Health-aware federated contrastive representation learning for generalizable state-of-health estimation of lithium-ion batteries

Accurate state-of-health (SOH) estimation of lithium-ion batteries is critical for safe operation and energy management in electric vehicles. However, battery data are naturally distributed across different owners, heterogeneous operating conditions and degradation patterns limit model generalization, and privacy constraints prevent centralized training. To address these challenges, this paper proposes a health-aware federated contrastive representation learning framework (HA-FCRL) for directly generalizable SOH estimation without centralizing raw charging data. The framework computes a local degradation rate for each charging cycle to dynamically adjust pairwise similarity thresholds in contrastive learning, tightening thresholds during rapid decay phases to capture subtle health differences and relaxing them during plateau periods to tolerate measurement noise. The method constructs a representation learning objective that integrates temporal and instance contrastive regularization, SOH-guided contrastive learning, and adaptive SOH ordinal ranking constraints, enabling the feature space to explicitly encode battery health similarity and ordering relationships. At the server side, a reliability-aware composite aggregation strategy simultaneously considers sample quantity and relative local training loss to accommodate heterogeneous clients. On the MIT battery dataset, HA-FCRL achieves a root mean square error (RMSE) of 0.0104 on unseen batteries, representing a 20.2% reduction compared to the best baseline and a 25.1% reduction compared to standard federated learning (FedAvg). On real-world vehicle battery data, HA-FCRL achieves an average RMSE of 0.0169, an 11.2% improvement over FedAvg, demonstrating practical value in heterogeneous real-world scenarios.