Data-Driven Quantification of Battery Degradation Modes via Critical Features from Charging

Abstract

Battery degradation modes influence the aging behavior of Li-ion batteries, leading to accelerated capacity loss and potential safety issues. Quantifying these aging mechanisms poses challenges for both online and offline diagnostics in charging station applications. Data-driven algorithms have emerged as effective tools for addressing state-of-health issues by learning hard-to-model electrochemical properties from data. This paper presents a data-driven method for quantifying battery degradation modes. Ninety-one statistical features are extracted from the incremental capacity curve derived from 1/3C charging data. These features are then screened based on dispersion, contribution, and correlation. Subsequently, machine learning models, including four baseline algorithms and a feedforward neural network, are used to estimate the degradation modes. Experimental validation indicates that the feedforward neural network outperforms the others, achieving a root mean square error of around 10\% across all three degradation modes (i.e., loss of lithium inventory, loss of active material on the positive electrode, and loss of active material on the negative electrode). The findings in this paper demonstrate the potential of machine learning for diagnosing battery degradation modes in charging station scenarios.