Learning Battery Health from Degradation Indicators: A Critical Review of Data-Driven SOH Estimation

Abstract

The accurate estimation of the state-of-health (SOH) of lithium-ion batteries is essential for reliable battery management, safety assessment and remaining service-life planning. Conventional SOH estimation methods often rely on electrochemical models, equivalent-circuit representations or controlled diagnostic measurements, which may require detailed battery knowledge, parameter identification or operating conditions that are difficult to maintain in practical applications. Data-driven approaches provide an alternative by learning the relationship between measurable battery signals and degradation behaviour from experimental or operational data. The reviewed approaches are analysed across conventional machine-learning methods, deep-learning architectures, hybrid and ensemble models, transfer and domain-adaptation strategies, and uncertainty-aware approaches. Particular attention is given to the role of input data, feature representation, model complexity, interpretability, computational requirements and generalisation capability. Feature-based machine-learning methods remain relevant where informative degradation indicators and limited computational resources are available, whereas deep-learning approaches provide greater capacity for automatic representation learning but require sufficiently diverse and representative datasets. Hybrid and knowledge-guided approaches offer potential advantages by combining complementary information sources, although their additional complexity must be justified through improved reliability or transferability. The review further identifies data quality, uncertain SOH labels, operating-condition variability, domain shift, feature instability, inadequate validation and deployment constraints as major barriers to practical implementation. A central finding is that reported prediction accuracy alone is insufficient for evaluating data-driven SOH estimators, as model performance is strongly influenced by dataset composition and validation design.