Bridging expert knowledge and sensor measurements for machine fault quantification with large language models

Data-driven fault diagnostics and failure prediction in the industry are dominated by numerical data from sensors. Other data, such as inspection notes, are often discarded due to complexity in feature extraction, fluctuations in information quality, and subjectivity due to human factors. However, for systems where degradation evolves slowly, these data could offer critical insight into the fault diagnostic process. Particularly in the task of quantifying the degradation level of inspected components, the observations made by technicians could hold critical information. Extracting useful information from text data has long remained a challenging task, especially with industrial reports. Therefore, this study presents a method utilizing large language models (LLMs) for extracting machine degradation information from brief on-site technician inspection notes. Industrial proprietary texts serve as background knowledge, with LLMs acting as the embedding medium. The proposed method’s performance is investigated through a real industrial case study, where identification of machine degradation level is enhanced using text data and domain knowledge.