Multisource Feature Separation and Weighted Network for Cross-Conditional Capacity Estimation of Lithium Batteries
提出多源特征分离与加权网络，通过分离公私特征、对抗学习域不变特征及动态加权各源域，解决多源域适应中的负迁移问题，在MIT和XJTU数据集上均显著降低均方误差和平均绝对误差。
Accurate prediction of lithium-ion battery capacity is a critical task in BMSs. However, existing multisource domain adaptation methods often ignore the different contributions of each source domain, focusing solely on aligning the global distributions of source and target domains. This limitation can result in negative transfer. To address this issue, this article proposes a multisource feature separation and weighted (MFSW) network for lithium-ion battery capacity estimation. First, private and common features of both source and target domains are disentangled through feature separation. An adversarial mechanism is employed to guide the common feature extractor to learn domain-invariant features. Then, the features are further aligned using a multiorder metric. Finally, a multisource dynamic weighting method is introduced to adaptively adjust the weight of each source domain. Compared with other multisource domain adaptation methods, the proposed method reduces the average MSE and MAE by 56.3% and 28.8% on the MIT dataset, and by 44.0% and 38.6% on the XJTU dataset, respectively. Extensive experimental results demonstrate that the proposed method effectively mitigates negative transfer and exhibits superior performance and robustness.