Deep learning model for remaining useful life prediction with reduced labeling data dependency

Accurate prediction of Remaining Useful Life (RUL) is critical for battery health assessment, yet conventional data-driven approaches exhibit limited generalizability across diverse operating conditions and chemistries owing to insufficient labeled training data. To address this challenge, this paper proposes battery masked autoencoders (BMAE), a pre-training and fine-tuning framework that leverages large-scale unlabeled datasets to extract transferable electrochemical features while enabling precise RUL predictions with minimal supervision. This methodology is rigorously validated across five public datasets encompassing 292 battery cells and 201,090 charge cycles, spanning four dominant chemistries (NCA, NCM, NCANCM, LFP), operating temperatures (25–55 °C), charge rates (0.25C–8C), and capacities (1.1–48 Ah). Experimental results demonstrate that BMAE achieves state-of-the-art performance with a root mean square error of 28.74 cycles (3.3 % relative error), requiring merely 20 % of the labeled data demanded by conventional supervised models. This framework significantly enhances prediction efficiency for industrial applications, offering transformative potential for battery lifecycle management in manufacturing, operation, and recycling sectors. Furthermore, the pre-trained model (110 million parameters) has been released as an open resource to facilitate the accelerated development of battery management algorithms beyond RUL estimation.