Online state-of-health prediction of lithium-ion batteries with limited labeled data

State-of-health (SOH) plays a vital role in battery health management and power system stability. This process can be achieved by capacity estimation. However, in practice, the capacity of a battery is difficult to obtain online given that it cannot be determined with general sensors. This means that the capacity is only known for the limited cycles of the batteries. To address this issue, we propose a novel semi-supervised learning framework to estimate the capacity of unlabeled data to achieve better SOH prediction. First, four indirect features are extracted from the charging profiles. Then an improved locally linear reconstruction method is used to determine the capacity distributions of the unlabeled data. Combined with the oversampling method applied to generate a series of data by the estimated distributions, a support vector regression model is utilized to predict the RUL of the batteries given the threshold values of the batteries. A case study with two types of cellular phone lithium-ion batteries is presented to illustrate the effectiveness of the proposed method for the prediction of the remaining useful life of different batteries and different starting points. The experimental results prove that the performance of the proposed method is better than the K-nearest neighbor method and locally linear reconstruction method in terms of accuracy and robustness.