Cross-domain transferable few-shot learning for state-of-health estimation of lithium-ion batteries with physics-consistent neural network

Accurate estimation of the state-of-health (SOH) of lithium-ion batteries is crucial for ensuring safety and longevity in electric vehicles. Given the varied operating conditions, data-driven methods have emerged as the predominant approach for SOH estimation due to their model-free characteristics. Nevertheless, challenges such as the lack of physical insights, limited data samples, and cross-domain discrepancies compromise the accuracy, reliability, and robustness of these methods. In this context, this paper proposes a cross-domain transferable few-shot learning physics-informed neural network for battery health evaluation. Specifically, the dynamics of battery degradation is embedded in the deep neural network. Then, a gradient-enhanced method is introduced to improve the physical consistency of the model, and thus enhance its transferability for cross-domain applications, which can effectively overcome the convergence stagnation problem in the late training stage due to the diminishing residual gradients. Finally, the pre-trained residual network is frozen to retain the cross-domain universal degradation equations, and fine-tuning with few-shot learning is performed to facilitate transfer learning from the source to the target domain. Experiments involving various battery types cycled under diverse loading profiles, along with feature importance analysis, are conducted to verify the effectiveness of the proposed method. The results demonstrate that after single-battery fine-tuning, the root mean square errors decrease to 0.0683 and 0.0157, representing relative reductions of 72.9% and 39.8%, respectively, compared to independent training in the target domains.