OpenSense OpenSense Karl Aberer, EPFL Boi Fal6ngs, Alcherio Mar6noli, Mar6n Ve<erli, EPFL Lothar Thiele, ETH Zürich
OpenSense Overview • Research challenges • Research progress and results • Deployments • Conclusion
OpenSense Air Pollu6on • Air pollu6on in urban areas is a global concern – aﬀects quality of life and health – urban popula6on is increasing • Air pollu6on is highly loca6on-‐ dependent – traﬃc chokepoints – urban canyons – industrial installa6ons
OpenSense Air Pollu6on Monitoring • Precise loca6on-‐dependent and real-‐6me informa6on on air pollu6on is needed • Oﬃcials – environmental engineers: loca6on of pollu6on sources – municipali6es: crea6ng incen6ves to reduce environmental footprint – public health studies • Ci6zens – advice for outside ac6vi6es – assessment of long-‐term exposure – pollu6on maps
OpenSense Opportunity • Monitoring today – few sta6onary and expensive sta6ons – models that extrapolate from pollu6on sources Nabel sta6on Zürich – data mostly inaccessible to the public • Opportuni6es mobile nodes wireless ﬁxed nodes – wireless communica.on: deploy larger Nabel sta6on Zürich numbers of sta6ons – mobility: deploy mobile sta6ons – mobile devices: gather context informa6on and deploy applica6ons for GPRS ci6zens GPS
OpenSense Value of Dense Measurements • Tradi6onal approach • Recent results – Few sta6ons – Massive deployment of – Low resolu6on interpolated sta6ons (150) at street-‐level es6mates of pollutant (2008/2009 New York City concentra6ons across Community Air Quality massive regions Survey) – Pollutants of interest heavily concentrated along roads with high traﬃc densi6es
OpenSense Overview • Mo6va6on • Research progress and results • Deployments • Conclusion
OpenSense Research Challenge SENSING SYSTEM INFORMATION SYSTEM NANO From many wireless, mobile, … to reliable, understandable and TERA heterogeneous, unreliable raw Web-‐accessible real-‐Dme informaDon measurements … mobile wireless ﬁxed nodes Nabel sta6on Zürich nodes sensor network control opDmizaDon of data acquisiDon GPRS informaDon disseminaDon GPS • More data, more noise, but also more redundancy – Can we produce be<er quality data? • Exemplary use case for other environmental phenomena – Radia6on, noise, energy
OpenSense Technical Challenges • Wireless sensing devices – energy eﬃciency, data transmission and compression, sensors control • Mobile sensors – sampling under mobility, data collec6on and dissemina6on with mobile devices, freshness of data, stream data management • Community sensing – privacy protec6on, trustworthiness of data, relevance of data gathered and informa6on produced • Modelling – behaviour and mobility of sensing devices è sensor, device and mobility models – air quality informa6on from raw data è air quality models – behaviour, interests and mobility of informa6on consumers è privacy, trust and acDvity models
OpenSense What is the problem? • A measurement system such • Illustra6on as OpenSense is a complex 1. Node decides individually system depending on its state, e.g. – layers energy – dependencies 2. Nodes communicate WSN and coordinate – dynamicity 3. Base sta6on schedules nodes • Op6miza6on becomes a 4. Mobility model: a third node complex task arrives, don’t measure! – mul6ple op6miza6on 5. Air quality model: don’t need dimensions measurement! – many system components and 6. Privacy model: node 1 should layers measure! – feedback 7. Applica6on model (e.g. health no6ﬁca6on): no Two mobile nodes: measurement needed! who should measure?
OpenSense U6lity-‐based Control ApplicaDon model: Relevance and cost User acDvity model: Mobility and user state Trust and privacy model: Reliability and security Control: translate high level Data: translate low level u6lity to low level u6lity Air quality model: data to high level informa6on Sampling and correlaDon Mobility model: PredicDon Wireless sensor network: Local coordinaDon Sensors: Individual state
OpenSense Testbed Sensors Deployments • CO2, infrared based • Lausanne: buses • CO electrochemical • Zürich: trams • NO2 electrochemical • SO2 electrochemical • Basel: sta6onary wireless • O3 silicon based network • Fine par6cles mechanical Power suppliers pDr1000: ultraﬁne par6cles Sensorscope (FH Nordwestschweiz) DataLogger SHT75: air temp and humidity Telaire T6613: C02 Langan T15n: CO Sensorscope Smart Interfaces
OpenSense Sensor Behavior Open sampling Closed sampling Sensors directly exposed to Sensors exposed to measurand inside environmental measurand controlled chamber Beneﬁts: Beneﬁts: • simple & slim solu6on • absolute measurements • con6nuous sampling • noise due to environment ﬁltered Drawbacks: Drawbacks: • no absolute concentra6on values • complex & bulky • noisy signal • non-‐con6nuous sampling Typical response: Typical response: IDEA: Combine the two approaches and get the beneﬁts of both.
OpenSense On-‐the-‐Fly Calibra6on • Challenge: – Supplied calibra6on may not match project requirements – Baseline driq due to sensor aging • Approach: – Ini6al calibra6on using sta6onary, high quality instruments – When deployed periodic recalibra6on using mobile sensor nodes Original calibra6on performs with an Aqer recalibra6on average error of the average error 30ppb drops below 3ppb
OpenSense High-‐resolu6on measurement Interpola6ng measurements of Planned work two Opensense sta6onary • Measurements obtained along sta6ons the road network + anisotropic • A diﬀerence of 10m from road is diﬀusion on lines, tuned by traﬃc and popula6on density considerable (from mobile sensors) Sta6on 1058 Sta6on 1059
OpenSense Mobility Modeling Goal Appropriate tool: microscopic traﬃc • Simulate realisDc trajectories of simulators (SUMO, AIMSUN) vehicles Tes6ng diﬀerent control strategies before deployment • What is the marginal beneﬁt of adding an addi6onal vehicle/line to the system • Knowing the traﬃc pa<erns, is the system coverage suitable for regions with ﬂuctua6ng traﬃc (emissions)? • What is the eﬀect of a traﬃc event on the coverage of the system? 3D view of traﬃc simulaDon run in front of Lausanne Train StaDon, using SimLo model (LAVOC, EPFL)
OpenSense Route Scheduling • Given – Area of interest Ω (Zurich) – N measurement instruments • Each has a limited budget E – M tram and bus tracks • Ques6ons – Which subset of tracks (and trams) gives the best coverage of the city? – Which tram should measure over shared track pieces? • The program is NP-‐Complete
OpenSense Air Pollu6on Models • Forward Reasoning – Spa6al and temporal interpola6on of pollu6on levels – Advanced warning for dangerous levels • Backward Reasoning – Crea6ng an emission inventory – Iden6fying previously unknown sources • Meta-‐Reasoning – Op6mal sensor placement – Sparse sampling
OpenSense A Region-‐Based Model • Exis6ng grid-‐based models – computa6onally expensive for ﬁne grids – do not dis6nguish streets • Pollu6on dispersion is not uniform within a grid – Ground-‐level air pollu6on is heavily inﬂuenced by streetscape and land use – A region-‐based model may be more appropriate for OpenSense ADMS-‐Urban, London 2010
OpenSense Mul6-‐model Query Processing in Mobile Geosensor Networks • Approach Con$nuous Moving Queries – Middle layer produces a model Give a (in car) pollu6on update cover from a set of regression Aggregate Queries every 30 mins models on an area COX emi<ed yesterday in – Con6nuous sensor updates Lausanne center – Con6nuous and ad-‐hoc queries • Advantages Model-‐based middle layer – Handling spurious updates to the data base – Minimizes data storage – Query results useful from DBMS applica6on perspec6ve (storage of raw sensor values) Mobile Sensor Data Mobile Sensor Data (Pollu.on Values) (Pollu.on Values)
OpenSenseModel-‐Based Query Processing Over Uncertain Data what is the probability that Bob is at room 4 at $me 1? original data stream ↓ inference of Dme-‐varying probability distribuDon (dynamic density metrics) ↓ creaDng probabilisDc views (Ω-‐View builder)
OpenSenseModel-‐based Anomaly Detec6on original data stream ↓ approximaDon using user-‐selected models ↓ detecDng anomalies ↓ user conﬁrmaDon: anomaly is an actual error?
OpenSense Cloud-‐based Time Series Management • TimeCloud: A Cloud System for Massive Time Series Management • Key features – manages large-‐scale 6me series in the cloud – scalable, fault-‐tolerant – built upon Hadoop and Hbase – adap6ve data storage through par66on-‐and-‐ cluster – model-‐based cache for fast model-‐based views – model-‐coding join for fast distributed join based on bitmap representa6on of 6me series.
OpenSense Sensor Context Extrac6on Objec6ve: Automa6cally annota6ng trajectories of diﬀerent types of moving objects (cars, people) bus metro walking Seman$c trajectory home oﬃce market home Seman$c Annota$on Middleware Hidden Spatial Map Markov Join Matching Model region road network point of interest e1 e2 e3 e4 e5 e6 e7 GPS episodes
OpenSense User Privacy vs. Data Reliability • Mobile devices with sensing • Privacy protec6on mechanisms try capabili6es to break the link between data and its source – ParDcipatory sensing – E.g. environmental sensing, • Thus, there is a clear trade-‐oﬀ health-‐care monitoring, etc. between privacy and • Incen6ves for par6cipa6on trustworthiness of data sources – Privacy concerns • Iden6ty • Loca6on – Trustworthiness Sensor, air polluDon, mobility, behavior models used to esDmate reliability of data
OpenSense Privacy Protec6on Approach • Trust authority (e.g. telco) knows iden6ty and Aggregation trustworthiness of users Server• Aggrega6on server receives Trust trust-‐rated but privacy-‐ Scores Ratings preserving data – Anonymize data sources – Obfuscate data, loca6on-‐ or 6me-‐stamps Trust – Hide/add events Authority Honest and malicious Entropy as measure for measurements uncertainty about user clearly dis6nguished data remains high
OpenSense Overview • Mo6va6on • Research challenges • Research progress and results • Conclusion
OpenSense Deployment Status Basel/Sapaldia Sapaldia study Status • Swiss Tropical and Public Health • Calibra6on tests performed in Ins6tute of Basel University 2010 • Es6mate individual exposure • Sta6onary sta6ons will be indoors and outdoors delivered on May 18 Sapaldia will use sta6ons for indoor air quality monitoring
OpenSense Deployment Status Lausanne 2 prototype sta6onary sta6ons and 1 prototype mobile sta6on • Currently under tes6ng at EPFL • Mobile sta6on will be mounted on a bus on May 23 Measured parameters • NO2, CO (2 sensors), Humidity, Temperature, CO2 (only mobile sta6on) Power • Solar panel (sta6onary sta6ons) • Bus power (mobile sta6on) Data • Transmission via GPRS to a central server Sta6on 1058 Sta6on 1059
OpenSense Deployment Status Zürich • 1 node @NABEL sta6on in Dübendorf (for reference measurements): – Communica6on: GSM, WLAN – Sensors: 2 x O3, CO, temperature/humidity – GPS • 1 node on top of Tram in Zürich is in prepara6on 14 (mid. July 2011): – Communica6on: GSM, WLAN – Sensors: O3, temperature/humidity – GPS – Accelerometer • 2 further nodes in construc6on (July)
OpenSense Calibra6on of CO Sensor @EMPA Lab gas bo<le empty Ini6ally not calibrated calibrated
OpenSense Installa6on @NABEL Dübendorf Originally calibrated O3 sensor: correct trend, but wrong absolute value. Calibra6on required.
OpenSense OpenSense Visualiza6on Portal Visualiza6on Server GSN sensor data cache Signiﬁcant Change Condi6on Interpola6on Image grid cache
OpenSense OpenSense CrowdMap The data from NABEL sta6ons are already integrated. It is possible to add data via SMS, Email or online Form. Based on open source plaworm. OpenSense CrowdMap is not yet publicly available.
OpenSense Overview • Mo6va6on • Research challenges • Research progress and results • Deployments
OpenSense Conclusion • End-‐to-‐end system view crucial – Inves6gate all system layers: sensor – user interfaces – U6lity-‐based framework as integra6ve approach • Results applicable beyond air pollu6on – Complex, distributed, par6cipatory measurement • Involvement of Nokia – Personalized health applica6on • For more informa6on: opensense.epﬂ.ch
OpenSense Team • Karl Aberer, EPFL-‐LSIR, project leader • Alcherio Mar6noli, EPFL-‐DISAL, PI – Thanasis Papaioannou, postdoc – Chris Evans, PhD – Dipanjan Chakraborty, (on leave from – Emanuel Droz, engineer IBM Research India), visi6ng researcher – Adrian Arﬁre, PhD – Hoyoung Jeung, postdoc • Lothar Thiele, ETH Zürich, PI – Rammohan Narendula, PhD – Olga Saukh, postdoc – Mehdi Riahi, PhD – Jan Beutel, postdoc – Zhixian Yan, PhD – Jayashree Ajay-‐Candadai, PhD – Soﬁane Sarni, engineer – Alex Arion, PhD – Saket Sathe, PhD • Mar6n Rajman, EPFL-‐LIA, coordinator • Boi Fal6ngs, EPFL-‐LIA, PI – Jason Jingshi Li, postdoc • Mar6n Ve<erli, EPFL-‐LCAV, PI – Guillermo Barrenetxea, postdoc – Andrea Ridolﬁ, postdoc – Heather Miller, PhD