NBD-Net: A structured degradation memory and local-global bridging network for lithium-ion battery remaining useful life prediction

Abstract Remaining useful life (RUL) prediction of lithium-ion batteries from measured capacity is a core function of battery management, but public datasets contain few cells, a predictor must generalize to unseen cells, and training cells often miss the late-life regime. This paper proposes NBD-Net, a Nonstationary-Bridged Degradation Network for next-cycle capacity prediction, and examines which elements make it reliable on unseen cells. NBD-Net normalizes each input window (RevIN), encodes it with a local stabilization branch, a causal detrending followed by learned causal convolutions, in parallel with a recurrent trend encoder, and couples the two branches through an attention memory (SDMC) and a bidirectional cross-attention bridge (LGBI). Under a leave-one-battery-out protocol covering all eleven cells of the CALCE, NASA, and TJU datasets, NBD-Net has the lowest capacity RMSE of nine models on every unseen cell, 9%–27% below the strongest baseline of each dataset, also after nested tuning of all models. When the training cells cover only the first half of life, it remains the most accurate model on every cell, and its error grows by 29% where that of PatchTST and TCN grows by 160% and 222%. Controlled experiments attribute these two results to different elements: with complete training data the per-window normalization carries the advantage, lowering the error by 22%–34% on every cell, so that a normalized recurrent predictor reaches comparable accuracy; with incomplete training data the local stabilization branch provides 84% of the gain over the recurrent backbone, whereas SDMC and LGBI add smaller gains at regeneration events and in late life. The results hold for next-cycle prediction under one cycling condition per dataset; its RUL relative error statistically matches that of the baselines, and its accuracy gain holds while input noise stays comparable to the scatter of the training data. The source code is publicly available at https://github.com/luxing-just/NBDNet_open_source.