State of health estimation of lithium-ion batteries using Autoencoders and Ensemble Learning

Abstract

This paper focuses on the state of health (SOH) estimation of lithium-ion batteries, which is critical for the reliable operation of electric vehicles. First, a convolutional autoencoder (AE) and a recurrent AE are designed to automatically extract health features (HFs), which are the low-dimensional mappings of the charging profiles at different aging stages. This feature extraction method with two AEs can avoid the artificial definition of HFs and the additional operation on the charging profiles. On this basis, an ensemble learning (EL) method is proposed to improve the SOH estimation accuracy, which consists of a series of sequentially trained gate recurrent unit (GRU) networks. This pattern of sequential training makes that the current GRU network can focus on the poor-performing samples of the previous trained network. Finally, four battery datasets under different cycling test conditions are used to verify the efficiency of the proposed SOH estimation method with the two AEs and the EL. The experimental results reveal that the proposed method can provide accurate battery SOH estimation, with the root mean square error and mean absolute error of Leave-One-Out cross validation (LOO) are 1.04% and 0.77%.