A data-driven methodology for early-stage estimation and second-life applicability assessment toward lifecycle refined management of power batteries

Escalating environmental costs and performance demands are pushing battery management toward refined, lifecycle approaches. A critical advancement involves extending state of health (SOH) assessment from automotive to second-life applications while reducing dependence on training data. This requires addressing temporal nonlinearity in degradation and mitigating data fragmentation issues. This study investigates feature effectiveness evolution across the battery lifecycle using a composite aging dataset with fast-charging and low-temperature conditions. By integrating Transformer and Bi-directional Long Short-Term Memory (BiLSTM) architectures with transfer learning, we identify optimal feature-model combinations and extend the SOH assessment window to early stages and second-life applications. Experiments show that in second-life, SOH estimation RMSE remains below 2 %. Moreover, transfer learning enables the valid assessment range to 30 % SOH using only 15 % target data, offering a critical approach to reduce carbon costs and enhance lifecycle management.