15 October 2008

'Hall of mirrors' key to wireless reliability

Dr. Kris Pister, professor of electrical engineering and computer sciences at the University of California Berkeley, had a wild academic idea that turned into a commercially relevant technology—at least it is getting there. The concept of smart dust resonated with academics, the military, and analysts, who predicted fast success—an $8.1 billion market for wireless sensor networks by 2007. While the success has not happened as fast in industry as they had hoped, Pister and his team have forged ahead. He has moved from academics to chief executive and now chief technical officer of his own company, Dust Networks.

“I’ve had a passion for the past 20 years to make micro-robots,” he said. “The MEMS in micro-robots lead to a workshop on the future of size, power, and costs going down, following an exponential curve.” Step one was trying to get it to work.

“In 1999, we built sensor platforms. We were also working on a miniaturization path. In 2001, we built the smallest wireless sensor node to date,” he said. “We put a lot of effort from graduate students into system integration and miniaturization.”

What Pister learned was his technological innovation had no commercial relevance, “but it also forced us to start thinking of issues surrounding doing ultra-low power of system integration,” he said. “At the same time of miniaturization, we continued with the off-the-shelf approach. We used radios—cheap, easy, off-the-shelf radio frequency (RF) systems. There was a big interest in cheap, easy, RF with research centers and industry.”

With low-costs sensors, the application space can cover everything from home electronics and healthcare to industrial process control. When Pister coined the phrase smart dust, I had no idea industrial process was where it would see its first commercial success.

Companies were saying, ‘If you build it, we will buy it.’ Unfortunately, the marketing side took over more than the engineering side. Standards are coming out,” he said, “and yet for some reason, it wasn’t taking off. So I claimed with proof that wireless could actually change everything.” So why didn’t it happen?

“Sensors and computations are riding these Moore’ Law curves to zero power cost, reducing the cost over time,” he said. The problem today is conduit, copper, and labor are going up every year, he said.

On the plus side though, people were excited. And with good reason. It was happening, it was just happening slower. Academics were building on a tiny system. We called them Berkeley motes. And this got industry people interested.

Multi-hopping key

The key to building a reliable network is forming a multi-hop communication infrastructure. “You can pull the battery out and for a second, a piece of the tree would disappear, but then it would reform and be a self-healing, self forming network,” he said.

The secret is to make wireless sensors look like the Internet. Form multi-hop networks. Respond to problems in the network, and deliver answers to those problems. “You can now go in and stick sensors in places you want them to be and not worry about site surveys and directional antennas,” he said. We are hearing from customers, they are seeing 90% reduction in installation costs in sensors. So what are the barriers to industry adoption?

The number one concern is reliability. After that is standards, ease of use, and power consumption. Wireless communication is great, people think, but if they have to put in an infrastructure of power routers, etc., power consumption kills it.

“I set out to solve this as an academic. Node size was totally irrelevant. But implications for power were important,” he said.

Save power; turn radio off

It is easy to save power by turning off the radio. The challenge is knowing when to turn the radio back on. “Because it’s time synchronized, it’s more straightforward to turn a radio off and on when you need it,” Pister said.

“A lot of people understand what a mesh is,” he said. “There’s lots of connectivity. When people start out with meshes, they end up with trees. So you have lots of single-point failures. You have to build meshes, not trees.” The point is you will have a reliable network if you have many links to choose from.

If mote A needs to communicate to mote B, if they share a network, A knows B wants to talk to it on a schedule. A knows when it expects to see the first packet from B. It wakes up a little bit early (a millisecond or so) and stays awake a little bit afterward. If it has not heard the first bit, it knows B did not have anything to say or there was interference. So A can go back to sleep. So A does not spend a whole lot of power waking up and listening. You get this instant-on capability. The network looks like it is on. That lets A reduce power when it is listening. If it does hear from B, it measures when that packet arrives. And A transmits back an acknowledgment: “Yes I got the packet,” or “Yes I got it but I can’t hold it.” A knows what time B thinks it is based on when the packet arrives.

That is how you keep your network synchronized to 100 microseconds or so.

“This environment looks like a hall of mirrors,” he said. “If you have a mesh and don’t need to see that one particular device, if you can see anybody, you can get your data out. But that mirror lets you get more information out. If the facility changes—if a truck drives by or a door closes, you’re continuously optimizing that mesh, continuously monitoring your reliability. You can’t guarantee you’ll get every path through. But you keep trying. The key is diversity.”

And while security was not on the first list of the survey, “once you deal with reliability and power consumption, then next up is security,” he said. The good news is the IEEE standard built that in from day 1. After 9/11, security was on everyone’s mind. The top three are confidentiality, integrity, authenticity. 

You also have to make it easy to use. Make sure the amount of security does not interfere with what people do. “People say I want to be able to turn security off. I try to tell people they don’t want that. I don’t ever want to ship a network where somebody turns off security and something bad happens. All out networks have security all the time,” he said.

First deployments

The first deployment happened in the university, “where people cared more about data than networks,” Pister said. There was an earthquake engineering center—a shaker table, with a full-scale house had 50 regular wired accelerometers and two weeks of wiring. “We put in twice as many sensors at a lower cost in the afternoon. They found the wireless sensors gave the same amount of data, but there was twice as much data.”

In 2000, California was having energy problems. “We needed to do something energy relevant. In an afternoon, we put in 50 temperature sensors in the electrical engineering building. We set up a multi-hop network and started reporting data. The cost was $100; we were putting in new wired sensing network at the time. We showed we could use that data and simulate a power emergency check; there were no temperature problems in the building. We showed all that worked in 2000.”

—Ellen Fussell Policastro