Fine-Tuned Large Language Model Empowered State of Health Evaluation of Lithium-ion Batteries based on Charging Sequence Matching

Accurate, convenient, and reliable state of health evaluation is critical for safe and efficient operation of lithium-ion batteries in electric vehicles. However, complex operating conditions, stringent feature engineering, and data scarcity severely limit the applicability of conventional estimation methods in engineering practice. Here, we infuse classification into state evaluation, breaking the existing mindset for battery degradation identification. Considering the potential limitations of data-driven technologies, this work employs emerging large language models from natural language processing to replace smaller models. Over 80,300 cyclic samples from 311 cells across large-scale datasets underpin this investigation. Firstly, incomplete charging data is scanned, fused, and dimensionally reduced to transform into a lightweight charging sequence. Thereafter, the current charging sequence is combined with the truncated candidate sequences to empower sequence matching, bypassing inherent degradation feature extraction. Furthermore, we devise various model construction strategies to fine-tune the model, seamlessly integrating prior knowledge with domain-specific insights. The validation results demonstrate that the proposed paradigm reliably exhibits robust evaluation performance, with an overall accuracy exceeding 99%. This work highlights the potential of large language models in battery intelligent management without requiring additional sensors, opening avenues for further interdisciplinary exploration. Infusing classification into state evaluation. • LLMs have the potential for interdisciplinary exploration in battery intelligent management. • SOH is a critical indicator supporting the safe and efficient operation of electric vehicles. • Innovative infusion of classification into state evaluation to tackle regression tasks. • Parameter-efficient fine-tuning is employed in LLMs to optimize accuracy and generalization. • Stable evaluation performance on large datasets including 311 cells.