An Interpretable Semi-Supervised Learning Approach for Battery Lifespan Early Prognostic

Early estimation of battery lifespan holds significant importance in efficient battery health management. This paper presents an Interpretable Semi-supervised learning approach for estimating battery lifespans during their initial degradation stage. Specifically, a Physics-informed Autoencoder is introduced in the pretraining stage to extract features reflecting the early degradation mode of batteries by incorporating an electrochemical model into data-driven method. Moreover, to improve interpretability, an aging mechanism-based attention neural network is proposed to integrate electrode-level degradation mode with cyclic-based features for early battery lifespan prognostics. Additionally, a Bayesian Neural Network is incorporated into the model for uncertainty estimation. The effectiveness of the proposed method is validated using the LiFePO
4
dataset, demonstrating high performance in the early estimation of battery lifespan.