Interpretable event-window transformer for lithium-ion battery state-of-health estimation under randomized usage

Lithium-ion batteries support a wide range of transportation, energy-storage, and industrial applications, where reliable knowledge of degradation is required for safe operation and effective asset management. State-of-health (SOH) assessment is nevertheless constrained by the intermittent availability of diagnostic capacity measurements during routine operation. This study presents an event-window Transformer for SOH estimation using the National Aeronautics and Space Administration (NASA) Randomized Battery Usage dataset. Random-walk charge and discharge events are represented by interpretable descriptors of voltage response, current loading, duration, charge throughput, energy, power, and local charge–voltage behavior. Chronological windows of 32 consecutive events provide the model input, while reference-discharge capacities supply diagnostic SOH targets through an explicitly defined next-reference labeling strategy. Previous SOH values, previous capacity measurements, future observations, and reference identifiers are excluded from the model inputs. In the primary battery-level transfer experiment, the Transformer was trained on RW9 and RW10 and evaluated on the held-out RW11 cell, achieving an event-window R² of 0.9551, a mean absolute error of 2.387, and a root mean square error of 3.039 SOH percentage points. Aggregating the same predictions at diagnostic checkpoints produced an R² of 0.9664 and a root mean square error of 2.535 percentage points. Leave-one-cell-out evaluation across RW9, RW10, and RW11 yielded a mean event-window R² of 0.9637 ± 0.0075 and a mean diagnostic-checkpoint R² of 0.9772 ± 0.0094. Gradient SHAP and Local Interpretable Model-agnostic Explanations identified voltage-window characteristics, local charge–voltage behavior, event duration, and electrical loading as influential predictive evidence. These results demonstrate the ability of chronological event representations to support interpretable SOH estimation under randomized operating conditions.