State of health estimation of lithium-ion battery with automatic feature extraction and self-attention learning mechanism

Abstract

Accurate state of health (SOH) estimation is significantly important to ensure the safe and reliable operation of lithium-ion battery. Most existing data-driven estimation methods are based on feature engineering and rely heavily on expert experience and manual operation. However, manually extracting qualified health features requires rich prior knowledge, and these highly-designed features for one specific application may not generalize well to other situations. In this work, an automatic feature extraction method combining convolutional autoencoder and self-attention mechanism is proposed for battery SOH estimation. With preprocessed data fed into the convolutional autoencoder, efficient features characterizing battery health are automatically extracted without human intervention. A self-mechanism module is then further employed to map these high-dimensional abstract health features into battery SOH. Finally, experimental study of battery aging is implemented to demonstrate the proposed method, and comparisons of the proposed method with existing data-driven approaches and the manual feature-based methods have also been presented. With the help of the convolutional autoencoder and self-attention module, the proposed method replaces the conventional manual feature engineering with automatic feature extraction, and reaches 0.0048 average test root-mean-squared error (RMSE) and 0.46% mean-absolute-percentage error (MAPE) on our dataset and 3.69% on the NASA public dataset.