Post on 17-Jan-2016
description
Model-Based Monitoring for Model-Based Monitoring for Early Warning Flood DetectionEarly Warning Flood Detection
Elizabeth A. Basha, Computer Science and Artificial Elizabeth A. Basha, Computer Science and Artificial Intelligence LaboratoryIntelligence Laboratory, Massachusetts Institute of
Technology
Daniela Rus, Computer Science and Artificial Intelligence Daniela Rus, Computer Science and Artificial Intelligence LaboratoryLaboratory, Massachusetts Institute of Technology
Sai Ravela, Earth Atmospheric and Planetary Science Sai Ravela, Earth Atmospheric and Planetary Science Massachusetts Institute of TechnologyMassachusetts Institute of Technology
OutlineOutline
• MotivationMotivation
• Previous WorkPrevious Work
• Prediction ModelPrediction Model
• Sensor Network ArchitectureSensor Network Architecture
• Installation and ResultsInstallation and Results
• ConclusionConclusion
• Pros&ConsPros&Cons
MotivationMotivation• River flooding detectionRiver flooding detection• Deployment target: rural and developing Deployment target: rural and developing
countriescountries• Requirements:Requirements:
– Withstanding hardware to river flooding and Withstanding hardware to river flooding and stormsstorms
– Monitor and communicate over 10000km^2 Monitor and communicate over 10000km^2 basinbasin
– Predict flooding autonomouslyPredict flooding autonomously– Limit costs allowing feasible implementation Limit costs allowing feasible implementation
in development countryin development country
IntroductionIntroduction• Flood Prediction Algorithm is based Flood Prediction Algorithm is based
on a regression model.on a regression model.• Nearly as good as that used by Nearly as good as that used by
hydrology researchershydrology researchers
Previous work (1/2)Previous work (1/2)
• Sensor network for environmental Sensor network for environmental monitoringmonitoring
• Redwood tree (air temperature, humidity, Redwood tree (air temperature, humidity, solar radiation). solar radiation). – Off-line data analysisOff-line data analysis
• Light intensityLight intensity– Communication via ZigbeeCommunication via Zigbee
• James reserve (humidity, rain, wind)James reserve (humidity, rain, wind)– Deployment in Bangladesh rice paddy to Deployment in Bangladesh rice paddy to
measure nitrate, calcium and phosphatemeasure nitrate, calcium and phosphate• VolcanoVolcano
– Seismic and acoustic dataSeismic and acoustic data
Previous work (2/2)Previous work (2/2)
• None above envision system None above envision system requirements:requirements:– Minimalistic number of sensor availableMinimalistic number of sensor available– Real-time need of dataReal-time need of data– Computational autonomyComputational autonomy– Complexity necessary to perform Complexity necessary to perform
predictionprediction
Sensor networks for flood Sensor networks for flood detectiondetection
• Castillo-EffenCastillo-Effen– Suggested an architecture but unclear on Suggested an architecture but unclear on
basin characteristics and no hardware basin characteristics and no hardware detaildetail
• HughesHughes– Gumstix sensor nodes, linux OSGumstix sensor nodes, linux OS– Tested in the lab but no field testTested in the lab but no field test– Planned deployment of 13 nodes along Planned deployment of 13 nodes along
1km riverside without flood prediction 1km riverside without flood prediction model.model.
Operational systems for flood Operational systems for flood detectiondetection
• US Emergency Alert SystemUS Emergency Alert System• Volunteer and limited technologyVolunteer and limited technology• MIKE 11-based flood forecasting MIKE 11-based flood forecasting
systemsystem
Computational requirementsComputational requirements• SAC-SMASAC-SMA
– Modeling different methods of rainfall Modeling different methods of rainfall surface runoff to determine how much surface runoff to determine how much water will enter the riverwater will enter the river
– Complex equations to establish the Complex equations to establish the modelmodel
– Not easily running on sensor networkNot easily running on sensor network
Prediction ModelPrediction Model
• Rainfall-runoff model:Rainfall-runoff model:– Computational burden, difficult to Computational burden, difficult to
customized for individual basincustomized for individual basin• Statistic model:Statistic model:
– Based on observed recordsBased on observed records– Intrinsically self-calibrated, real-timeIntrinsically self-calibrated, real-time– Used in other areas such as hurricane Used in other areas such as hurricane
intensity forecastingintensity forecasting– Linear regression models assume a linear Linear regression models assume a linear
equation can describe system behaviorequation can describe system behavior– Weighting the past N records of relevant Weighting the past N records of relevant
inputs at time T to produce prediction at inputs at time T to produce prediction at T+tT+t
– Past prediction errors are allowedPast prediction errors are allowed
Flood prediction algorithmFlood prediction algorithm
Test data and setupTest data and setup
• Use 7 years of rainfall, temperature Use 7 years of rainfall, temperature and river flow data for Blue River in and river flow data for Blue River in OklahomaOklahoma
• Compare data to DMPICompare data to DMPI• 3 criteria for the quality of algorithm:3 criteria for the quality of algorithm:
– Modified correlation coefficientModified correlation coefficient– False positive and negativeFalse positive and negative
Model CalibrationModel Calibration
• Training window: 1/3/6/9/12 monthsTraining window: 1/3/6/9/12 months• Optimal values of inputs: Sweep the Optimal values of inputs: Sweep the
order for each input of past order for each input of past predictionprediction
• Pick the best input values with high Pick the best input values with high MCC and low false positive/negativeMCC and low false positive/negative
• Other approaches: climatology, Other approaches: climatology, persistencepersistence
• 1/24 hours prediction1/24 hours prediction
Sensor network architecture Sensor network architecture (1/2)(1/2)• Monitor events over large geographic regions of Monitor events over large geographic regions of
10000 km^210000 km^2• Provide real-time communication of Provide real-time communication of
measurements covering a wide variety of measurements covering a wide variety of variables contributing to the event occurrencevariables contributing to the event occurrence
• Survive long-term element exposure, the Survive long-term element exposure, the potential devastating event of interest, and potential devastating event of interest, and minimal maintenanceminimal maintenance
• Recover from node lossesRecover from node losses• Minimize costsMinimize costs• Predict the event of interest using a distributed Predict the event of interest using a distributed
model driven by data collectedmodel driven by data collected• Distribute among nodes the significant Distribute among nodes the significant
computation needed for the predictioncomputation needed for the prediction
Sensor network architecture Sensor network architecture (2/2)(2/2)• 2-tier communication network2-tier communication network
– Long-range communication node Long-range communication node transmits on the order of 25 km using transmits on the order of 25 km using 144 MHz radio144 MHz radio
– Low power sensing node operates at Low power sensing node operates at 900 MHz900 MHz
– Office and communication nodes for UIOffice and communication nodes for UI
Base systemBase system
• Base system:Base system:– ARM7TDMI-S microcontroller core for ARM7TDMI-S microcontroller core for
LPC2148 from NXPLPC2148 from NXP– Using photovoltaic charging of lithium-Using photovoltaic charging of lithium-
polymer battery at 3.7Vpolymer battery at 3.7V
Base system hardwareBase system hardware
• An ARM7TDMI-S microcontroller coreAn ARM7TDMI-S microcontroller core• Extend to 8 serial ports by adding Extend to 8 serial ports by adding
Xilinx CoolRunner-II CPDLXilinx CoolRunner-II CPDL• Mini-SD card and FRAM supply data Mini-SD card and FRAM supply data
and configuration storageand configuration storage• Running software package developed Running software package developed
in C using WinARMin C using WinARM
CommunicationCommunication
• AC4790 900MHz modules operate at AC4790 900MHz modules operate at 76.5 kb/s76.5 kb/s
• Modem uses MX614 Bell 202 Modem uses MX614 Bell 202 integrated circuitintegrated circuit
Sensing nodeSensing node
• Measuring rainfall, air temperature, Measuring rainfall, air temperature, water pressurewater pressure
• Log dataLog data• Compute data statistic over each Compute data statistic over each
hourhour• Analyze data for potential sensor Analyze data for potential sensor
failuresfailures
Communication nodeCommunication node
• Computation of predictionComputation of prediction• Maintain a record of all values and Maintain a record of all values and
examine data correctionexamine data correction• Request data if encountering Request data if encountering
prediction model uncertaintyprediction model uncertainty
Office and community nodeOffice and community node
• Maintained by governmental Maintained by governmental agenciesagencies
• Display malfunctioning nodesDisplay malfunctioning nodes• Provide data and prediction Provide data and prediction
regarding the event of interestsregarding the event of interests• Community nodes provide a simpler Community nodes provide a simpler
UI and do not supply detailed UI and do not supply detailed information regarding node status information regarding node status and private data and private data
Installation and resultsInstallation and results
• Test the flood prediction algorithmTest the flood prediction algorithm using a large set of physical river flow using a large set of physical river flow
datadata• Demonstrate long-term data Demonstrate long-term data
collection of river flow data with a collection of river flow data with a sensor networksensor network
• Test the networking capabilities of 2-Test the networking capabilities of 2-tier sensor network in a rural settingtier sensor network in a rural setting
Blue River testingBlue River testing
• Use a large data set to test prediction Use a large data set to test prediction algorithmalgorithm
• 7 years of data measured from 1 river flow 7 years of data measured from 1 river flow and 6 rainfall sensors and a weather stationand 6 rainfall sensors and a weather station
• Autocorrelation at 24 hours: 0.627Autocorrelation at 24 hours: 0.627
Blue River testingBlue River testing
Dover field testDover field test
• 5 weeks data5 weeks datacollectioncollection
Honduras field testsHonduras field tests
• Collaboration with FSAR to install the Collaboration with FSAR to install the system and understand deployment system and understand deployment issuesissues
• Radio antennas need line-of-sight Radio antennas need line-of-sight high in the airhigh in the air
• Possible water measuringPossible water measuring systemsystem
ConclusionConclusion
• Described a complete architecture for predictive Described a complete architecture for predictive environmental sensor networks over large environmental sensor networks over large geographic areasgeographic areas
• Nodes-limited and cost constraintsNodes-limited and cost constraints
• Implementation of flood prediction algorithm and Implementation of flood prediction algorithm and evaluationevaluation
Pros&ConsPros&Cons
• ProsPros– A complete studyA complete study– Use off-the-shelf devicesUse off-the-shelf devices– Detailed hardware descriptionDetailed hardware description
• ConsCons– No real flooding occurred during evaluationNo real flooding occurred during evaluation– Energy consumption problemEnergy consumption problem