NASA's new way of tracking battery life

A NASA-led team has shown how a novel statistical algorithm could be applied to better predict the life span of batteries.

"Batteries represent complex systems whose internal state variables are either inaccessible to sensors or hard to measure under operational conditions," the researchers wrote in an article published in the August 2008 issue of IEEE Instrumentation & Measurement Magazine.

Anyone who runs a laptop PC on battery power can certainly testify that this is the case. The software meter could indicate that hours of battery life remain but only minutes later indicate that the battery is almost completely drained.

"Estimating the remaining life of any process is difficult because you don't know exactly how the component will be used in the future," said Kai Goebel, the senior scientist at the NASA Ames Research Center who led the research.

Goebel is leader of the NASA Prognostics Center of Excellence, which focuses on better ways of predicting the health of components of complex systems.

"Depending on the load conditions and depending on the environment conditions, a component can last longer or not so long," he said.

NASA system builders know that trouble all too well. And they also know that it is doubly difficult to estimate the remaining life in batteries under unusual operating conditions, such as extreme cold or heat, or under highly variable workloads.

Goebel and his team tested how well a statistics-based method, called the particle filter (PF) model, would work for estimating remaining battery life. They also tested more traditional techniques. The various algorithms estimated the life span of a set of 18650-size lithium-ion cell batteries that were subjected to fluctuations in temperature and use. (The dataset can be found at the center's online repository.)

They found that the PF model predicted a battery's future life with the greatest accuracy. What's more, the model also asserts the uncertainty of its prediction. For example, it can say that there is a 90 percent certainty that the battery will last for another four days. Such predictions would help system planners make the most of the remaining life of a battery, Goebel said.

"If you have a remaining-life estimate but your uncertainty bounds are incredibly wide, then the estimates in the worst case could be almost useless depending on the risk tolerance that the user is willing to take," Goebel said.

Although used elsewhere in industry and academia, the PF model has not been widely applied to estimating remaining battery life, Goebel said.

The PF model was designed to predict the future state of dynamically changing systems based on all available performance data. It uses Bayesian inference and a series of Monte Carlo simulations to generate predictions about future baseline states.

In addition to the PF model, the team also tested a series of simple statistics- and probabilistic-based models, all using some form of regression analysis. Regression is the art of finding an underlying pattern of behavior in a given set of data. The other models were good at predicting battery lifetimes in cases in which use was steady, but they did not provide adequate results with more varied use and conditions.

The downside of the PF approach is that it is more computationally intensive than the other methods, Goebel said. If you do not need your predictions to be highly accurate or you are limited by the amount of processing power you have, one of the other approaches might be more suitable.

The PF model could be particularly useful in cases in which a battery will undergo variable, or nonlinear, operating conditions, Goebel said. For example, battery-operated electric cars could benefit from that form of estimation.

A car could be driven from San Francisco, which has a mild climate, through Sacramento, which can be quite hot, to Lake Tahoe, which can be cold. Batteries lose their power more quickly at higher temperatures and when the vehicle is climbing a hill. As a result, a driver along this route might not have a good estimate of how much longer the battery might last.

A PF model, which could be calculated by an onboard computer on an ongoing basis, could provide a more accurate prediction of how much driving time the battery could offer, Goebel said.

The researchers are now working on ways to narrow the uncertainty bounds of PF-based battery life estimations to increase their usefulness. They are also running additional tests to check the formulas against a wider set of conditions.

About the Author

Joab Jackson is the senior technology editor for Government Computer News.

inside gcn

  • power grid (elxeneize/Shutterstock.com)

    Electric grid protection through low-cost sensors, machine learning

Reader Comments

Please post your comments here. Comments are moderated, so they may not appear immediately after submitting. We will not post comments that we consider abusive or off-topic.

Please type the letters/numbers you see above

More from 1105 Public Sector Media Group