Multi-stage adaptive evaluation and diagnosis strategy for backup battery lifecycles in distribution networks oriented towards operation and maintenance decision-making

To address the limited dynamic adaptability of single-stage state-of-health (SOH) models for backup batteries in distribution-network automation terminals, this study proposes a lifecycle-aware diagnostic framework that combines strictly causal stage detection, stage-specific SOH tracking, capacity–power dual-dimensional diagnosis, probabilistic GPR output, and hierarchical safety decision logic. For online tracking, the first 20 observed target-cell cycles establish a frozen robust log-curvature baseline; the prediction at cycle k is generated before the observation at cycle k is ingested. Stage II is activated only after consecutive threshold exceedances, and the model-to-model continuity offset does not use the measured target capacity. In the B0005-to-B0007 case, the proposed method reduces RMSE from 0.0307 for SVR and 0.0294 for single-stage GPR to 0.0207. Across six directed source–target pairs, the mean RMSE is 0.0740, compared with 0.0795 for the strongest single-model baseline, Random Forest, corresponding to a 6.92% reduction. The proposed method also achieves the lowest mean MAE and MAPE and the highest mean R 2 , with lower RMSE in five of the six pairs. Causal Stage II activation occurs 2–11 cycles after the retrospective curvature reference. The temperature-rise-based \(SO{H}P\) proxy crosses 0.70 at cycle 68 in the B0007 Exp2-derived case while \(SO{H}C\) remains 0.8900. The Cell25 experiment is reported as a case-level demonstration: the capacity-retention threshold is first crossed at cycle 7, while Isolation Forest is retained only as an auxiliary warning signal. These results support the feasibility of lifecycle-aware diagnosis at the dataset level; field calibration and validation remain necessary before engineering deployment.