Cross-domain state-of-health estimation for lithium-ion batteries via meta-transfer learning

The inherent cell-to-cell variations of lithium-ion batteries, coupled with the complexity of real-world operating conditions, pose substantial challenges for cross-domain state-of-health (SOH) estimation. Existing cross-domain approaches have improved adaptability to some extent, but they fall short of achieving adaptability under real-world complex conditions. Motivated by these issues, this paper proposes a unified Meta-Transfer Learning Framework (MTLF). The framework embeds a battery-oriented Model-Agnostic Meta-Learning (MAML) strategy in both source-domain meta-learning and target-domain meta-adaptation. In the source domain, the model learns generalized initialization parameters from multiple battery tasks. In the target domain, only the first 20% of the cycling data are used for bi-level adaptation, while the remaining 80% are reserved exclusively for testing. Each battery is treated as an independent task, and its aging data are partitioned chronologically into support and query sets to preserve the temporal evolution of battery degradation. The source and target domains are strictly separated at the cell level, and the designated cross-domain factor or factors are non-overlapping in each scenario. In addition, to further enhance the representation of health features, we design a one-dimensional convolutional neural network (1-D CNN) and a Transformer Encoder in sequence to jointly capture local features and global dependencies from relaxation voltage (RV). This study systematically investigates seven cross-domain scenarios, covering all combinations of three factors, namely material system, ambient temperature, and charge/discharge rate. On the Tongji and the private dataset, MTLF achieves average root mean squared error (RMSE) of 1.34%, 1.43%, and 1.68%, and average mean absolute error (MAE) of 1.08%, 1.16%, and 1.38% across the seven scenarios, respectively. It outperforms three state-of-the-art cross-domain methods. These results demonstrate superior accuracy and strong adaptability under challenging cross-domain scenarios.