Cross-conditions capacity estimation of lithium-ion battery with constrained adversarial domain adaptation

Accurate estimation of lithium-ion battery capacity is important for battery management systems. Traditional deep learning algorithms assume in advance that the training and test data satisfy independent identical distribution (IID). However, this ideal assumption reduces the generalizability of related methods because the battery operating conditions are often diverse. To address this issue, an unsupervised constrained adversarial domain adaptation method based on causal analysis, attention mechanism and Mogrifier-LSTM (CAM-LSTM-DA) is proposed. First, causal analysis is used to select health indicators (HIs) that are intrinsically associated with capacity degradation, ensuring that the constructed model is valid for the target domain. Then, we adopt Mogrifier-LSTM with key-value pair attention mechanism as the primary network, forcing the learned embedding to have rich degradation information. Finally, to avoid the negative transfer brought by traditional domain adaptation methods, we propose a constrained adversarial domain adaptation method that uses a self-supervised learning module with dynamic temperature and a semantic information constraint module to constrain feature alignment in terms of temporal and semantic information, respectively. The extensive cross-conditions experiments validate the generalizability and prediction performance of the proposed method.