Lithium-ion battery RUL prediction based on multi-modal historical data

A comprehensive assessment of satellite lithium-ion batteries is essential for ensuring the operational reliability of in-orbit satellite platforms. The modeling of newly developed satellite lithium-ion battery models often encounters challenges due to limited historical data, which complicates rapid assessment necessary for minimizing battery testing costs and shortening the installation interval before deployment. This paper introduces a novel approach utilizing multi-modal historical data for predicting the remaining useful life (RUL) of new satellite lithium-ion battery models. The proposed method comprises three main stages: Initially, a Transformer network-based model is trained using data from older models to establish a base model that captures the general degradation trajectory. Subsequently, a transfer learning model is developed by fine-tuning the base model with data from the new models and aligning the model parameters with the unique degradation characteristics of the new batteries. Lastly, this transfer learning model is applied to provide accurate RUL predictions for new lithium-ion battery models. Experimental validation on two distinct datasets demonstrates that the proposed method significantly enhances the accuracy of RUL predictions, achieving minimal max percentage error and mean absolute error of 4.63% and 10.7, respectively, with limited historical training data.