Di t ib t d I t lli t S t Distributed Intelligent Systems ...
Transcript of Di t ib t d I t lli t S t Distributed Intelligent Systems ...
Di t ib t d I t lli t S t W10Distributed Intelligent Systems – W10:An Introduction to Wireless
Sensor Networks from a Di t ib t d I t lli t Distributed Intelligent Systems PerspectiveSystems Perspective
OutlineOutline
• Wireless Sensor Networks (WSN) as ( )a special class of DIS
• Motivating applicationsT h l• Technology
• Tools used in this course– Mica-zMica z– Zigbee-compliant module for e-puck
robots– Webots extensionsWebots extensions
• Closing the loop with multi-robot systems
ll i d i i b h k– Collective decisions as a benchmark– Multi-level modeling for WSN
A Special Class of Di t ib t d I t lli t Distributed Intelligent
Systems: Wireless Systems: Wireless Sensor NetworksSensor Networks
WSN and DISWireless sensor networks:
– are spatially distributed systemsare spatially distributed systems– exploit wireless networking as main inter-node interaction
channel– typically consist of static, resource-constrained nodes– energy saving is a crucial driver for the design of WSN
h d hi h ( i ll h i l– have nodes which can sense, act (typically no physical movement), compute and communicate in an unattended mode
Are WSN a special class of Distributed pIntelligent Systems?
WSN and DISThe potential is there but currently we observe:- Limited embedded intelligence/adaptation:Limited embedded intelligence/adaptation:
- sensing data are typically only collected for a particular application and rarely used to take local actions (e.g., change of activity pattern in a node)
- no emphasis on local real-time perception-to-action loopactivity pattern (sensing computing networking) are typically- activity pattern (sensing, computing, networking) are typically a priori scheduled
- static nodes face lower unpredictability than mobile onesp y
- Limited control distributedness- the fact that WSN are spatially distributed does not necessarily
mean distributed control: the existence of a sink allows for centralized control which in turn often promote energy saving
WSN vs Networked Multi Robot SystemsWSN vs. Networked Multi-Robot Systems
Networking is common sensor nodes = mobile robotsNetworking is common, sensor nodes mobile robots without wheel or mobile robots = sensor nodes with wheels. So minimal difference? Not really …y
• Mobility changes completely the picture of the problem: y g p y p pmore unpredictability, noise, … .
• Self-locomotion even more so: real-time control loop at h d l l b d b kd di llthe node level + energy budget breakdown radically
different• Typically different objective functions and performance• Typically different objective functions and performance
evaluation metrics
Motivating Applications
Motivation
• Micro-sensors, on-board processing and wirelessprocessing, and wireless interfaces all feasible at very small scale– can monitor
phenomena “up close”
• Will enable spatially and ll d
Seismic Structure response
Contaminant Transport
temporally dense environmental monitoring
• Will enable precise, real-time alarm triggering
Marine Microorganisms
Ecosystems, Biocomplexity
time alarm triggering• Embedded networked
sensing will reveal previously unobservable p yphenomena
Adapted from D. Estrin, UCLA
Pioneering deployments –The EPFL-UNIL Bird tracking Project
(Freitag Martinoli Urzelai 1995 1999)(Freitag, Martinoli, Urzelai, 1995-1999)Goals • Understanding better the overall• Understanding better the overall
behavior of migratory Wrynecks (endangered species) and therefore actively intervene for improving his survivability
• Monitoring nest passages, hunting g p g , gmovements, environmental cues (e.g., temperature inside and outside the nest)the nest)
[Freitag, Martinoli, Urzelai, Bird Study, 2001]
EPFL-UNIL Biotracking ProjectEPFL UNIL Biotracking ProjectMonitoring Units• Nest and hunting zone• Bird tags: RFID
transponders (nest) andtransponders (nest) and active emitter (hunting)
• Energy management based on rough estimation ofon rough estimation of bird’s habits
• Male/female identification• Data collection with HP
calculator/laptop • 1 week energetic autonomy1 week energetic autonomy• No wireless networking
LessonsG l fi ld iGeneral field experience• Birds do not usually play the game as we would like to (e.g.,
camouflage issue)• Packaging: major issue (waterproof case, connectors, …)• Not low-stress monitoring (bird captured with nets, …); very
invasive technique but still … tagless techniques?q g q• Still much better than standard human-guided radio-telemetry
Specific field issuesSpecific field issues• Limited observation window: 1 month/year for testing the
equipment in the field with tagged Wrynecks; no failure admitted
Issue due to inexistent unit networking• Manual collection of data from local data loggersgg• No remotely controlled operation possible• No collaborative or centralized sink-based power-aware algorithms
Pioneering deployments –g p yThe WISARD Project
(Flikkema NAU 2001 )(Flikkema, NAU, 2001 -)
Microclimate meas ring in– Microclimate measuring in the Redwood forestImpact of fine scale– Impact of fine-scale ecological disturbances on diversityd ve s ty
– Micro-measurement of energy, water, carbon fluxesgy, ,
Pioneering deployments –Pioneering deployments Great Duck Island
• originally a 9 month deployment (2002)• 32 nodes: light temp humidity barometer32 nodes: light, temp, humidity, barometer• in 2003, added ~200 nodes with various
sensors; still activesensors; still active– www.greatduckisland.net
Application 1 - Permasense
• What is measured:– rock temperaturerock temperature– rock resistivity– crack widthcrack width– earth pressure
water pressure– water pressure
Pictures: courtesy of Permasense
Application 1 - Permasense
• Why:“[…] gathering of[…] gathering of
environmental data that helps to understand the processes that
t li tconnect climate change and rock fall in permafrost areas”in permafrost areas
Pictures: courtesy of Permasense
Application 2 - GITEWSGerman Indonesian Tsunami Early Warning System
• What is measured:– seismic events
y g y
seismic events– water pressure
Pictures: courtesy of Deutsches GeoForschungsZentrum (GFZ)
Application 2 - GITEWS
• Why:To detect seismicTo detect seismic events which could cause a Tsunami. Detect a Tsunami and predict its propagation.
Pictures: courtesy of Deutsches GeoForschungsZentrum (GFZ)
Application 3 - Sensorscope
• What is measured:– temperaturetemperature– humidity– precipitationprecipitation– wind speed/direction
solar radiation– solar radiation– soil moisture
Pictures: courtesy of SwissExperiment
Application 3 - Sensorscope
• Why:Capture meteorologicalCapture meteorological events with high spatial density. y
Pictures: courtesy of SwissExperiment
Application 4: Ecosystem Monitoringpp y gScience• Understand response of wild populations (plants and animals) to habitats
iover time.• Develop in situ observation of species and ecosystem dynamics.
TechniquesTechniques• Data acquisition of physical and chemical properties, at various
spatial and temporal scales, appropriate to the ecosystem, species and habitat.habitat.
• Automatic identification of organisms(current techniques involve close-range human observation).)
• Measurements over long period of time,taken in-situ.
• Harsh environments with extremes in temperature, moisture, obstructions, ...
Source: D. Estrin, UCLA
WSN Requirements for H bit t/E h i l A li tiHabitat/Ecophysiology Applications
• Diverse sensor sizes (1-10 cm) spatial sampling intervalsDiverse sensor sizes (1-10 cm), spatial sampling intervals (1 cm - 100 m), and temporal sampling intervals (1 ms -days), depending on habitats and organisms.
• Naive approach → Too many sensors →Too many data.– In-network, distributed information processing
Wi l i ti d t li t t i thi k• Wireless communication due to climate, terrain, thick vegetation.
• Self-Organization to achieve reliable, long-lived, operation g , g , pin dynamic, resource-limited, harsh environment.
• Mobility for deploying scarce resources (e.g., high l ti )resolution sensors).
Source: D. Estrin, UCLA
E bli T h l i Enabling Technologies and Challengesand Challenges
Enabling TechnologiesEmbed numerous distributed devices to monitor and interact with physical world
Network devices to coordinate and perform higher-level tasks
Embedded NetworkedControl system w/ ExploitControl system w/Small form factorUntethered nodes
collaborativeSensing, action
Sensing& Actuation
Tightly coupled to physical world
Exploit spatially and temporally dense, in situ, sensing and actuation
Source: D. Estrin, UCLA
Sensor Node Energy RoadmapSource: ISI & DARPA PAC/C Program
Sensor Node Energy Roadmap10,00010,000
1,0001,000
r (m
W) • Deployed (5W)
• PAC/C Baseline
Rehosting to Low Rehosting to Low Power COTSPower COTS(10x)(10x)
100100
1010ge P
ow
er • PAC/C Baseline
(.5W)
• (50 mW) --SystemSystem--OnOn--ChipChipAd PAd P1010
11Avera
g --Adv Power Adv Power ManagementManagementAlgorithms (50x)Algorithms (50x)
20002000 20022002 20042004
.1.1
(1mW)
20002000 20022002 20042004
Communication/ComputationSource: ISI & DARPA PAC/C Program
Communication/Computation Technology Projectiongy j
1999 (Bl t th 2004(Bluetooth
Technology)2004
(150nJ/bit) (5nJ/bit)C i ti (150nJ/bit) (5nJ/bit)1.5mW* 50uW
~ 190 MOPSComputation
Communication
Assume: 10kbit/sec. Radio, 10 m range.Assume: 10kbit/sec. Radio, 10 m range.
(5pJ/OP)Computation
Large cost of communications relative to computation Large cost of communications relative to computation continuescontinues
Free Space Path Loss
• Signal power decay in air:
• Proportional to the square of the distance d• Proportional to the square of the frequency f
– high frequency = high loss– low frequency = low bandwidth
Friis LawsFriis Laws
• Basic Friis law (open environment) Pr = received powerP t itt dBasic Friis law (open environment) Pt = transmitted powerGt = gain transmitting antennaGr= gain receiving antennaλ = signal wavelength g gR = distance emitter-receiver
f = c/λ!
• Modified Friis law (cluttered, urban environment)
n between 2 and 5!
Hardware and Software Modules used in this
Course
MICA mote family
• designed in EECS at UCBerkeley• manufactured/marketed by Crossbowmanufactured/marketed by Crossbow
– several thousand producedused by several hundred research groups– used by several hundred research groups
– about CHF 250/piecei t f il bl• variety of available sensors
MICAz
• Atmel ATmega128L– 8 bit microprocessor, ~8MHz
k k– 128kB program memory, 4kB SRAM– 512kB external flash (data logger)
• Chipcon CC2420• Chipcon CC2420– 802.15.4 (Zigbee)
• 2 AA batteries– about 5 days active (15-20 mA)– about 20 years sleeping (15-20 µA)
• TinyOS
Sensor board
• MTS 300 CA– Light (Clairex CL94L)Light (Clairex CL94L)– Temp (Panasonic ERT-J1VR103J)– Acoustic (WM-62A Microphone)Acoustic (WM 62A Microphone)– Sounder (4 kHz Resonator)
TinyOS: description
• Minimal OS designed for Sensor Networks• Event-driven executionEvent driven execution• Programming language: nesC (C-like syntax
but supports TinyOS concurrency model)but supports TinyOS concurrency model)• Widespread usage on motes
– MICA (ATmega128L)– TELOS (TI MSP430)
• Provided simulator: TosSim
The epuck ZigBee-Compliant Radio p g p
• Custom module designed specifically for short range• Software controllable (~5cm-5m)• TinyOS radio stack • Interoperable with MICAz, etc.
33
[Cianci et al, SAB-SRW06]
802.15.4 / Zigbee
• Emerging standard for low-power wireless monitoring and control– 2.4 GHz ISM band (84 channels), 250 kbps data rate
• Chipcon/Ember CC2420: Single-chip transceiver– 1.8V supply
• 19.7 mA receiving• 17 4 mA transmitting17.4 mA transmitting
– Easy to integrate: Open source drivers– O-QPSK modulation; “plays nice”
with 802.11 and Bluetooth
Comparison to other standardsComparison to other standards
Communication Plug-In for WebotsCommunication Plug In for Webots
• OmNET++ engine• OSI framework• Custom Layers
- 802.15.4 Zi B- ZigBee
• Physical communication model:- semi-radial disk with noisesemi radial disk with noise- channel intensity fading
362008-11-06 : Cianci
Collective DecisionsCollective Decisions
Collective Decisions
• A general benchmark for testing distributed intelligent algorithmsg g
• Feasible without mobility relying exclusively on networkingexclusively on networking
Understanding Collective Decisions• Local rules and appopriate amplification and/or
coordination mechanisms can lead to collective decisionsdecisions
• Modeling to understand the underlying mechanisms and generate ideas for artificial systems
Modeling
Ideas forIndividual behaviors and local interactions
Global structuresand collective
decisions
Ideas forartificialsystems
Example 1: Selecting a Path (W1)Example 1: Selecting a Path (W1)
Choice occurs randomly
(Deneubourg et al., 1990)
l 2 S l i d S ( 1)Example 2: Selecting a Food Source (W1)
Example 3: Selecting a ShelterExample 3: Selecting a Shelter• Leurre: European project focusing on mixed insect-robot
• A simple decision-making i 1 2 h l
societies (http://leurre.ulb.ac.be)
scenario: 1 arena, 2 shelters• Shelters of the same and different
darknessdarkness• Groups of pure cockroaches (16),
mixed robot+cockroaches (12+4)• Infiltration using chemical
camouflage and statistical behavioral model
[Halloy et al., Science, Nov. 2007]
behavioral model• More in week 12
Example 4: Selecting a DirectionExample 4: Selecting a Direction
Converging on the direction of rotation (clockwise or anticlockwise):• 11 Alice I robots• local com, infrared based• Idea: G. Theraulaz (and A. Martinoli); implementation: G. Caprari, W. Agassounon
Set up and Collective Decision AlgorithmSet-up and Collective Decision Algorithm
[Cianci et al, SAB-SRW 06]
Some Results
[Cianci et al, SAB-SRW 06]
Alternative Scenario: Networking S N d d R b tSensor Nodes and Robots
[Cianci et al, SAB-SRW 06]
Example 5:Example 5:Assessing Acoustic Events
• Non-trivial event medium– Unpredictable in both
space & time– Generality
• Applicable to other similar media• Applicable to other similar media
– Highly localized• Facilitates experimentation
– No or weak assumptions about the underlying acoustic field
472008-11-06 : Cianci
acoustic field
[Cianci et al., ICRA 2008]
The Physical Set-upe ys ca Set up
R li i f h• Realization of the general case– 1.5 x 1.5m tabletop arena
• Multiple elements– Nodes (Robots)– Radio
S d– Sound– Independent event source
[potentially mobile]
482008-11-06 : Cianci
[p y ]
Submicroscopic Model (Webots)Submicroscopic Model (Webots)• Realistic simulation in
WebotsWebots• Calibrated modules
internal to nodesinternal to nodes– e-puck
(wheels, distance sensors,…)
– Sound propagation (Image-Source)
R di i i– Radio communication (OmNET++; semi-radial disk with noise, channel intensity
492008-11-06 : Cianci
fading, OSI layers)
Measurement Confidence in Acoustic Event Detection
• In some situations, event d i i ldetection may trigger a costly process – (i.e. human intervention, fire ( ,
brigade, etc…)• A simple consensus mechanism
may help limit false positivesmay help limit false positives– Require k nodes to agree on
detection before reportingHere k=2 shown– Here, k=2 shown
• Test against “desirable” & “undesirable” event sources of diff l i i i i
Detected by more nodes = louder
M i i t it ith bidifferent relative intensities– Decoy = {50%, 75%, 95%}
* Target intensity
– Measuring intensity with a binary sensor
Hierarchical Suite of ModelsHierarchical Suite of Models• Microscopic
N ti l ( W k 8)• Non-spatial (see Week 8)• Discrete-spatial
C i i l• Continuous-spatial • Submicroscopic (called module-based in ICRA 2008)
A area of interestN number of nodesN number of nodesD(N) distribution of nodesE events in the environmentPdet(r,Ie) probability of detection
51
det( e) p yPcom(r,It) probability of communication
Performance Metric
⎟⎟⎞
⎜⎜⎛−+⎟⎟
⎞⎜⎜⎛−+⎟
⎟⎞
⎜⎜⎛−+⎟⎟
⎞⎜⎜⎛
= fpE
PSEEM 1/
1)(
1),,,( det δγβαδγβα
false negatives false positives measurements messages
⎟⎠
⎜⎝ ⋅⋅⎟
⎠⎜⎝ ⋅⋅⎟
⎠⎜⎝
⎟⎠
⎜⎝ msstotfptot
E FTNLFTNEEE /),max(),,,( γβγβ
Edet : the number of events reportedEtot : the total number presentedEfp : the number of false positives reportedN : the number of nodes in the networkS : total number of measurements (whole network)( )T : length of experiment (time)Fs : sampling frequencyL : samples per measurement
52
Ls : samples per measurementP : total number of messages (whole network)Fm : maximum message transmission rate
Validation ResultsValidation Results• All four modeling levels
presented agree quite closely with the results from the real system– Avg & Std over 20 runs
shown for each:• 100 events (real & module-
based)based)• 1,000 events (continuous &
discrete spatial)• 10,000 events (discrete , (
non-spatial)– Additional experimentation
using the models should therefore remain applicable( )⎟⎟⎠
⎞⎜⎜⎝
⎛−+=⎟
⎠⎞
⎜⎝⎛ fpdet
E EEE
EEM
max1
21
210,0,
21,
21
53
therefore remain applicable to the target system
( )⎟⎠⎜⎝⎠⎝ totfptot EEE ,max2222
[Cianci et al., ICRA 2008]
Comparison of Execution TimesModel Speed Factor
Non-spatial Microscopic (Matlab) 90.81xDiscrete Spatial Microscopic
(Matlab)23.27x
Continuous Spatial Microscopic (Matlab)
17.07x
S b i i (C) 1 36Submicroscopic (C) 1.36xPhysical System 1.0x
54
Potential 10x for Matlab -> C implementation
Conclusion
Take Home MessagesTake Home Messages• WSNs represent a very promising technology for a
number of applicationsnumber of applications• Commonalities and synergies between distributed,
networked multi-robot systems and WSNs are appearing but their potential need still to be further investigated and formalized
• Collective decisions represent interesting benchmarks• Collective decisions represent interesting benchmarks for testing distributed intelligent algorithms on WSNs
• A first multi-level modeling attempt for WSN has beenA first multi level modeling attempt for WSN has been carried out using a framework similar to that used for swarm robotic systems, an effort potentially allowing f f l i ti ti f liti dfor formal investigation of commonalities and synergies of hybrid robotic/static WSNs
Additional Literature – Week 10• Permasense http://www.permasense.ch• GITWES – the German Indonesian Tsunami Early Warning System
http://www.gitews.de ttp://www.g tews.deftp://ftp.cordis.europa.eu/pub/fp7/ict/docs/sustainable-growth/workshops/workshop-20070531-jwachter_en.pdf
• Sensorscope http://www.sensorscope.ch/• Mobicom 02 tutorial:
http://nesl.ee.ucla.edu/tutorials/mobicom02/• Course list:
http://www-net.cs.umass.edu/cs791 sensornets/additional resources.htmhttp://www net.cs.umass.edu/cs791_sensornets/additional_resources.htm• TinyOS:
http://www.tinyos.net/• Smart Dust Project
htt // b ti b k l d / i t /S tD t/http://robotics.eecs.berkeley.edu/~pister/SmartDust/• UCLA Center for Embedded Networking Center
http://www.cens.ucla.edu/• Intel research Lab at Berkeleyy
http://www.intel-research.net/berkeley/• NCCR-MICS at EPFL and other Swiss institutions
http://www.mics.org