The Emerging Challenges in Sensor Network Research
|
"Imagine high-rise buildings in downtown Los Angeles that could detect their structural faults, then alert authorities on corrective action," begins Deborah Estrin, a professor of Computer Science at UCLA and a member of Calit²'s Advisory Board. This was the opening scientific challenge Estrin posed at a seminar she gave October 21 at the UCSD Computer Science and Engineering Department.
"Or what if Belmont School could reliably measure toxic levels at very low concentrations and trace the transport of a contaminant back to its source? Or what if buoys along the coast could alert surfers, swimmers, and fishermen to dangerous levels of bacteria?"
Back in the fall of 2001, Estrin, preparing for an impending site visit that was to take place shortly after 9/11, wondered presciently, "What if a building razed by an earthquake could be infiltrated with robots and sensors to locate signs of life?"
All of these could be potential applications for work underway in Estrin's new 10-year enterprise supported with $40M of NSF funding: The Center for Embedded Networked Sensing (CENS). In the words of one of Estrin's colleagues, "ENS' charge is to reveal previously unobserved phenomena." Their applications include studying seismic structure response, marine microorganisms, contaminant transport, and ecosystems/ biocomplexity.
"When we talk about 'embedded sensors,' that actually has a dual meaning," says Estrin. "Sensors can be embedded in the world doing in situ monitoring. But they're also 'embedded' in the traditional computer science meaning of that word, that is, they can be largely autonomous and independent of human interaction with the system."
The seismic application - structural response in the field and in the structures themselves - is their most developed application. Using, as a centerpiece, a 17-story, steel-frame building that's instrumented with 72 seismometers, Estrin's group is developing a 100-node grid of seismometers at 100-meter spacing across the UCLA campus and surrounding the building. Prof. Paul Davis, lead geophysicist on this project, seeks to use this grid to do dense monitoring of the ground and correlate the characteristics of the seismic wavefield with structural response.
"All data will be public-domain - all online and accessible," Estrin is quick to point out. "We're building systems to serve the scientists, but we want to move increasingly toward understanding how to build engineered systems that can be deployed more generally."
|
Video of Talk
|
 |
|
|
|
|
|
|
|
Their early work in this applications area has been lab-based. Says Estrin, "Harmon has carefully designed a 'sandbox' with various combinations of soil to collect experimental data upon which a model can be developed. Through the use of miniaturized nitrate sensors developed with Prof. Jack Judy, Harmon is now able to measure data at 100 collection points. Unless you go to that density of monitoring, you can't develop an accurate model." To underscore this point more generically, CENS' early focus in their various projects is on experimental work, not simulation, because they need to collect empirical data describing the environments for which they want to develop models.
Estrin points out that sensors can also be applied usefully in a non-networked environment. "Speaking of contaminant transport, why don't we all have sensors in our water filters at home?" she wonders. The main point is to position sensors close to the source of a potential problem, where the phenomenon is likely to have the greatest impact on people.
Next Estrin describes ecosystem monitoring: "Society makes decisions related to the ecosystem all the time," she says, "but those decisions are typically based on poorly understood processes. Ecology, as a discipline, has lacked the measurement tools to be highly quantitative, but they believe sensor technology will provide the discipline with some of what is necessary to make great advances."
Estrin points to the current issue of Forbes magazine (October 28, 2002) and the article titled "Sensors Gone Wild," which describes work underway by CENS at the James Reserve, located near Idyllwild, CA, where they are implementing a microclimate monitoring system. "It currently has just one tower, like at the Santa Margarita Ecological Reserve," run by the San Diego State University Field Station Program and affiliated with Calit². "We hope to use SMER as the next environment to instrument at cubic-meter resolution," she says. "The sensors are available, and that scale is right for what the scientists want to do."
But it's important to move beyond a focus of just collection of data from a lot of point sources. "We need to be able to identify points in time during a 24/7 period that are interesting," she says. "That motivates trying to build network systems that can identify anomalous phenomena in place and time, then focus image analysis on those points." Taking that one step further, Estrin points out that it's the network as a whole that acts as the sensor (Mages & Smith, ORNL, 10/98).
Estrin points to an important design theme cutting across the various projects in CENS: Developing long-lived systems that can be untethered (wireless) and unattended. That still leaves the problem of power, however: Devices, even though wirelessly enabled, are still energy-constrained, and communication is commonly regarded as the primary consumer of energy. How much more energy is needed to transmit one bit vs. that needed to compute one instruction might be a matter of debate, but everyone agrees that the ratio is large. The point is that the processing needs to be pushed out to where the data is being collected, then transmit back higher-level events or less data, filtering out what is not of interest. "We refer to this as 'in-network' data processing,'" says Estrin.
CENS is implementing self-configuring systems that have irregular configurations or network topologies that change over time. "Hand configuration will fail," asserts Estrin. "The solution must rely on local adaptation and redundancy. To do that, we're faced with addressing many challenges: Localization, time synchronization, calibration, information aggregation, storage, event detection, and developing appropriate programming models."
With respect to localization, it turns out that identifying location in time is easier than in space. Location in space, though, matters to enable building "contour maps" of the received data. In many contexts, GPS doesn't work; for example, it can be hindered by foliage or the fact that the sensing device is placed indoors, which means that the sensor can't get a good "fix" on the relevant satellite. "We're trying an acoustical approach to do ranging and multilateration, but it requires the support of fine-grained time synchronization," says Estrin.
CENS is experimenting with a tiered system design using iPAQs and "dust motes" developed at UC Berkeley. Localization of the sensors is done using acoustic ranging. The iPAQs being used are randomly "placed," and they "listen" for the motes' chirp to determine the separation (by measuring the time difference between when the chirp was emitted and received) between iPAQs and ultimately their locations on a two-dimensional coordinate system.
"This has been a very interesting collaboration between EE and CS students - combining acoustic signals and localization of the source," says Estrin. "We've done it in an area about half the size of a standard seminar room, and we're not yet sure how it will scale over larger distances.
"So, looking at the big picture," poses Estrin, "what if we want to task a 1,000-node dynamic sensor network to conduct a complex, long-lived task? We need to identify spatio-temporal, multi-modal events; and address scalability, energy, and communication issues. Is this a database question, an Internet peer-to-peer question, a parallelization question, an ad hoc network question, or…? Lots of people are looking at this from various perspectives, which has created a very intellectually stimulating environment for everyone."
And how is it possible to inject more complex behaviors into the network? Estrin's group has tried a directed diffusion approach, experimenting with a data- vs. a network-centric routing. "We're tried putting more of the processing and contingent behavior out in the network," says Estrin, "when the communication range is short. Using an address-centric approach to opportunistically aggregate data seems to be a 'win,' but trying to manipulate the tree to move aggregation points closer to the source of the data doesn't buy you much."
A sensornet can also be seen as a database problem, that is, using the sensornet as a database in the sense of supporting distributed querying and optimization, on the one hand, vs. storage, on the other. "In fact," says Estrin, "some of the more interesting research questions relate to distributed representation and storage of information."
Another issue is the need to know the state of the environment globally but not having the luxury of taking readings at each node and bringing all the data bits back to a central place. "So one of our goals is to build a multi-resolution storage architecture that captures interesting features but also allows drilling down to observe features in more detail as warranted by a given research question," says Estrin.
CENS has various systems under construction related to microclimate monitoring, triggered image capture, "canopy-net," and the seismometer grid at UCLA. They are also applying robotics to this endeavor because robotic elements have the potential to extend system lifetime, speed, and efficiency. They are also working on developing selected types of sensing devices, such as miniaturized devices that can detect nitrate, a common type of agricultural runoff.
"What we're trying to do," says Estrin in conclusion, "is team CS, EE, and sensor researchers to build common systems that can serve a variety of applications. This is very hard to do theoretically because of the fairly uniform lack of physical experimental data across the applications. So we're starting out by building carefully instrumented systems to collect the data."
Related Links