Multi-Agent Data Cleaning Framework of Solid-State Batteries using Large Language Models

Accurate data cleaning is important to ensure the quality of operational data for solid-state batteries (SSBs), improving the reliability of prognostics and management for electric vehicles. However, SSBs often show large short-term fluctuations due to relaxation or polarization, causing normal data to be misidentified as anomalies during data cleaning process. Besides, manual labeling of abnormal operation data is time-consuming. To address these challenges, large language model (LLM)-based multi-agent data cleaning framework for SSBs is proposed, which consists of different agents for operational indicators construction, anomalies recognition and revision in data cleaning tasks according the inputs and system prompts. The framework also develops a two-stage ensemble anomalies detection method (TETL) based on semi-supervised learning and designed multi-scale local outlier factor to address the anomalies recognition of SSBs. Additionally, 861818 samples from two types of SSBs (LiNiCoMnO2 and LiNiCoAlO) are collected to validate effectiveness of the proposed framework. The f1score that is the harmonic mean of precision and recall, is higher than 0.98 while other methods fail. The developed framework shows competitive results in anomalies recognition experiments, hyperparameters experiments, and anomalies revision experiments when compared to alternative methods. This study highlights the promise of LLM in the management for SSBs. • An improved two-stage ensemble abnormalities detection method (TETL) is designed. • TETL is used to reduces misrecognition in solid-state batteries data cleaning. • large language model (LLM) is designed to handle several data cleaning subtasks. • A multi-agent data cleaning framework is developed based on TETL and LLM. • The designed framework is verified by experimental data of solid-state batteries.