A Novel Temporal Distribution Matching-Based Transfer Learning Method for State of Health Estimation of Li-ion Batteries

Accurate evaluation of the state of health (SOH) is crucial for ensuring the safe functioning of lithium-ion batteries and maximizing their potential for usage and optimization. The deep learning-based approach to assessing the SOH of lithium-ion batteries has difficulties in attaining strong generalization across many domains due to the wide range of battery types and usage conditions, as well as the high cost of collecting battery data. To address this problem, this paper proposes a temporal distribution matching-based transfer learning approach to enhance the temporal generalization ability on the basis of parameter transfer and feature transfer. First, multi-dimensional features are retrieved, and the feature expression score (FES) is computed using the gray relational analysis (GRA). The task with the greatest FES is chosen as the base model for transfer learning. During the transfer learning phase, the initial 25% of the target dataset is utilized to refine the base model. Simultaneously, temporal distribution matching is performed to improve temporal feature alignment by taking into account the output of each hidden state in the temporal network. The accuracy and robustness of the proposed model are effectively demonstrated through experiments done on two datasets. Our solution outperforms previous transfer learning algorithms in terms of performance.