Generalizing capacity estimation for cross-domain lithium-ion batteries with deep multi-domain adaptation

Abstract

Accurately estimating the available capacity to correctly reflect the state of health (SOH) of lithium-ion batteries is crucial for ensuring their safe and efficiency. However, capacity estimation models confront major challenges in generalizing across lithium-ion batteries with different chemistry, types, and varying operating conditions because of the distribution shifts. To address the issue, this paper proposes a deep multi-domain adaptation (DMDA) method to tackle distribution shift problems and to enhance the generalization of capacity estimation for various batteries under dynamic operating conditions. First, cycling experiments with prismatic cells are conducted and compared with public datasets to form a mathematical problem statement. Second, marginal and conditional distribution shifts between multi-domains are aligned by the proposed hybrid loss function and the novel optional kernel strategy during a semi-supervised training process. Meanwhile, we present a training strategy by embedding Bayesian Optimization. Finally, the superiority of the proposed method is verified by comparing with the state-of-the-art transfer learning methods on two case studies. Dimensionality reduction and visualization analysis are further conducted to enhance the interpretability of the results. This work contributes to broaden the existing capacity estimation model to encompass a wide range of lithium-ion batteries.