Improving the state-of-health estimation of lithium-ion batteries based on limited labeled data

Lithium-ion batteries, the most promising and widely used power source, require accurate age-related failure assessments for safe and efficient operation. As a critical battery age indicator, state of health (SOH) estimation is a pivotal function of battery management systems. This study proposes two machine learning (ML) methods with data augmentation (Method 1 and Method 2) for predicting batteries' SOH. In Method 1, data augmentation is performed using limited labeled data, and an ML model is employed to predict batteries' SOH throughout their life cycle. Method 2 comprises two ML models: the first ML model predicts early-life SOH online, while the second predicts mid-to-late-life SOH online utilizing augmented labeled data. To address the big data requirement problem of ML, a linear relationship between the equivalent circuit model features and battery SOH is found and used to generate much augmented training data from limited labeled data during batteries' early-life. The proposed method is validated using three types of batteries, comprising 118 cells with 45,948 data units. The results indicated an excellent improvement in predictive performance with an increase in limited labeled data. Specific application scenarios for the two methods are discussed. Additionally, if online early-life data are labeled, they can be used for data augmentation for further prediction accuracy improvement when using Method 2.