Abstract
Accurate lithium-ion battery state-of-health (SoH) estimation is essential for second-life screening, predictive maintenance, and circular-economy routing decisions such as reuse, remanufacture, or recycling. Short pulse-voltage tests provide a fast diagnostic alternative to full capacity tests, but their interpretation remains challenging because pulse responses depend jointly on chemistry, state of charge (SOC), nominal capacity, degradation history, and operating protocol. This challenge becomes especially critical when a model is deployed at SOC conditions that were not observed during training.This paper proposes the \emph{Unified Sparse Diffusion Transformer} (USDT), a chemistry-conditioned sequence architecture for pulse-based battery SoH estimation and second-life decision support. USDT represents the pulse response as a sequence of voltage tokens enriched with SOC, capacity, pulse-width, and chemistry information. The architecture combines chemistry-conditioned token embedding, lightweight sequence mixing, sparse expert routing, decoupled temporal modeling, and a decision-oriented SoH regression head. An enhanced USDTPlus variant further exposes voltage-shape information through local difference channels, attention pooling, and tabular skip connections.Experiments are conducted on the processed PulseBat dataset, comprising 270 retired lithium-ion cells across LFP, LMO, and NMC chemistries. We evaluate three complementary regimes: standard battery-disjoint interpolation, seen-SOC testing, and unseen-SOC generalization. The obtained results shows that under the standard battery-disjoint interpolation protocol, strong tabular baselines remain highly competitive, with RBF-SVR achieving the best standalone regression accuracy. However, under the stricter unseen-SOC protocol, where models are trained on SOC values of 5, 25, and 50\% and tested on previously unseen SOC values, USDT-family models clearly outperform the tabular baselines. USDTPlus achieves the best unseen-SOC performance, reducing MAE by 42.5\% relative to RBF-SVR and improving $R^2$ from 0.7215 to 0.9193.Beyond regression accuracy, we evaluate reuse-threshold decision quality at $\mathrm{SoH}\geq0.80$ and quantify retained reusable energy capacity and avoided CO$_2$ under scenario-based environmental assumptions. In the unseen-SOC setting, USDTPlus substantially reduces false rejects compared with RBF-SVR, preserving more viable cells for second-life use. Under the central environmental scenario, USDTPlus increases retained reusable energy capacity and net avoided CO$_2$ relative to the strongest tabular baseline. These results show that the value of chemistry-conditioned sequence modeling is most evident when SoH estimation is used not only as a fixed-split regression task, but as a deployment-oriented decision tool for circular battery management.

This publication has 1 reference indexed in Scilit: