Physics-informed degradation diagnostics of lithium-ion batteries using practical charging segments

The degradation modes (DMs) of lithium-ion batteries provide detailed health information beyond capacity and resistance for reliable battery management. However, most existing methods for diagnosing battery degradation have two major limitations: reliance on large amounts of labeled data for effective machine learning (ML) and limited adaptability to practical applications owing to the inaccessibility of features in practice. To address these challenges, a physics-informed machine learning (PIML) method is developed for battery DM diagnostics on the basis of a physics-based model (PBM). The PBM is calibrated to reproduce both voltage responses and DMs of batteries with high fidelity and is subsequently validated by capturing the consistent evolution of incremental capacity (IC) curves during battery aging. This approach involves generating large-scale synthetic data that cover various aging states of batteries by adjusting the aging-related parameters of the PBM. The synthetic data are then refined by screening out invalid voltage response outputs via the PBM, with the mean Euclidean distance as the screening criterion. The features that are used to indicate battery aging are extracted from practical voltage responses (generally within the 21.26 % to 98.11 % range of the state of charge, SOC) under 0.3 C-rate charging. The relationships between these features and battery DMs are then systematically analyzed and modeled using Gaussian process regression (GPR). The experimental results for silicon/graphite batteries show that GPR, which is trained only on synthetic data, can accurately diagnose battery degradation throughout a battery’s lifespan. Among the 16 batteries that are tested via various discharge strategies and at various temperatures, the root mean square error of DM diagnosis is <2 % in most scenarios.