Multi-Stage State-of-Health Estimation for Lithium-Ion Batteries Based on an Adaptive SHAP-Guided Dual-Attention LSTM

Accurate state-of-health (SOH) estimation across nonlinear battery aging is challenged by condition-dependent feature relevance. This study combines aging data up to a defined 80% SOH operational end-of-life threshold with an adaptive SHapley Additive exPlanations (SHAP)-guided dual-attention long short-term memory (LSTM) network. Current, voltage, state of charge (SOC), and a pulse-derived short-time resistance indicator from hybrid pulse power characterization (HPPC) are used as inputs. Training-set SHAP importance supplies a weak global feature prior, whose influence is adjusted for each input sequence by an adaptive gate. Residual feature attention recalibrates the inputs, while temporal attention aggregates representations from a shared LSTM encoder. Stage windows and labels are permuted synchronously during training to avoid fixed-position shortcuts. Under held-out SOC conditions, the complete model achieves MAE = 0.3418 percentage points (pp), RMSE = 0.5652 pp, MAPE = 0.3825%, and R2 = 0.9945, reducing MAE and RMSE by 14.1% and 12.4% relative to the Plain Shared LSTM. The results support the coordinated architecture within the tested SOC-holdout protocol; cross-cell performance is assessed separately under a different UDDS setting.