Penn State engineers have developed neural network pattern recognition software -- "a tool that learns" -- to predict the performance deterioration and failure of long-life, rechargeable batteries, the type used in laptop computers, cell phones and other portable electronic devices.
University Park, Pa. --- Dr. Mirna Urquidi-Macdonald, associate professor of engineering science and mechanics, says that the new software requires only minutes worth of laboratory test data to predict long-term performance and life of the battery. In addition, the new software can be used to aid manufacturers in producing better batteries, she says.
The software was developed with data from lithium-ion secondary batteries being considered for use in the U.S. space program.
Her results are described in a paper, "Predicting Failure Of Secondary Batteries Before It Happens" in the July 15 issue of the Journal of Power Sources. Her co-author is Penn State alumnus Neil A. Bomberger, one of her former undergraduate students.
The Penn State researchers conducted their study using a set of battery data supplied by the National Aeronautics and Space Administration (NASA).
Urquidi-Macdonald explained that NASA needs especially reliable information on battery failure for uses in survival situations.
"Normally, NASA will take half of a family of batteries manufactured at the same time and test them to failure," she says. Then, a failure average time is obtained and a safety factor is used to calculate the optimal number of battery cells they will send into space, knowing that they will have the amount of energy and power required for the space mission."
The new software provides a faster, cheaper way to obtain failure information. In addition, she says, "We know that the life of a battery is related to the fabrication process. With information provided by the new software, we should be able to build better batteries."
Urquidi-Macdonald says the new software is built on artificial neural nets, which are mathematical tools with excellent capabilities in pattern matching, recognition and classification. Inspired by the biological behavior of brain cells, neural nets can "learn" when given a large body of data on which to "train."
Using the NASA data and the neural net technique, the Penn State researchers identified those variables during the life testing of the batteries that were most associated with predicting the future performance of the batteries. These variables included operational temperature and current/time signatures imposed on the batteries. Then, using those variables, they "trained" the software to predict the output voltage at a given time. Finally, after training on five data sets, the software was tested with data on which it had not been trained and accurately predicted the output voltage and time of battery failure.
Urquidi-Macdonald says, "Using this technique we can predict the response of the battery over two hours when the net was trained on only 10 minutes of actual testing data. Now, we'd like to apply the neural net technique to larger data sets and to different kinds of batteries. We have a data set measured for nickel-cadmium battery cells that includes 10 years of observations."
The research was supported, in part, by a contract from the U.S. Department of Energy.
EDITORS: To contact Dr. Urquidi-Macdonald call (814) 863-4217 or to mumesm@engr.psu.edu by email.
Journal
Journal of Power Sources