LLM-TSFD: An industrial time series human-in-the-loop fault diagnosis method based on a large language model

Industrial time series data provides real-time information about the operational status of equipment and helps identify anomalies. Data-driven and knowledge-guided methods have become predominant in this field. However, these methods depend on industrial domain knowledge and high-quality industrial data which can lead to issues such as unclear diagnostic results and lengthy development cycles. This paper introduces a novel human-in-the-loop task-driven approach to reduce reliance on manually annotated data and improve the interpretability of diagnostic outcomes. This approach utilises a large language model for fault detection, fostering process autonomy and enhancing human–machine collaboration. Furthermore, this paper explores four key roles of the large language model: managing the data pipeline, correcting causality, controlling model management, and making decisions about diagnostic results. Additionally, it presents a prompt structure designed for fault diagnosis of time series data, enabling the large language model to realize task-driven. Finally, the paper validates the proposed framework through a case study in the context of steel metallurgy.