An Exponential Transformer for Learning Interpretable Temporal Information in Remaining Useful Life Prediction of Lithium-Ion Battery

Abstract

Accurately predicting the remaining useful life (RUL) of lithium-ion batteries (LIBs) is crucial for ensuring the safety and health maintenance of electric vehicles (EVs). Although conventional Transformers have demonstrated superior performance in various scenarios, they encounter significant challenges when applied to RUL estimation. Specifically, these models often lack the ability to decompose and interpret raw temporal data effectively, leading to underutilization of critical information. In addition, the prolonged prediction times associated with these models can increase the risks associated with safe battery operation. To address these issues while preserving the strengths of Transformer-based temporal modeling, this article introduces a novel exponential Transformer designed to learn interpretable temporal features. Given the declining trend in battery capacity data and the presence of fluctuating information, this article first extracts the level component and the fluctuation component from the original data, thereby decomposing the capacity information into interpretable time-series features. Furthermore, we propose an exponential attention mechanism to replace the self-attention mechanism, which enhances the RUL prediction accuracy and efficiency by focusing on similar data features through exponential decay. Extensive experiments conducted on several aging datasets demonstrate that the proposed network achieves effective results in both single-step and multistep prediction tasks. Compared with other well-established methods, our method reduces the mean absolute error (MAE) and root mean square error (RMSE) by an average of 81.61% and 77.76%, respectively. The code will be made publicly available at: https://github.com/wch1121/RUL-Prediction.