SCIENTIFIC AMERICAN www.sciam.com Smart
Sensors to Network the World By David E. Culler and Hans
Mulder
An emerging class of
pillbox-size computers, outfitted with sensors and linked together
by radios, can form perceptive networks able to monitor a factory, a
store-even an ecosystem. Such devices will more intimately connect
the cyberworld to the real world.
Today we
coddle our computers. They are fragile and expensive, so each
typically belongs to an owner who looks after it. When we need to
connect many of them into single system, we hire experts and set
aside large amounts of time and money for the job. The sheltered
cyberworld of computers still hardly intersects with the real world
of birds and trees, ships and bridges.
Where the two worlds
do connect, it is often because people have carefully altered
objects and methods of work to be computer-friendly. Stores stick
bar codes on everything they sell or ship. Warehouse clerks attach
radio-frequency identification (RFID) tags to pallets. Tagged goods
must then funnel through a few scanners so that the computers can do
their accounting.
A new class of microelectronic devices
frees us to mix computers much more freely with the objects and
places of everyday experience. Our research groups at the University
of California at Berkeley and Intel, as well as at start-up firms
and other universities, have joined simple computers to radio
transceivers and sensors to form small autonomous nodes that we call
"motes." Running an operating system known as TinyOS, each mote
links up with its neighbors from the moment it is turned on.
Although these smart sensors have limited power and processing
capabilities, an assembly of hundreds of them can spontaneously
organize into a perceptive network that is spread throughout the
physical world, able to perform tasks no ordinary computer system
could.
These wireless gadgets are affordable and sensitive
enough, for example, that dozens have been strapped to redwood limbs
to form a new kind of scientific instrument-we might dub it a
"macroscope"-that records the microclimate around an entire tree in
each of several parts of a forest. The battery-powered motes are
small enough that this past summer biologists placed 150 of them
within and outside the nests of seabirds to help ecologists learn
why they brood their eggs where they do. In addition to collecting
and processing data, wireless nodes deduce how to route information
through their neighbors so that it efficiently reaches an
Internet-connected base station. That capability allows Intel to
envision placing thousands of such sensor nodes in its manufacturing
plants to monitor critical machinery and prevent costly outages.
It is easy to imagine that as the price of motes fall and
their capabilities rise along with the rest of semiconductor
technology, this novel class of machines will be put to myriad uses:
boosting productivity, opening fresh avenues for scientific
research, and enabling creative ways to prevent and respond to
emergencies, environmental troubles and military engagements. But we
do not underestimate the difficult engineering required to realize
this potential. A mote is not a miniaturized PC; every aspect of the
system, from the way it runs programs to the way it communicates
data, must be optimized to conserve power, space and cost. A rule of
thumb in designing motes and their networking protocols for
long-lived applications is that each device should sleep 99 percent
of the time and do its energy-consuming work in the remaining 1
percent.
The natural world is not
computer-friendly. To function outdoors and in industrial settings,
computers must be "hardened" with enclosures to protect the
electronics from weather, soil, wild animals, and jolts. But sensors
must be exposed to the environmental conditions they monitor. Motes
have small, inexpensive shells and use redundancy to increase their
reliability.
They are designed to be inexpensive enough for
deployment in large numbers to gather very detailed information
about the environment. Networks of them are dense enough that it is
acceptable if some fraction die and smart enough that the overall
system can adapt to the loss and keep working. Designing for loss
and the uncertainty of the physical world presents new challenges
but allows perceptive networks to be economical, portable and
unobtrusive.
While designing successive generations of motes
and their networking capability, we have conducted pilot projects to
help identify how the technology needs to evolve to be most useful
for various applications. Several years ago, for example, we began
working with biologists on studies of flocks of about 18,000 petrels
that live at sea but fly inland every summer to lay eggs and rear
their chicks on Great Duck Island, a small, uninhabited isle off the
coast of Maine. The birds nest in underground burrows, which cluster
around particular places on the island. Understanding why they
choose the brooding spots they do may improve coastal wildlife
conservation strategies.
As with many aspects of biology and
ecology, that matter are local environmental conditions. A petrel
does not dig a burrow where it does because of the average
temperature or wind speed on the island but because of how warm or
windy it is at that particular spot. Other variables are probably
important, too, so biologists would also like to measure humidity
levels and the amount of light-both inside each burrow and just
outside of it. And investigators want to observe these factors over
the nesting season to learn how they correlate with the presence of
eggs and the habits of parent birds.
Since 2002 we have been
using motes to study the petrels' nesting behavior. The biologists
are asking a lot of the technology: to work well for this
application (and many others like it), each mote must carry a suite
of sensors. In this case, temperature, atmospheric pressure and
humidity sensors record microenvironmental conditions, while passive
infrared sensors detect the presence of warm birds and eggs. Yet the
device must be no more than a few centimeters in size so that it
does not disturb the bird and its chicks Clearly, it must be
wireless, because it is not feasible to string power and network
cables over acres of nesting grounds. So the device must carry its
own energy, enough to power the electronics for the annual nesting
season. And it must keep running and communicating its information
through other nodes in the network without any human contact.
Many of the system's design constraints boil down to power.
A single bulb on a strand of Christmas tree lights consumes about
half a watt. Whether they use batteries, solar cells or gadgets that
harvest energy from vibrations, as self-winding watches do, motes
must operate on 1/10,000 of this power on average.
A solar
cell that is one square centimeter in size generates about 10
milliwatts (thousandths of a watt) in full sunlight; solar cells
work poorly indoors and not at all inside burrows. A typical
coin-size battery stores about three watt-hours of electrical
energy. Microcontrollers generally burn about 10 milliwatts of
power; low-power radios burn about 20 milliwatts. Many useful
sensors consume similar amounts of power. Even running at a mere 30
milliwatts, however, such a battery will last for less than five
days.
That is why motes spend about 99 percent of their
lives "sleeping" in a standby mode that drops the power consumption
to a few millionths of a watt. Several times each second, the device
flicks on its radio to check for incoming messages, but if there are
none, the radio is shut off within milliseconds. Similarly, the
sensors usually take their readings of the temperature, light level
and so on only once every few minutes.
Most techniques for
saving energy exploit the intelligence within the device to perform
local processing and to turn off unneeded resources. We often use a
simple, low-power sensor to turn on others in response to a
preprogrammed stimulus. When a bird enters a nest, for instance, the
temperature rises quickly. A heat-sensitive circuit could take
readings once a minute and trigger a camera or other power-hungry
sensors on the mote to start recording whenever the burrow warms
rapidly.
The onboard processor offers other ways to save
power. Communicating one bit of data through the radio transceiver
costs as much energy as executing roughly 1,000 processor
instructions. The mote can conserve power by storing and aggregating
sensor readings, rather than sending them out immediately. The
processor can also compress information before it is sent and can
summarize the sensor logs with an average or the high and low values
if the details are not crucial. Nodes may swap sensor data with one
another, identify important observations and then send simplified
descriptions out to the user. There is no way around certain
network-protocol conversations between nodes, but these messages can
be held until there are sensor measurements to transmit and then
stuffed into the same "envelopes" as those packets of data.
The project on Great Duck Island successfully tested these
and other ideas for making the most of wireless sensor networks on
this scale. And in the 2002 breeding season alone, the macroscope
there was able to take more than a million measurements, adding far
more detail to biologists' picture of a key scene in the life cycle
of petrels. Just as important, the technology allowed scientists to
observe the birds without alarming them by a human presence.
OVERVIEW/PERCEPTIVE
NETWORKS
Thumb-size computers called
motes combine microprocessors and memory with radio
transceivers, onboard power supplies and a variety of
sensors.
Motes are inexpensive enough to
deploy by the thousands in factories, farms or wildernesses.
Each mote can collect and analyze sensor readings
independently but can also link up with neighboring motes in
a meshlike perceptive network.
Motes are already being
manufactured by Crossbow, Intel and others. Early prototype
systems have helped biologists study seabird nests and
redwood groves. Perceptive networks are also being developed
to monitor vibrations of manufacturing equipment, strain on
bridges, and people in retirement homes.
Compared with a handheld PDA, an
individual mote
is a computational weakling. Each mote has a microcontroller that
can handle four million to 10 million instructions a second, whereas
a palmtop can whip through about 400 million a second. But unlike
PDAs, motes can join forces in ad hoc networks to form a system that
has greater computational power than its parts.
ANATOMY OF A
MOTE
Smart nodes combine processing and
memory capabilities with sensors, wireless communications and
a self-contained power supply. A prototype produced by Intel
Research is drawn below. Motes are typically designed in
stackable layers so that a processing layer can be connected
to a wide variety of sensors and power sources to suit a range
of applications.
PROCESSING AND
COMMUNICATIONS
Standard connectors allow
various combinations of processing, sensing and power layers
Integrated microchip contains a
12-megahertz processor, 64 kilobits of RAM and 512 kilobytes
of flash memory
Radio antenna is designed to
exchange data at 200 to 600 kilobits per second over a range
of up to 30 meters, using a 2.4-gigahertz frequency and the
Bluetooth protocol, which has posed an interesting technical
challenge
Multicolor LED indicates status
of iMote
POWER
Lithium-ion battery pack stores
approximately two watt-hours of electrical energy
SENSING
Temperature and humidity sensors
are integrated on a single silicon microchip. Sensor boards
are available to measure many phenomena-including vibration,
acceleration, sound, and atmospheric pressure-as well as to
read RFID tags and to interact with other wireless systems.
OTHER MOTE
PLATFORMS
Mica mote is being used in some
500 research projects, including a "Robomote" made at the
University of Southern California. Using motes that control
actuators, perceptive networks can operate machinery,
regulate indoor environments, and change the position of the
sensors in the system.
Mica2Dot mote made by Crossbow
incorporates a 900-megahertz radio transceiver, 640
kilobytes of memory and a coin-size three-volt battery.
Sensor layers connect to the processing board using pins on
the circumference of the device. These motes formed the
redwood-and seabird-monitoring networks.
Smart Dust prototype, developed
at Berkeley, performs many TinyOS functions in hardware
rather than in software. Thanks to its ultraefficient radio
and analog-to-digital converter, the five-square millimeter
device would be able to run on energy harvested from ambient
light or vibration.
In
April we assembled such a system by strapping 120 plastic-encased
motes to the trunk and limbs of redwoods at a grove near Sonoma in
northern California. The goal is to build a detailed picture of how
the microclimate enveloping such trees changes and how the trees
shape the local environment through their shade, respiration and
water transport. For this project, cost determines the density of
measurement, and power determines the lifetime. The network will run
for several weeks on AA-size lithium batteries. The larger
challenges in this case are collecting data from devices that are so
high up that they are out of radio contact with the ground and
reprogramming the motes as needed to test different hypotheses about
the interaction between the forest and the environment.
The
low-power, silicon microchip radios in the devices can transmit and
receive data about as fast as a dial-up modem, but their range is
limited to less than 30 meters-sometimes much less. In a forest, wet
wood and needles attenuate the signals. A mote stuck to a tree trunk
often cannot communicate directly with a neighbor on the other side
of the trunk just a meter or two away. To cope with these
limitations, a mote might beam its sensor readings to a mote on a
higher limb. From that node, the data packets could travel to motes
in the treetop and then continue hopping from one device to the next
down the far side of the tree, over to other trees on the edge of
the grove, and finally out for storage and analysis on a more
powerful computer. The sensor-network macroscope in Sonoma is
designed to relay its redwood measurements to a PDA-like cellular
device on the ground and then through the Internet to a server in
Berkeley, 70 kilometers away.
When a deployment involves
hundreds of motes, it is not practical to set up such multihop
networks by configuring each device individually, as is done in a
typical office or cellular network. For many applications of
perceptive networks-monitoring equipment, raw materials, and
products in a factory or on a farm, for example-the arrangement of
motes will be constantly changing. So the motes self-organize into
networks. Special algorithms running in each sensor node determine
how many hops it is from the server and which of its neighbors
offers the most efficient path to that collection point at any given
moment.
A
SELF-ORGANIZING SMART SENSOR NETWORK
A
perceptive network of smart, wireless sensors called motes
could help customs officials prevent weapons or contraband
from being smuggled in through ports. Each cargo container
might hold numerous motes able to self-organize into wireless
networks. Those on pallets inside each container could link up
with a node on the container wall, and that device could in
turn share data with motes on all the other containers on the
ship in an efficient, treelike network. The port official's
laptop computer thus needs to communicate with only one of the
containers to retrieve a summary of all cargo on the ship and
more detailed sensor readings for any anomalous containers
that might warrant manual inspection.
1. A mote on each pallet could use a
built-in RFID reader to record the identity and origin of each
box it carried. A node mounted on the wall of the container
could aggregate data from the pallets inside and se its own
sensors to note whether the container became too hot, cold or
humid; whether it was dropped or bumped; or whether material
was, suspiciously, added or removed during the voyage.
2. The mote nearest the laptop serves
as the root of the treelike network. It transmits the
official's request to nearby nodes, which answer and also
forward a copy of the request to their more distant neighbors.
As the devices trade messages, each calculates how many "hops"
its neighbors are from the root. Motes generally send their
data through the neighbor closest to the root, but if that
device is malfunctioning or too busy, the mote is intelligent
enough to choose an alternative path.
3. A program called TinyDB ("DB" for
database), created by Intel Berkeley and U. C. Berkeley, runs
on each mote and effectively hides the complexity of the
network from the user. For example, the customs official might
request that the network report each container's
identification number, its origin and destination, and the
dates it was loaded and last opened. Those opened en route
could be flagged. A shipping agent might use the system
differently, querying the network to identify any cargo that
was exposed to high temperatures or humidity, potentially
damaging the goods.
Mote-to-mote
communications are coordinated by an operating system on each mote
as well as by an application program that can run in pieces, with
different pieces on different nodes in the network. Standard
operating systems, such as Windows or Unix, are much too large and
processor-intensive for these tiny devices. That is why Culler's
group at Berkeley created TinyOS, an extremely compact,
network-centric operating system that is now "open source" and
maintained by a community of programmers using it in their own work.
TinyOS is stingy with power; it forces mote programs to shut
down except when certain events occur that warrant action. The
operating system is also highly modular. If a program needs only
certain functions from TinyOS, the nonessential parts of the
operating system are automatically removed from the mote. This
modular approach ensures that the program code fills as little
memory as possible, leaving more room for sensor data. Modules also
enhance the robustness of the devices by limiting how the distinct
parts of the software interact.
Perhaps the most challenging long-term
question raised by perceptive networks is how we can most
efficiently and reliably program the thousands of smart nodes that
may coexist in a system. This scale is no idle conjecture: Intel has
begun installing prototype nodes called iMotes on pumps and other
machinery at its Jones Farm fabrication plant in Hillsboro, Oregon.
About 4,000 places in such a facility hold equipment that should be
monitored for signs of wear and failure-so many locations that
currently engineers can check only selected pieces every one to
three months. That is not frequent enough. Not long ago a device
failure occurred between two vibration inspections at an Intel
plant, causing a costly interruption of operations. An entire system
of 4,000 iMotes could be created now for well under $1 million that
could provide hourly updates on the health of the plant's
infrastructure, with no need for roving engineers. But we have had
to think carefully about how to program and debug the network so
that it remains manageable as it grows to include thousands of
sensor nodes.
Because of the tight constraints on power use
and processor speed, a perceptive network functions differently from
the Internet and office LANs, where computers have individual names
and addresses and most messages are sent from one machine to a
specific recipient machine. In sensor networks, one node generally
broadcasts messages to many, with the intended recipients identified
by attributes such as their physical location or sensor value range.
Recently a team at Intel and Berkeley created software
called TinyDB that makes a perceptive network system function much
like a database. A user can "query" all the smart nodes at once with
a request for, say, any vibrations between 40 and 120 hertz stronger
than a certain level. The request enters the network at its "root"
node, which forwards copies to its neighbors and so on until all
sensors have received the command.
Motes that lack vibration
sensors may ignore the message; others may turn on their sensors if
they have been sleeping; still others may run a series of
calculations on the data logged in their memories, extract readings
that meet the requested criteria, and pass that information back to
the root mote for collection. All the user sees is a
spreadsheet-like list of the relevant measurements and locations.
Software running on a high-powered server could then perform a wider
analysis of the trends to determine which machines require
maintenance.
In the redwoods, biologists are most interested
in the dramatic temperature and humidity fronts that move up and
down the tree every day, creating powerful gradients that may drive
the flow of nutrients. To track these fronts, motes pool their data
and search for spatial patterns. As scientists and engineers learn
from their observations through the macroscope, they periodically
change the tasks the network performs.
To replace the
software on motes with updated versions, we have drawn on lessons
from Internet viruses and worms. A new program is packaged in a
special form and delivered to the root mote, which installs it and
"infects" its neighbors with the package. The upgrade makes its way
through the network like an epidemic, but it does so in a more
controlled fashion that avoids redundant communications and adapts
to the way that the motes are scattered in space.
This
reprogramming model immediately suggests one of the harder problems
in sensor network design: how to secure them against hackers,
viruses and eavesdroppers. TinyOS has built-in algorithms that can
authenticate the identity of motes. But for the system to work well,
keys must be distributed to a large number of small nodes in
reliable and uncomplicated ways. Malefactors can attack perceptive
networks using strategies that are different from what is generally
seen on the Internet. One promising way to defend the networks is to
treat the effects of an attack as essentially another form of noisy
sensor data, so the perceptive network as a whole will still
function even if a small fraction of nodes has been compromised. But
as with all forms of computer security, the protection of mote
systems will be a constant battle of wits.
As we gain
experience with this new kind of tool, we find that it fails in
unfamiliar ways. A sensor network is unlikely to crash outright, but
as some nodes die and others generate noisy or corrupt data, the
measurements of the overall system may become biased or
inconsistent. We and other computer scientists are working on
techniques to judge the health of a perceptive network by perturbing
the system in a controlled way and observing how the sensors
respond.
Over the next decade or so, wireless sensor nodes
and perceptive networks will probably evolve into a much less
distinct and less visible form. Devices will gradually migrate out
of their little boxes and will instead be incorporated directly into
various materials and objects. Many will draw energy directly from
the environment in which they operate. To the extent that these
kinds of computers infiltrate homes, workplaces, farms,
transportation terminals and shopping sites and are able to sense
the presence, motion and even physiological states of individuals,
they will raise substantial privacy concerns. Indeed, a discussion
about such technology has already begun over the use of passive RFID
tags [see "RFID: A Key to Automating Everything," by Roy Want;
Scientific American, January]. Privacy issues are quite
straightforward for many valuable applications-such as monitoring
vibrations in pumps, fatigue in beams or microclimate in forests-but
in other domains a careful balance must be struck to ensure that the
technology properly empowers the individual.
With
appropriate debate, these matters will surely be surmounted-mote
technology is too useful to ignore. By connecting us to the physical
world in ways not previously possible, it promises to advance
scientific pursuits and the businesses of manufacturing,
agriculture, construction and transportation.
PROTOTYPES
OF PERCEPTIVE NETWORKS
PURPOSE
SENSORS
NODES
ORGANIZATION
Observes weather and nesting
behaviors of seabirds on Great Duck Island, ME
Temperature, humidity,
infrared
150
Intel, Berkeley
Analyzes activity of residents
in elder care facilities in Portland, OR and Las Vegas,
NV
Motion, pressure,
Infrared
130
Intel
Antitank mines communicate and
reposition themselves to close gaps in a mine
field
Location, orientation,
acceleration
96
DARPA
Collects readings on
microclimates surrounding redwood trees
Monitors the performance of pump
and scrubber motors in a microchip factory
Vibration and RPM
70
Intel, Berkeley
Maps growth conditions and
susceptibility to fungal infections in a
vineyard
Temperature
65
Intel
Listens for gunshots and then
triangulates shooter position
Sound, shock wave,
location
45
DARPA, Vanderbilt
Records microclimates within
James San Jacinto Mountains Reserve, CA
Temperature, humidity, rainfall,
light, wind
30
U.C.L.A.
Monitors movement of
acceleration design
Vibration,
acceleration
Under design
Berkeley
DAVID E. CULLER and HANS MULDER have
collaborated for many years on wireless sensor node research. Culler
is professor of computer science at the University of California,
Berkeley, and was the founding director of Intel Research Berkeley.
For the past decade his research has focused on ways to combine
large numbers of computers to work in a highly coordinated fashion.
Hans Mulder is associate director of Intel Research and director of
the Intel Research network of university labs. He initiates and
drives research on ubiquitous computing and distributed systems.
Mica: A Wireless Platform for Deeply Embedded Networks. Jason
Hill and David Culler in IEEE Micro, Vol. 22, No. 6, pages 12-24;
November/December 2002.
Query Processing in Sensor Networks. Johannes Gehrke and
Samuel Madden in Pervasive Computing, Vol. 3, No. 1, pages 46-55;
January 2004.
The Emergence of Networking Abstractions and Techniques in
TinyOS. David Culler et al. in Proceedings of the First USENIX/ACM
Symposium on Networked Systems Design and Implementation. USENIX,
2004.
Great Duck Island monitoring network:
http://greatduckisland.net
TinyOS: www.tinyos.net
U.C.L.A. Center for Embedded Networked Sensing:
http://cens.ucla.edu