Generalized foundation model for lithium-ion battery state-of-health prediction with distribution metric learning

With the rapid proliferation of electric vehicles and lithium-ion batteries, accurate estimation of the State of Health of batteries has become increasingly critical. However, variations in battery types and operating conditions often lead to distribution discrepancies among datasets, and the scarcity of labeled data within a single dataset makes SOH prediction particularly challenging. To address these issues, this paper proposes a generalized foundation model for few-sample battery SOH estimation, trained with data from multiple source domains. The proposed method employs Maximum Mean Discrepancy to minimize the feature distribution gap among different domains, and integrates a ResNet-Transformer hybrid architecture with the Kolmogorov–Arnold Network to enhance nonlinear feature representation. By leveraging voltage, current, and time data from multiple domains, the model effectively learns generalized knowledge of battery degradation and transfers it to target domains for robust SOH prediction. Experimental results demonstrate that the proposed generalized foundation model achieves outstanding generalization performance and maintains high prediction accuracy even with limited labeled data. This indicates its strong potential for practical applications in battery health management and lifespan prediction of electric vehicles.