Power control
- Interference noise (no multi-user det.)
- Arbitrary delay
17Correlated Traffic
- Many (most?) sensor network traffic models are
highly correlated - Correlation can be exploited with distributed
compression (coding) when transmitting to a
common destination Slepian Wolf 1973 - fundamental limit on data reduction
- requires known correlation model
- Many-to-One Transport Capacity
- Even with optimal (Slepian-Wolf) compression
assumed, flat architecture with single collector
does not scale Marco, Duarte-Melo, Liu, Neuhoff,
2003 - Leads naturally to routing schemes, e.g., trees,
data aggregation - Scaglione, Servetto, 02, 04
- Development of practical distributed coding
schemes continues - e.g., Pradhan, Kusuma, Ramchandran, 02
18Mobility brings Diversity
- Dramatic gains in capacity limit if mobility is
introduced, i.e., network topology is
time-varying Grossglauser Tse 02 - store and forward paradigm, delay finite but
arbitrary - throughput can now be , i.e., not
decreasing with n - Delay Capacity tradeoff in mobile ad hoc
networks - e.g., mobile network capacity can exceed that of
stationary network, even with bounded delay Lin
Shroff 04 - iid mobility model
- Mobility (time / channel diversity) can greatly
increase throughput in random access schemes
(e.g., ALOHA), when channel knowledge or
multi-packet reception is utilized,
e.g., Tong Naware Venkitasubramaniam 04
19Time Synchronization
- Levels of Timing
- (carrier phase, symbol boundary)
- data fusion, event detection, state update
- MAC scheduling / duty cycling, TDMA slots
- Message frequency vs timing accuracy
- exploit piggy-backing, broadcasting
- extrapolation possible (forward and backward)
- Pairwise vs global synch
- e.g., iterative global LS solution
- several protocols devised in literature
- comms update rates critical
- micro-secs accuracies reported experimentally
circa 1908
20Oscillator Accuracy
Accuracy Power Lifetime with AA battery AA 10,800 J (3 W-Hrs)
GPS 10-8 -- 10-11 180 mW 16.7 hrs beacon, outdoor, cost
DARPA chip-scale atomic clk 10-11 30 mW 100 hrs program goals
MCXO 3 x 10-8 75 mW 40 hrs large, aging drift
TCXO 6 x 10-6 6 mW 500 hrs (21 days) gt1 PPM
Watch clock 200 x 10-6 1 micro W 342 yrs Temp (98.6 ), aging
o
- Increased network timing accuracy increases
lifetime and throughput - With high duty cycling, clock becomes dominant
energy consumer - Low power GPS clocks likely to be developed,
but - Beacons must be robust for DoD application
21Clock Drift and Resync Times
22Hardware Trends
- Sensing, signal processing, radio
- clock, PA, receiver complexity
- State transitions
- duty cycling off, idle, SP, listen, communicate
- turn-on consumes energy, balance against length
of off-time - Performance energy tradeoffs
- dynamic voltage scaling yields variable latency
- slow DSP clock to accommodate time allowed for
the job - multiple DSP bit-widths, i.e., FLOPS at different
quantizations - domain-specific DSP suite
- Energy harvesting
- vibration, solar, thermal
ARL Blue Radio
23An Energy Model
- Coarse energy consumption
- receiver energy may dominate
- idle listening vs duty cycling synch on
receive - scheduling multiple listeners vs perfect
scheduling - short range desirable, but node density high
(application?) - Definition of Network Lifetime? - application
node density dependent - (i) first (or j) node failures
- (ii) first (or k) network partitions appear
Total will incorporate duty cycles
24Power Amplifier Efficiency
- Power control vs PA efficiency
- variable voltage supply to maximize PA use
- PAPR an issue with non-constant modulus
modulations (OFDM)
25Localization Calibration
Where are my nodes? Location, orientation,
calibration.
- Employ internal / external beacons
- Deploy beacons within network GPS limitations
cost - Self-localization use radio or exploit sensor
modality - RF requires sufficient TB product, acoustic /
other possible - Mixed modality possible, e.g, rcvd signal
strength (RSS) AOA mix - Fundamental limits CRB analysis Garber Moses
2003 - desired sensor connectivity approx 5
- always have residual uncertainty
- Relative vs absolute location
- Anchored network (e.g., GPS)
- Sensor calibration
- Temperature, aging
26Medium Access Control (MAC)
How do we efficiently share the common medium?
- Scheduling duty cycling to eliminate idle
listening (TDMA) - Deterministic (peer-to-peer), perhaps
pseudo-random, in clusters - Issues
- scalability
- latency vs energy (duty cycle rate)
- time variation (new joins, drop outs, channel
changes, mobility) - synchronization (clock drift)
- broadcasting (mode switch)
- Random access (e.g., ALOHA)
- Issues collisions energy loss, idle listening
- Slotted employs scheduling (hybrid random access
TDMA) - Optimal duty cycle possible
- low energy to find neighbor dominates
- high energy spent listening dominates
27Medium Access Control (MAC)
PHY / MAC cross-layer design
- Multi-user detection significantly enhances
random access performance (2 or 3 users,
relatively simple SP), e.g., Adireddy, Tong, 02 - Dual-channel transceiver
- e.g., busy-tones in random access (CSMA-MA)
- Further issues
- broadcasting
- monitoring, heartbeat synch, maintain
connectivity - polling from clusterhead vs event driven
- adaptive frame size heavy-tailed (bursty)
traffic
28Medium Access Control (MAC)
- MAC typically comes with large range of tunable
parameters - Analysis challenging, reliant on simulations
small experiments - Optimality measures?
- Scalability?
- Markov model for energy consumption, e.g.,
Zorzi, Rao, 03 - Optimality depends on variable factors
- Applications traffic models
- Node density (perhaps highly varying in same
network) - QoS required? (may be time varying, e.g., how
when to ACK?) - Latency required? (see QoS above)
Solutions provide various tradeoffs. Provable
performance elusive. Adaptability and flexibility
important if variety of service desired.
29Sampling MAC - 1
Consider field reconstruction fidelity under 2
sampling schemes.
Random Access
Deterministic Scheduling
Performance a function of Poisson sensor
distribution sensor density SNR MAC
throughput (finite collection time)
probability no sensor in interval
Processing Steps 1 sensor snapshot 2
information retrieval 3 field reconstruction
Dong, Tong, Sadler, 02, 04
30Sampling MAC - 2
A Mobile Collection Architecture
- Move network functions away from sensors to
mobile APs - Network via mobility
- Connect only when needed
- Design for fraction of packets, from fraction of
sensors (no one sensor is critical)
31(1-D) Signal Field Reconstruction
Sampling MAC - 3
- The signal field (Gaussian, Markov)
- Poisson sensor field with density
- Signal reconstruction via MMSE smoothing
- Performance measure average maximum distortion
of reconstruction (pair-wise sensor spacing
critical)
32Sampling MAC - 4
- MAC Assumptions
- Slotted transmission in a collision channel
- Fixed collection time M slots
- of packets collection is a r.v.
(1) Random Access
(2) Deterministic Scheduling
Sensor Outage Probability (no sensor in interval)
MAC Throughput packets/slot
Schedule one packet per resolution interval of
length
33Sampling MAC - 5
r distortion ratio of random access to
scheduling
- Relative performance depends critically on
- (scheduling less robust)
- Random access may be easier to implement
34Sampling MAC - 6
Deterministic scheduling
random access
- If expect of sensors in interval gt ,
then - scheduled collection is preferred
- Or, given sensor density , choice of
dictates - appropriate collection regime
35Routing
Some rough classes of algorithms
- Energy-aware cost
- parameters delay, range, hop count, battery
level, etc - heterogeneous nodes with highly variable
energy resources - Directed Diffusion
- Query-based, data-dependent routes, controlled
flooding (establish gradients), e.g., tracking - Clustering algorithms
- Supports hierarchical signal processing
- Geographically-based (e.g., geographic
forwarding)
Issues route discovery, scalability Santivanez
et al 02, global vs local, provably good
performance, comms load (energy), mobility
36Odds and Ends
- Security, authentication, encryption
- Broadcasting
- Node management maintenance
- Collaborative transmission
- Relay
- regenerative and non-regenerative
- analog vs digital
- Antennas, propagation
- Iterative distributed detection estimation
- Tracking
37Conclusions
- Its all about energy
- Reduce idle listening, new adaptive hardware,
accurate low power clocks - SP, MAC, and Routing are fundamentally
interrelated - application dependent, cross-layer design
- Large scaling is problematic
- Common channel interference, correlated traffic
flows, leads naturally to clustering - Exploit mobility, heterogeneous nodes
- No Moores Law for batteries (ever?)
- Energy harvesting
- Local vs global SP tradeoffs
- Maximum performance with minimal communications
38Conclusions Cross-Layer Design
- Layered architecture
- takes long term view
- facilitates parallel engineering, ensures
interoperability - lowers cost, leads to wide implementation
- Tension between performance and architecture
Kawadia Kumar 2003 - cross-layer tangled spaghetti ?
- What architecture for low-energy sensor nets?
- limits on performance
- optimal layer interaction feedback
- what information is passed?
- provable stability needed
- widely varying application space
Transport
Network
Link
Physical
OSI Wired World
- Wireless Sensor-Net World
- Multi-antenna
- Multi-user detection
- Synchronization
- Beacons robust comm
- Adapt. modulation coding
- Geolocation
- Hierarchical distr. SP
- Mobility
- Variable QoS
- Routing metric
- Non peer-to-peer
39Sensor Networks, Aeroacoustics,and Signal
ProcessingICASSP 2004 TutorialEnd of Part I
Overview of Sensor NetworksBrian M.
SadlerRichard J. Kozick17 May 2004