Lithium-ion batteries have become a key energy storage solution due to their long lifespan, high energy density, and low self-discharge rate. Long-term use leads to inevitable performance degradation and possible structural damage. Accurate state of health (SOH) estimation is thus essential to ensure their safety and reliability in real-world applications.

In a study published in IEEE/ASME journal, Dr. LIN Mingqiang's group from the Fujian Institute of Research on the Structure of Matter of the Chinese Academy of Sciences proposed an ensemble convolutional neural network based transfer learning framework for achievable lithium-ion battery health perception in flexible charging procedure.

In this SOH estimation method, random charging process was divided into multiple local voltage segments from which the data of voltage and capacity were extracted as core indicators of battery health. Besides, a transfer learning approach based on domain adaptive neural network (DaNN) was employed to align features between the source and target domains, which significantly reduces the model’s dependence on target battery data and lowers the amount of training data required.

To estimate the SOH under arbitrary local voltage segments, an adaptive relevance vector machine ensemble model optimized with a blending strategy was developed. This model utilized multiple DaNN-based sub-models, each trained on a different voltage segment, as base learners. By leveraging their estimation capabilities on local segments, the ensemble model enhanced its adaptability to diverse charging conditions.

To further improve feature weight allocation during the model fusion stage and suppress the risk of overfitting, the Pearson correlation coefficient was introduced to analyze the relationship between features from each voltage segment and SOH. This served as a key criterion for feature fusion, enhancing the stability and generalization performance of the ensemble model.

The proposed method was validated in battery aging experiments conducted under various operating conditions. The results demonstrated that even when trained solely on data from a single source battery, the method maintained high estimation accuracy across different target conditions, showcasing strong cross-condition transferability and robustness.

This study proposes an SOH estimation method which enables accurate battery health assessment under random charging conditions and improving model generalization. The method demonstrates strong cross-condition transferability even with data from a single source domain.