Equivalent-circuit-model-guided feature selection for interpretable battery state-of-health estimation

Reliable state-of-health (SOH) estimation depends not only on the nonlinear mapper but also on whether its inputs are compact, informative, and physically interpretable. We develop an equivalent circuit model (ECM)-guided feature-selection strategy that uses fractional-ECM parameters offline as physical anchors for charging-response and incremental capacity analysis (ICA) features. The predictive evaluation comprises 4259 cycle samples from ten commercial nickel–cobalt–aluminum (NCA) cells tested at 25 °C, 35 °C, and 45 °C under a 0.5 C protocol. A 13-dimensional (13D) full input is compared with five equal-cardinality 8-dimensional (8D) schemes using gated recurrent unit (GRU) and temporal convolutional network (TCN) nonlinear mappers. Ten-fold leave-one-battery-out (LOBO) validation provides the primary evidence. Among the 8D schemes, the Pearson correlation coefficient (PCC-8D) gives the lowest GRU macro root-mean-square error (RMSE; 0.005272 ± 0.002064 ). Physically anchored ECM-8D gives the lowest TCN macro RMSE ( 0.005882 ± 0.003804 ). ECM-8D is therefore competitive rather than universally most accurate. The two retained charging-time ratios are deterministically related, and 7D removal sensitivities reveal battery- and architecture-dependent effects. ECM fitting is required only during offline feature design; inference uses charging, ICA, and temperature variables. These results support a bounded route to interpretable SOH input construction within the tested NCA protocol, not cross-chemistry or wider-regime generalization.