Distributing the Power Wealth

For optimal results in sensor networks, sometimes you spend most on the worst sensor, and sometimes all on the best

At a concert, a gaggle of people spread throughout a vast audience area is armed with portable microphones that can send the sound to only one place, though they can also receive and relay sound from other microphones. Far away, a central desk is trying to reproduce the music as clearly as possible.

A new paper by communication theorists from the Viterbi School explores an electronic version of this situation in the context of sensor networks. These sensor networks are schools of interconnected sensor units, each of which is capable of gathering data and sending it on to other sensors or to a central data point. The sensors can also relay the measurements of neighboring sensors.

According to the authors, Urbashi Mitra, a professor in the Ming Hsieh Department of Electrical Engineering, and graduate student Gautam Thatte, the solution is a curious blend.

“If your measurements are really good, give all your power to the sensor with the best channel,” says Mitra. “As the measurement quality degrades, you start allocating more power to sensors that have worse channels.

“But if my measurements are rather poor, I need to listen to all of them, so I want to give more resources to ones that have less power. In the first case, the winner gets everything. In the second, I give the most to the one who has the least, because I need everyone to participate.”

The central node to which the individual sensors are reporting makes the decisions, and the key to this decision-making is the limited amount of power available for reporting from sensor networks. Allocation of power is the subject of intensive research as sensor networks grow in importance, performing numerous tasks, military and civilian, that range from environmental monitoring to intrusion alarms.

Such networks have highly limited power available to be used both for data gathering (sensing) and for communication. Communication typically takes the lion’s share of the budget—up to 70 percent—but the universal goal in such sensor networks is to minimize that figure.

Coincident and part of the process of power distribution is the selection of a route for sending the signals back to the central processor. The title of the paper, “Sensor Selection and Power Allocation for Distributed Estimation in Sensor Networks: Beyond the Star Topology,” summarizes the choice. “The well-analyzed ‘star topology’ has all sensors reporting directly to a single ‘fusion center,’ where the data is analyzed. But many other routes are possible, including ones in which sensors relay data from one to another, like racers handing off a baton, to get it to the fusion center; or a modified branch-and-tree relay in which one sensor receives a group of three, four or five batons to pass along to the center.

“These different architectures imply different strategies for passing along the data—the branch-and-tree architectures demand a straight ‘most-for-the least’ strategy, while the many relay systems demand a modified one.”

Mitra gives credit to Thatte for developing a number of interesting theoretical analyses in the paper.

Exact solutions are not always easy to find, says Mitra. “However, we have developed some bounds for which closed-form solutions exist. When we compare the overall sensing accuracy of the network using the approximated power allocation, we are very close to optimal.”