Interpretable Approaches for Optimizing the Pulse Diagnostics and Formation for Lithium-ion Batteries

Abstract

Lithium-ion Batteries (LiBs) are widely used in electronic devices and energy storage systems owing to their high energy density, long lifespan, and low cost. To further improve their performance and safety, we focus on optimizing two necessary yet time-consuming LiB applications: pulse diagnostics that enable degradation mechanism-level diagnosis, and the formation step that can greatly impact the battery performance. Interpretability plays a key role in the optimization scheme. Interpretability allows us to break down result-oriented, time-consuming objective functions into tractable problems, enhances the extrapolation capability of the optimization solver, and determines the depth of physical insights we gain from the optimization scheme. Thus, we aim for interpretable approaches for optimizing the pulse diagnostics and formation step. First, we optimize voltage-pulse diagnostics using the fitness model, a mechanistic model that describes the behavior of degraded LiBs under voltage pulses, as a source of interpretability. The optimal set of diagnostic protocols is found from the Pareto front plot with two objective functions, each representing the practical identifiability of the degradation parameters and the total diagnostic time. We discuss on why film resistance is challenging to identify, and how to improve the practical identifiability using redundant voltage pulses. Next, we demonstrate how the formation protocols can be optimized using the systematic feature engineering framework as the source of interpretability. Using two features designed from our framework, we achieve reliable evaluation of new formation protocols during the formation step, with a < 10% cycle life prediction error. Combined with the domain knowledge, our feature engineering work guides the development of the mechanistic distributed-resistance model, providing physical meaning to the data-driven features. Finally, we present a reliable classification method of full-cell Acoustic Emission (AE) signals based on their acoustic sources, which is the first step for utilizing AEs for optimizing the formation step. AEs capture microscopic mechanical phenomena, such as gas generation or particle fracture. Using the simplest convolutional neural network structure for interpretability, we observe a possible transferability of the classifier from half-cell to full-cell AEs, and across similar cell chemistries.