Wireless sensor networks and environmental monitoring applications

Transcription

1 Wireless sensor networks and environmental monitoring applications LE BORGNE Yann-Aël ULB Machine Learning Group 1050 Brussels Belgium Group site: Personal site: WSN lab site: Work supported by the COMP 2 SYS project, sponsored by the Human Resources and Mobility program of the European community (MEST-CT )

2 ULB Machine learning group Head: Gianluca Bontempi 7 researchers Research topics: Machine learning, data mining, forecasting, modelling and simulation Theoretical research: Local learning, feature selection, model selection Speech recognition and text classification Bioinformatics Gene expression and cancer detection Computer-aided medicine Data processing in wireless sensor networks Facilities: Cluster of 16 computers 30 ultra low power wireless sensors Lego robotics lab (10 Mindstorms kits) More to come (mobile robot, Mindstorms NXT kits, 100 wireless sensors) Website:

3 Agenda Technology and applications Solbosch Greenhouse monitoring WSN: research challenges

4 Wireless sensor Tiny electronic system (aka mote) that can collect, process and communicate data 5 components Microprocessor Memory Radio Sensors Energy supply Smart dust project 2001 UC Berkeley Goal: scale wireless sensors down to 1mm3 Golem and deputy dust 16mm 3 circumbscribed volume Microprocessor: 4MHz Radio: 4kbps, 180m Sensors: Light and accelerometer Energy: Solar powered

5 Current research prototypes Examples Tmote sky (MoteIV) TI MSP430 8MHz 48KB prog flash, 10KB RAM, 512KB data flash 250kbps radio Light, temp, hum sensors Mica2Dot (Crossbow) Atmel AVR 8MHz 128KB prog flash, 4KB RAM, 512KB data SRAM 38.4kbps radio Eyes node (EU project) TI MSP430F149 5MHz 60KB prog flash, 2KB RAM, 4KB data flash 115kbps radio µchip (Particle computing) PIC12F675 4MHz 1.4KB prog flash, 64b SRAM, 128B data flash 19.2kbps radio Light, temp, movement

6 Wireless sensor networks Real-time monitoring of an environment Throw n play : WS scattered on a field Self-organization Data collection

7 Examples Habitat monitoring - Great duck island Coordinator: Intel Berkeley and College of Bar Harbor Study Leach s Storm Petrel nesting habits Over 150 Mica nodes deployed for 4 months Tungurahua Volcano Coordinator: Harvard University Real time survey of volcano activity 16 Tmote deployed for 2 months

8 Examples Ocean tracking network 2006 Coordinator : Dalhousie university - Canada Monitor endangered species Understand animal migration patterns CitySense Coordinator : Cambridge - Massachussets Monitor pollution in cambridge 100 wireless sensors operating on streetlamps

9 And more Applications in a wide variety of domains: Environmental monitoring Forest fires, pollution, agriculture, Civil engineering Building, pipeline monitoring Defense Battelfield surveillance Industry Failure prevention, failure diagnosis Medical healthcare Remote or non invasive monitoring

10 Solbosch greenhouses Greenhouses used by different research labs Locations of sensors within the greenhouses 3 greenhouses monitored with 18 Tmote Sky for a two day period

11 Data variation profiles Temperature ( C) Humidity (%) Sensors: Temperature, humidity, and light on two bandwidths Sampling period: 1 minute 1440 readings collected over 1 day PAR (raw) TSR (raw)

12 Data summary Variations within greenhouse 3 Range temp ( C) Differences up to: 12 C in temperature 15 C in humidity Range hum (%) Range PAR (raw) Main cause: Sun position Ventilation paths Range TSR (raw)

13 Data summary Variations between greenhouses 3 and 4 Diff temp ( C) Plots give the differences of averaged measurements between greenhouses n 3 and n 4 Diff hum (%) Greenhouse 3 takes more sunlight than greenhouse 4 Diff PAR (raw) Diff TSR (raw)

14 (Some of the) WSN challenges Mote Operating system: Low memory footprint, ease of programming, Network Self organization: synchronization, routing layer, Data Information extraction: Data compression, data prediction, Constraints: Limited computational, memory, network bandwith and energy resources If run continuously, Tmote s lifetime is about 5 days

15 Operating systems (OS) Resource constrained. Memory is about 100 s KB, forget Linux. Several OS research projects: TinyOS: Initiated at Berkeley university Mantis: Colorado University SOS: University of California Contiki: University of Sweden And some ad hoc solutions

16 Synchronization Energy consumption and duty cycle Component CPU Power consumption 3mW With 2AA batteries, operation time is about a few days if radio and CPU on Radio receive Radio transmit Flash read Flash write 38mW 35mW 7mW 27mW Sleep 15µW Tmote Sky power consumption Possibility to switch components on and off (CPU, memory, radio) Use of duty cycle: Motes are in a sleeping state most of the time Regularly wake up, take measurements, and forward or send packets Example: Duty cycle of 20% would increase mote lifetime by a factor 5

17 Routing Example: Routing tree Routing tree: Oriented graph overlaying the network Each node chooses a parent According to a given metric (nb of hops, energy, load) Retrieve data from its children, and forwards it to the parent Relies on synchronization for maximizing mote sleeping time

18 Data processing Approximate data queries In most cases, only an approximation ε to the real measurements is desired Temperature within ±0.5 C Humidity ±2%... Dual prediction scheme (DPS) Policy: Send a measurement only if it differs from the previous one by ±ε Sends current measurement s t and the prediction model parameters α Tries to guess upcoming measurements: Prediction model h : ŝ t+1 =h(s t,α) Retransmit or apply the prediction model h to get the measurements Prediction model: Examples Constant model (CM) : ŝ t+1 =s t Autoregressive model of order 2: ŝ t+1 = α 1 s t +α 2 s t-1

19 Example with greenhouse temperature data Out of 1440 samples (1 sample every minute for one day) ε=0.5 C CM: 105 updates AR2 : 75 updates temperature ( C) Real data Approximated data Update sent temperature ( C) Real data Approximated data Update sent

20 Example with greenhouse temperature data Out of 1440 samples (1 sample every minute for one day) ε=1 C CM: 21 updates AR2 : 45 updates temperature ( C) Real data Approximated data Update sent temperature ( C) Real data Approximated data Update sent

21 Adaptive model selection The goal of DPS is to reduce the network load The network load incurred by a model depends on: The average number of updates The number of parameters to send In most settings: No a priori knowledge on the best model to use Adaptive model selection (Le Borgne et al., 2006, submitted) : Simulation on the mote of several models, including the CM Online assessment of the competing models Use of the best performing one when an update is required Online selection procedure based on racing to ultimately reduce the number of competing models to one.

22 Data processing Data compression Data compression for distributing network load Network load unbalanced: Node further away transmits 1 packet/cycle Node next to the server transmits 6 packets/cycle Lossy compression, yielding k packets/cycle for all nodes. Le Borgne et al, 2007, submitted

23 References Handbook of sensor networks: Compact wireless and sensing systems. M Ilyas, I Mahgoub and L Kelly CRC Press, Inc. Wireless sensor networks: An information processing approach. F. Zhao and L. Guibas Morgan Kaufman publishers. Adaptive Model Selection for Time Series Prediction in Wireless Sensor Networks. Y. Le Borgne, S. Santini and G. Bontempi. Submitted to the Special Issue on Information Processing and Data Management in Wireless Sensor Networks, Journal of Signal Processing, Elsevier. Simulation architecture for data processing algorithms in wireless sensor networks. Y. Le Borgne, M. Moussaid, and G. Bontempi. Proceedings of the 20th Conference on Advanced Information Networking and Applications (AINA), pages IEEE Press, Piscataway, NJ, ULB MLG wireless sensor WebSite

24 Thank you for your attention Questions? Group site: Personal site: WSN lab site: