Ecosystem Inferential Models to Control Data Acquisition and

Assimilation

Paul Flikkema

Wireless Networks Research LaboratoryDepartment of Electrical Engineering

Northern Arizona University

January 2008

P. Flikkema

Outline of this talk

Introduction/Motivation Requirements & Applications

Example: WiSARDNet Design and capabilities Observations and lessons learned

Challenge: Available Energy is Limited Hardware/Software Networking Application

[Data + Model]-Driven Control of Network Activity Energy efficiency Data management

P. Flikkema

CA Coastal Redwoods

Fern mat

Wired sensing infrastructure

requires over 1 km of cable per tree

P. Flikkema

Wireless Ad Hoc Networkingof Physically-Embedded Sensors

Need:  Dense, minimally-invasive array of sensors to monitor microclimate variables such as temperature and light.  Standalone or wired sensor arrays are invasive and difficult to deploy and operate.

Opportunity: Wireless networking of the sensors dramatically improve coverage and spatial

density, and ultimately, our understanding of environments and ecosystems...

…while greatly reducing the total monitoring cost

P. Flikkema

Requirements –Environmental Sensing

Minimal invasiveness

Long battery life

Scientific accuracy Support of a broad

spectrum of probes Support transparent

incremental deployment

Scalable in network size and density

Ease of installation and maintenance

Support of internet connectivity

Terrestrial (cellular) Satellite

Rugged, weatherproof packaging

Low life cycle cost

P. Flikkema

Modular hardware design Dual-processor architecture Three-board stack

Gateway nodesuse memory/time board in place of sensor data acquisition board

2nd Generation WiSARD

Brain

Probe dataacquistion

Radio

P. Flikkema

G2 WiSARDNet Design

Communication and Networking 902 – 928 MHz ISM band Non-Coherent Binary FSK (NC-BFSK)

modulation Slow time/frequency hopping spread

spectrum via pseudo-random number generator

CRMA radio channel sharing algorithm Distributed control Local information Scalable

Power Management Monitor power status Report battery voltage Adaptive radio transmit power (under

development)

Scheduler Time-triggered s/w architecture Dynamic scheduling of communication

Dynamic Reporting Report on significant change only

User Interface Command line from PC User selection of ID, sample rates Rich on-line diagnostics

Self Organization Forms minimum power-cost tree from

gateway node Periodic search for new nodes Can add, move, or delete nodes

P. Flikkema

G2 WiSARD H/W Capabilities

Built-in probe interfaces 12-bit A/D conversion 4 temperature channels

thermocouple 4 light (PAR) channels

photodiode 2 general purpose probe

channels, two power outputs and two CCP modules (Capture/Compare/PWM)

Soil moisture Decagon Ech2oprobe

Serial communication withintelligent probes

Sap flux (Granier method)

Interface for multiple additional intelligent probes One-wire bus

Provision for external energy supplies Supports autonomous

switching between internal and external energy sources

Battery-backed solar

P. Flikkema

Temperature Transducer

Thermocouples Based on voltage potential between two bimetallic junctions Rugged; low thermal mass Accuracy: absolute error < 0.5 C (tests with water bath) Low cost

copper

constantin

Cold Junction

copper

copper

1-wire Digital Temperature Sensorfor cold junction compensation

Digital Output

ADCHot Junction

P. Flikkema

Light Intensity Transducer

Photodiode response tailored to photosynthetic active radiation (PAR) spectrum

LI-COR LI-190 Quantum Sensor (~$300)

Apogee Quantum Sensor ($180) Low-cost home-brewed probes using Hamamatsu GaAsP

photodiodes capped with acrylic diffusers ($50)

P. Flikkema

NAU PAR vs. reference standard

P. Flikkema

Soil Moisture Transducer Time domain reflectometer (TDR)

measures the dielectric constant of a medium via traversal time of electromagnetic pulse in transmission line buried in the medium

Expensive; relatively complex Low-cost probe (Decagon ECH2O)

ca. $80/probe measures the soil dielectric constant (proportional to volumetric water content) from

steady-state response to sinusoidal input can calibrate to soil type in lab using gravimetric water content.

Excitation

Signal

10 ms

Sample

P. Flikkema

Weather Station Peripheral

Vaisala WXT-510 smart transducer for above-canopy data

Wind Speed and Direction ultrasonic vector anemometer

Liquid/Hail Precipitation piezoelectric impact detector

Relative Humidity/VPD Temperature Barometric Pressure

P. Flikkema

Trial Deployment, April 2005: Grasslands site,

C. Hart Merriam Elevational Gradient

Microburst andbrief cloud cover

Data from One Site

Data from All Sites

P. Flikkema

SENSING Lots of dataloggers scattered in the environment They require energy Their measurements are error-prone

NETWORKING Transmitting the measurements to where they can be

used takes a lot more energy/batteries (than just taking a

sample) is very error-prone

Environmental sensor networks in 30 seconds

P. Flikkema

Sensors sample, communicate, and hibernate

Typically

iS = 500 iH

iC = 1000 iH = 2.5 iC

Sorry, another 30…

E v(iSqS iCqC iHqH )(t t0)

=>Software-level design strategies=>Hardware technology

P. Flikkema

1. Minimize useless radio operation transmitting when there is no relevant node to

receive listening when no relevant node is transmitting

Use communication energy only whentransmitting or receiving data

How can we improve the network?

2. Transmit only what is necessary to solve the problem of model/data inference

exploit spatio-temporal redundancy of the data use coding to protect data

Make sure the transmitted data is informative

P. Flikkema

Network-Level Strategies

How to communicate Modulation, synchronization, detection Power control

When to communicate Media Access

Who to send the information to Routing

Network Design Redundancy

P. Flikkema

Network-Level:Sharing the Wireless Medium

Key tradeoff: proactive vs. reactive coordination

Proactive coordination: avoid contention and/or collisions by advance assignment of resources Implies less contention, better energy efficiency

May not scale well Reactive coordination: recover from collisions or

react to contention Reactive mechanisms offer simplicity and poor energy

efficiency

P. Flikkema

Networking Algorithms - Summary

How to balance proactive and reactive coordination mechanisms? Energy efficiency Capacity

What is the computational overhead? How much information must be collected? What

is the cost? Is the algorithm scalable? How robust is it to interference and the varying

wireless medium?

P. Flikkema

Why not justuse satellites?

Need ground truth data Undercanopy measurements Biotic responses Resolution: individual organism

Sap flux Leaf-based measurements

P. Flikkema

How much does a sensor net cost?

Per Sensor Node Total

Node$70 (incl. packaging)

$70

Light2 x ($50 - $300)

$300

Temperature

(thermocouple)2 x $15 $30

Soil moisture $80 $80

Cost per node $480

One Station per Site

RH (VPD), rain, wind speed/velocity + solar pwr

$3000 $3000

Sap flux,

Trunk growth,

Leaf respiration,

Eddy

Covariance,

Seed baskets,

Tree ID/location,

Surveys,

Deployment…

P. Flikkema

Monitoring the planet…

Suppose we receive a grant of $100M(OCO ~ $300M)

Assume sensors cost $100 including installation

Assume have average of ~ 2 per hectare…(Current deployments have ~ 9 sensors/hectare)

P. Flikkema

Hardware Technology Prospects: Today and Tomorrow

Node hardware costs can be made small Leverage volume-manufactured chips

Transducers remain expensive Embedded in inhospitable 3D space

Possible to have redundancy in nodes, but not transducers

Installation/maintenance costs dominate Battery replenishment

P. Flikkema

Future

Node hardware costs can vanish Energy supply (provision and replenishment)

and harvesting difficult (but there is hope) Transducer costs will depend on size of

market Installation/maintenance costs fixed (?)

P. Flikkema

Need: Make the Network Design Application-Aware

Spatial-temporal modeling minimize average spatial density of nodes minimize per-node energy consumption

Models useful in collection, not just analysis

=> Injection of intelligence

(aka models) into network

P. Flikkema

Opportunities Abound…

P. Flikkema

Dynamic Inference of EcosystemData and Processes

Dynamic model/data-driven control of Sampling Communication Estimation and prediction Model inference

according to their relative values and costs:

Integrate spatio-temporal sensing with modeling and prediction in an adaptive framework

P. Flikkema

A Top-Level ArchitectureAccommodate limited in-network

resourcesComplement them with out-of-network

capability Relaxed energy constraints Massive processing power

available Latency

Both should support multiple concurrent models

ApproachAdaptive in-network joint estimation and coding(inner, fast-simple control loop)Supervisory out-of-network processing(outer, slow-complex control loop) – model inference

P. Flikkema

Energy

Fid

elity

Model-driven measures

Easy

Where to be

Goal: maximum predictive power at ecological model level for a given energy cost

Model-DrivenDynamic Optimization:

In- and Out-of-Network Control of Sampling and

Reporting

P. Flikkema

Towards Application Awareness

Example: multi-sensor time series of temperature.Can view as

streams of numbers use general-purpose source codes (e.g., delta modulation)

correlated spatio-temporal process use a parameterized statistical model to drive adaptive

space-time sampling and reporting high-level model input

sampling driven by the needs of ecological models e.g., leaf efficiency/tree growth model: sampling rate may

be high because of sensitivity of high-level model, even when dynamics appear slow

functional coding

P. Flikkema

A [brief] wordon ecological models

Non-linear differential/difference equations

Stochastic Coupled in space Coupled across

variables

nZds

dtR(t) ET[s(t)] LQ[s(t),t]

y(t) s(t) (t)

Embedded in hierarchical models Scaling in time and

space Initial step: known

multidimensional prior distribution

P. Flikkema

1. Minimize useless radio operation transmitting when there is no relevant node to

receive listening when no relevant node is transmitting

Use communication energy only whentransmitting or receiving data

Energy-Efficient Inference: Strategy

2. Transmit only what is necessary to solve the problem of model/data inference

exploit spatio-temporal redundancy of the data use coding to protect data

Make sure the transmitted data is informative

P. Flikkema

Application-Aware Design

Should be aware of communication costs Radios enforce simple channel models and have tricky performance

characteristics radio warm-up time & synchronization preamble are significant

overhead, so should look at censoring to save packets, not bits

Don’t send every measurement Will have multiple correlated measurements from which to infer the

measurement/model Take advantage of prior information when available

Consider energy/latency tradeoff

P. Flikkema

A First Step: Dynamic Reporting

+/-threshold: report measurement if

|x - x*| > Implies simple

prior distribution cost based on that distribution in-net computation

P. Flikkema

energy-constrained inference in sensor nets

fidel

ity

power consumption density (W/m2)

spatial granularity limit

slope = pwr density

efficiency limit

P. Flikkema

Managing Uncertainty and Errors

Network as instrument Measurement errors

Transducer noise Channel errors

bit errors and lost packetsA bit error can be worse than a lost packet

How can we manage them?

P. Flikkema

Across the network layers: coding

Removing and adding redundancy… Application-dependent goal: minimize energy consumption

Old way:Separate source and channel coding Suboptimum for short block lengths found in sensor

networks

Joint source-channel coding Optimize compression and channel error correction How to do this in a sensor network…

P. Flikkema

Example: Bayesian decoder/estimatorIntegrates Prior probabilistic spatial-temporal model Network characteristics: knowledge of

Symbol/packet failure rates Selective reporting (suppression/censoring)

Data Properties of FEC code

Exploit redundancy in data as encoded in model At destination, symbols are exploited as joint

information/check symbols

At symbol/bit level, received coded measurements must satisfy

Parity constraints from FEC codeEquality constraints from model

P. Flikkema

At the destination Received coded measurements form source-channel product codeword

SC code rate = f(FEC code, censoring policy)

FEC

checks

network

reliabiity

(symbol LLR’s)

Bayesian

decoder/

estimator

process/

data

modelcensoring

model

transducer

models

FEC

structure

rx coded

measurements

info

erasure

censor

P. Flikkema

Global source model + FECimproves inference

Bayesian

(MAP) limit

PCMP

Iterative

decoder

Channel bit errors

Source Model

MV Gaussian

3 measurements

2info = 1

2 = 0.4

FEC channel code

(7,4,3) Hamming

WSN Regime

with S. Howard

P. Flikkema

Review

Hardware/Software Network

Use radio only when communicating Communicate only what is informative

Application Management of uncertainty in an energy-constrained

distributed instrument Coding Use of information about sensed data, the network, and

models of both Joint management of data & models Inference in data collection, not just data analysis

P. Flikkema

Workshop Strategy

Focus on science needs Is the idea relevant to wirelessly networked

environmental sensing?

Understand the network architecture Data gathering networks Dynamically controlled DGN’s (commands)

Nonuniformly sampled models + data

P. Flikkema

Suggested Aims

Attack problem of “too much” modeling How do we handle need in 2030 for “raw” data NOT collected in

2010? Compendium of models

Non-uniform Space-Time Coupled Best practices

Model ontologies, interfaces, and standards Characterize

Transducer errors Communication errors

P. Flikkema

Enjoythe

Workshop!

Questions, Coffee, Discussion

P. Flikkema

Micro-ControllerPIC18F8720

PowerMgmt

MemorySRAMFLASH/FRAM

SPI+

SystemTime

BrainsBoard

Trans-ceiver Micro-

ControllerPIC18F452

RadioBoard Sensor

Board

AnalogI/O

1-

Wir

e

Tem

p(4

)

Genera

lPurp

ose

(2)

Lig

ht(

4)

PWM(2)

Pwr(2)

G2 WiSARD Functional H/W Design

Ext Power

Int Power

Enclosure

P. Flikkema

Mitigating Collisions

In CRMA, two types of conflicts: Soft conflicts occur when proximate cliques allocate the

same time-frequency-code slot Causes multiple-access interference that is mitigated by excess

bandwidth and randomized access A hard conflict can occur when the number of cliques

intersecting at a node exceeds the number of radios Nodes can use predictive conflict resolution (PCR) to look to the

future, identify hard conflicts, and resolve them

P. Flikkema

Sensor Networks: Design Space

How do realities of WSN implementations affect energy-efficient inference?

Star or multi-hop : logical or physical Is phase coherence possible? Most WSN radios are simple

Binary symmetric channel Is time synchronization maintained? Source and channel coding

P. Flikkema

Example: Precision vs. # of reporting sensors: linear estimate of a sampled random field

100

101

102

103

0

510

1520

0

5

10

15

# of reporting sensorscorrelation radius

Projection onto random

spatial basis

Minimal prior knowledge

(a priori mean and priori variance)

Flikkema ISCC 2006

P. Flikkema

Compressive Sensing

Uber-goal: use basis that gives sparse (efficient) representation (encoding) of information Project data onto basis

Compressive sensing theory shows that randomized basis can be efficient for data-gathering sensor nets

Excellent potential when no priors are available