Abstract
The rapid adoption of battery-powered vehicles and energy storage systemsover the past decade has made battery health monitoring increasingly critical.Batteries play a central role in the efficiency and safety of these systems,yet they inevitably degrade over time due to repeated charge-discharge cycles.This degradation leads to reduced energy efficiency and potential overheating,posing significant safety concerns. Accurate estimation of a State of Health(SoH) of battery is therefore essential for ensuring operational reliabilityand safety. Several machine learning architectures, such as LSTMs,transformers, and encoder-based models, have been proposed to estimate SoH fromdischarge cycle data. However, these models struggle with the irregularitiesinherent in real-world measurements: discharge readings are often recorded atnon-uniform intervals, and the lengths of discharge cycles vary significantly.To address this, most existing approaches extract features from the sequencesrather than processing them in full, which introduces information loss andcompromises accuracy. To overcome these challenges, we propose a novelarchitecture: Time-Informed Dynamic Sequence Inverted Transformer (TIDSIT).TIDSIT incorporates continuous time embeddings to effectively representirregularly sampled data and utilizes padded sequences with temporal attentionmechanisms to manage variable-length inputs without discarding sequenceinformation. Experimental results on the NASA battery degradation dataset showthat TIDSIT significantly outperforms existing models, achieving over 50%reduction in prediction error and maintaining an SoH prediction error below0.58%. Furthermore, the architecture is generalizable and holds promise forbroader applications in health monitoring tasks involving irregular time-seriesdata.