Explainable Neural Network for Sensitivity Analysis of Lithium-Ion Battery Smart Production

Abstract

Battery production is crucial for determining the quality of electrode, which in turn affects the manufactured battery performance. As battery production is complicated with strongly coupled intermediate and control parameters, an efficient solution that can perform a reliable sensitivity analysis of the production terms of interest and forecast key battery properties in the early production phase is urgently required. This paper performs detailed sensitivity analysis of key production terms on determining the properties of manufactured battery electrode via advanced data-driven modelling. To be specific, an explainable neural network named generalized additive model with structured interaction (GAM-SI) is designed to predict two key battery properties, including electrode mass loading and porosity, while the effects of four early production terms on manufactured batteries are explained and analysed. The experimental results reveal that the proposed method is able to accurately predict battery electrode properties in the mixing and coating stages. In addition, the importance ratio ranking, global interpretation and local interpretation of both the main effects and pairwise interactions can be effectively visualized by the designed neural network. Due to the merits of interpretability, the proposed GAM-SI can help engineers gain important insights for understanding complicated production behavior, further benefitting smart battery production.