Stryker: predicting battery life without running the battery to failure

Battery life is defined by the number of charge and discharge cycles a cell will complete. Establishing that number by testing means cycling the battery until it degrades, which takes as long as the battery lasts and consumes the battery in the process. At the volumes a manufacturer needs to test, that is prohibitive in both time and cost.

ThirdEye Data built a machine learning model that predicts remaining life from early cycle test data, so the answer arrives from the beginning of a test rather than the end of one. The work involved substantial feature processing, both selecting which signals carry information and constructing new ones from them, followed by training and evaluating several different kinds of regression model rather than committing to one approach at the outset.

Two further elements made the difference between a model and a working method. A reinforcement learning component guides the testing process itself, so the tests being run are the ones that will most improve the prediction. And an active learning approach decides which batteries are worth testing further, concentrating effort where uncertainty is highest rather than spreading it evenly.

That second part is the one that changes the economics. Predicting from existing data is useful. Deciding what to test next is what reduces how much testing has to happen at all.

Remaining battery life is predicted from early cycle data rather than established by running each battery to the end of its life.