A two-stage probabilistic framework with validity screening and physics-informed trend modeling for lithium-ion battery health estimation from short discharge transients

초록

Reliable lithium-ion (Li-ion) battery state-of-health (SOH) estimation remains challenging under cross-cell variation and operating-condition shifts. This paper proposes a deployment-aligned two-stage probabilistic framework that infers SOH from discharge-only measurements using short initial transient windows that are realistically available online. An identifiability-aware transient parameterization based on a reduced-order equivalent-circuit form (modified 1RC with a linear trend) is used to construct compact physics-interpretable features from such windows. In Stage 1, lightweight definition-based screening rejects settling and non-nominal cycles and defines the operational validity range for downstream prediction. In Stage 2, only valid cycles are forwarded to a physics-informed trend–residual Gaussian Process Regression (PI-GPR) model that combines a physically motivated trend component with a stochastic residual process to provide decision-facing uncertainty. The main distinction of the proposed framework lies in the explicit separation between validity screening and downstream probabilistic prediction, together with a short-transient physics-interpretable representation and a trend–residual structure tailored to small-data cross-cell deployment. To further improve conservativeness under distribution shifts, an out-of-distribution (OOD)-aware uncertainty inflation mechanism is incorporated for risk-aware interval estimates. Leakage-safe cross-cell experiments show that the proposed framework achieves stable SOH prediction performance and improved uncertainty reliability under the present cross-cell setting, while explicitly reporting the operational availability induced by validity-gated prediction. A runtime assessment of the present MATLAB implementation showed an average online latency of approximately 1.53ms per cycle, while Stage 2 PI-GPR inference itself required only about 0.01ms per query. These results suggest that the proposed framework is compatible with cycle-wise online deployment under the present implementation setting. © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

키워드