State of Health Estimation for Lithium-ion batteries Based on Extreme Learning Machine with Improved Blinex Loss

Abstract

The state of health (SOH) estimation for Lithium-ion batteries (LIBs) plays an important role in battery management system (BMS), and the data-driven based SOH estimate methods mainly depend on the measurement data which is usually corrupted by non-Gaussian noise or other random disturbances because BMS usually works under complex environmental conditions. To achieve more accurate and robust SOH estimation, a novel robust extreme learning machine (ELM) based SOH estimation method is proposed. An improved Blinex Loss (IB-Loss) function is defined to replace the mean square error (MSE) loss in traditional ELM, and a novel robust estimation method called ELM with IB-Loss (IB-ELM) is derived, which can reduce the effect of noise. Through the comprehensive analysis of the aging experimental data of LIBs, we extract health features (HFs) from the charging data as the input of the estimation model, and the gray relational analysis (GRA) is utilized to evaluate the correlation between HFs and SOH to determine the rationality of selected HFs. Finally, the battery datasets provided by NASA are used as the training set and testing set to verify the effectiveness of the proposed method, and experimental results show that it has higher estimation accuracy than other existing data-driven methods under non-Gaussian noise conditions.