Multisource Weighted Domain Adaptation for Cross-Conditions Capacity Estimation of Lithium-Ion Batteries

Abstract

Accurate online lithium-ion battery capacity estimation is essential for ensuring the safety of battery management systems (BMS). Lithium-ion batteries exhibit varying dynamic degradation processes across different operating environments and load conditions (i.e., different domains). In real engineering scenarios, unsupervised battery capacity estimation under multiple operating conditions becomes a challenging problem due to the absence of labeled target data and existing data distribution discrepancies. A novel transfer learning model, Multisource Weighted Domain Adaptation (MWDA), is proposed for unsupervised cross-conditions capacity estimation. First, a common feature extractor, called the CNN-Transformer, is constructed to learn domain-invariant degradation-sensitive features from dynamic battery monitoring data under the guidance of an adversarial mechanism. Second, a multi-order statistical metric is introduced to make the unsupervised feature alignment process more comprehensive and effective. Third, a multisource dynamic weighting method is proposed. By adaptively adjusting the weight of each source domain during the training process, the model can fully utilize the supervision information from related domains and minimize the risk of negative transfer. Compared with other multisource domain adaptation methods, MWDA reduces the average mean square error (MSE) and mean absolute error (MAE) by 61% and 43% on the MIT dataset, and by 32% and 28% on the NASA dataset. Extensive transfer experiments validate the superiority and effectiveness of the proposed method.

Topics & Concepts