Simultaneous Feature Extraction of Intra- and Inter-Cycle Changes for Knee Point Prediction of Lithium Ion Batteries

Early lifespan prediction is essential in battery prognostics, and data-driven approaches are prevalent. Previous data-driven approaches did not exploit the cyclic non-stationary patterns in the dataset due to the recurring charging and discharging operation. This study proposes a novel method to extract features from intra- and inter-cycle patterns simultaneously by applying a 2D convolutional neural network (CNN) to a data matrix structured in cycle and voltage dimensions. Results using a published partial discharging dataset show that the 2D CNN not only outperformed the 1D CNN and Elastic Net but also enabled a reduction of the input size for an early prediction.