State-of-health estimation of lithium-ion batteries from random charging segments using a discrete-time physics-informed neural network

To improve the accuracy and robustness of state-of-health (SOH) estimation for lithium-ion batteries under incomplete and discontinuous charging data, this paper proposes a discrete-time physics-informed neural network (DT-PINN) tailored to random charging segments. Within random state-of-charge (SOC) windows during constant-current charging, three-scale voltage features, including macroscopic positional, morphological-energy, and local differential-geometry descriptors, are constructed to improve inter-cycle comparability and aging sensitivity. In parallel, temperature, current rate, and time are embedded into a discrete-time degradation driver coefficient, which, together with a capacity regeneration gating factor, defines a physically constrained residual that regularizes the SOH trajectory of fragmented segments. The neural network and physical-constraint modules are jointly trained, and the framework is systematically evaluated on three datasets under full-sample, sparse-sample, and cross-dataset settings. Under the 100% training sample ratio, DT-PINN achieves MAE values of 0.1600%, 0.1379%, and 0.4079% on the three datasets, with a maximum R 2 of 0.9984. With 20% training samples, the MAE remains within approximately 0.4%–1.2% while preserving accurate estimation of early- and mid-stage SOH. Cross-dataset experiments show that fine-tuning with a single labeled target-domain cell reduces the cross-domain MAE from 2.8%–13.2% to approximately 0.2%–0.5%, suggesting that the discrete-time physical constraints can serve as a useful regularization component under the adopted lightweight target-domain adaptation strategy. Compared with the strongest data-driven method under the 20% training sample ratio, DT-PINN reduces MAE by 19.9%. Overall, the proposed method enables accurate SOH estimation without requiring complete life-cycle trajectories and demonstrates methodological effectiveness for incomplete charging fragments on public laboratory datasets.