LLM-MSIformer: A method for lithium-ion battery state of health estimation based on a large language model and multi-time scale interval transformer

Accurate state of health (SOH) estimation of lithium-ion batteries (LIBs) is critical for the safe and stable operation of battery system. Current SOH prediction models generally suffer from three critical limitations: insufficient robustness under extreme operating conditions, inadequate utilization of multimodal data, and poorly characterized coupling mechanisms of multi-timescale degradation features. To bridge this gap, this study proposes an innovative SOH prediction framework combining Large Language Model (LLM) with a Multi-Time Scale Interval Transformer (MSIformer). Firstly, the incremental capacity analysis (ICA) curves are derived from the charging voltage and capacity of LIBs. Then, health factors (HFs) are extracted from the charging data and ICA curves using Pearson correlation coefficient analysis. Subsequently, contextual prompts and textual metadata are fed into the pre-trained LLM, which generates denoised text embeddings through its noise-resilient semantic reasoning and contextual inference capabilities. After that, a hierarchical fusion mechanism combines the numerical features and LLM-derived contextual embeddings, enhancing interference immunity through complementary information integration. The MSIformer then hierarchically decouples multi-timescale dependencies from fused features via dilated convolutions and temporal attention, explicitly modeling cross-modal aging dynamics. Finally, cross-attention dynamically recalibrates multimodal features by aligning cross-modal multi-timescale dependencies with historical SOH trends, and the refined features are directly utilized for SOH prediction. The proposed framework is validated on a commercial 21700 and 18650 LIBs dataset, demonstrating statistically significant improvements over nine state-of-the-art benchmarks. Quantitative results show minimum reductions of 66.42% in RMSE, 70.19% in MAE, and 70.33% in MAPE, confirming both accuracy and operational robustness.