Battery lifetime prediction by pattern recognition application to lead-acid battery life-cycling test data

A novel approach to battery lifetime prediction has been evaluated by application to life-cycling data collected for 108 ESB EV-106. golf cart batteries (tests conducted by TRW for NASA-Lewis). This approach utilized computerized pattern recognition methods to examine initial cycling measurements and classify each battery into one of two classes: long-lived or short-lived. The classifier program was based on either a linear discriminant or nearest neighbor analysis of a training set consisting of: each member of the EV battery set which had failed; the relative lifetime of each member--normalized with respect to test conditions; and a set of features based on measurements of initial behavior. The raw data set included capacity trends over the first 8 or 9 cycles and records of specific gravity and water-added for each cell after initial cycling. Features defined from these raw data included the individual data items as well as transformations and combinations of these data. All features were represented as standardized variables. It was shown that lifetime prediction of batteries within the two categories defined could be made with about 87% accuracy. It is concluded that for a similarly-manufactured battery set, relative lifetime prediction could be based on initial measurements of the same type examined here.