A Contrastive Learning Battery State of Health Estimation Method Based on Self-supervised Aging Representation
Jiaqi Li, Jingzhe Zhu, Ziying Huang, Guodong Fan, Xi Zhang
Abstract
Accurate state of health (SOH) estimation is crucial to improve the performance and reduce the maintenance cost of electric vehicles. Existing data-driven SOH estimation methods mainly focus on using sufficient capacity labels to train deep learning models, which tend to establish the mapping between health indicators and residual capacity. However, the number of available supervised labels in the actual driving data is usually limited, resulting in substantial deterioration of model performance. To address this label-insufficient issue, this article proposes a novel self-supervised SOH estimation method based on contrastive learning. First, a two-stage relation network composed of a feature encoder and a relation module is proposed to model the similarity between different battery samples in a contrastive learning paradigm. Meanwhile, a representation space is obtained where embeddings from different samples aggregate with battery aging. Then, the difference of cycle number in a single battery sample is taken as the self-supervised aging pseudo label. Rather than establishing the conventional mapping between SOH and its indicators, the proposed framework directly learns the similarity between batteries in different aging states, and thus manages to eliminate the dependence on large number of supervised labels. Finally, experimental results show that compared with end-to-end mapping, the root mean square error of the proposed method is reduced by 33% and 36% on average under insufficient and sufficient label conditions, respectively.