Battery State of health estimation with fewer labelled data: a semi-supervised approach

Accurate estimation of battery state of health (SOH) is indispensable for reliable battery management. While machine learning methods are playing an increasingly important role, they generally require profuse training samples which consist of input data and measured capacities. To alleviate this issue, we present a semi-supervised approach that can draw on easily available training samples without measured capacities to train deep neural networks (DNNs) with high SOH estimation performance. First, a label propagation strategy is proposed to generate pseudo capacities for unlabelled training samples by resorting to the similarity between input data. Then, a training strategy is designed to efficiently train the DNN using the training samples with measured and pseudo capacities while taking into account the label propagation errors. A large battery degradation dataset is developed for method validation. End-to-end SOH estimation using is carried out based on a typical long short-term memory (LSTM) DNN. The validation results based on electrochemical impedance spectra demonstrate that reducing the number of training samples deteriorates the performance of the supervised DNN. In contrast, the proposed method can achieve higher accuracy than the supervised DNN and another two machine learning models with fewer labelled training samples. Our results provide an efficient and general approach to developing data-driven SOH estimation models with reduced data collection efforts.