An improved co-training architecture for Lithium-ion batteries state of health estimation with semi-supervised learning

Abstract

State of Health (SOH) estimation plays a critical role in ensuring lithium-ion battery lifetime safety. However, obtaining labeled battery data is time-consuming. Conventional semi-supervised methods reduce the dependence on labeled data but still require abundant labeled battery data for model training. To address this issue, an improved co-training architecture with semi-supervised learning is proposed, achieving promising SOH estimation accuracy using only unlabeled data and 10% labeled data from one battery. The architecture consists of an Extreme Learning Machine (ELM) with traditional features calculated from the charging curves and a Bi-directional Gated Recurrent Unit (Bi-GRU) with deep features extracted from an encoder. The co-training algorithm facilitates the mutual learning of two models to improve overall estimation accuracy. Pseudo-labels are generated for unlabeled data and filtered through a selection mechanism to supplement scarce labeled data. A fine-tuning stage then leverages these pseudo-labels to augment supervised knowledge. Extensive experiments demonstrate the superiority of the proposed architecture. Under the scenario of 10% sparsely labeled training data from one battery, the proposed method achieves Root Mean Square Error (RMSE) improvement around 56%, 52%, 56% over SVR-KNN and 70%, 74%, 46% over Dual-NARX on average in the Oxford, CALCE CX2 and Tongji NCA battery dataset.