Towards reliable energy storage: a review of machine learning-driven SOH estimation for electrochemical energy storage systems

State-of-health (SOH) estimation is a prerequisite for the safe and efficient operation of electrochemical energy storage systems (ESS), but practical deployment remains limited by incomplete observability, inconsistent health labels, domain shift, and insufficient trustworthiness under changing operating conditions. This review synthesizes machine-learning-based SOH estimation across electrochemical ESS, covering lithium-ion, sodium-ion, and aqueous batteries, supercapacitors, and battery-supercapacitor hybrid configurations. The literature is organized along an end-to-end machine-learning-driven pipeline, including health target definition and labeling, field data and observability, health-indicator representation, model learning and generalization, benchmarking and trustworthiness, and deployment-oriented integration. The review shows that reliable SOH estimation does not depend solely on model architecture, but on the alignment among target semantics, signal availability, feature representation, supervision regime, evaluation design, and operational objective. It further highlights that health semantics are device-dependent and should not be transferred mechanically across batteries and supercapacitors. Representative learning strategies are discussed, including structured regression, temporal deep models, transfer and online adaptation, and physics-guided or uncertainty-aware learning. The review also extends the discussion from component-level diagnosis to system-level integration in batteries, supercapacitors, and hybrid ESS. Future progress will depend on observability-aware target design, robust evaluation, and trustworthy deployment rather than algorithmic novelty alone.