A Large Language Model Based Knowledge Mining Method for Improving the Reliability of Fire Water Systems

The fire water system plays a critical role in protecting both infrastructure and human lives. An essential aspect of enhancing the reliability of this system is fault diagnosis. However, the current fault diagnosis methods primarily rely on data-driven approaches, which often result in a high threshold for application due to their lack of interpretability. To tackle this challenge, this paper introduces a novel approach based on large language models for knowledge mining from textual data to extract fault information related to the fire water system, thereby enhancing the interpretability of data-driven fault diagnosis methods. The methodology followed in this paper consists of two main steps: firstly, analyzing the characteristics and principles of fire water system faults to develop a fault ontology, and secondly, creating a knowledge mining model using a large language model guided by the established fault ontology. Experimental findings indicate that the proposed model achieves an F1 score of 0.944, meeting the necessary criteria for effective knowledge mining in fire water system fault analysis. Furthermore, a comparative experiment was conducted to evaluate the performance of various encoder models, including GRU, BiGRU, LSTM, BiLSTM, and pre-trained large language model BERT. The results revealed a significant improvement in performance with the BERT encoder, showing increases in F1 scores of $22.12 \%$, $2.27 \%, 17.41 \%$, and $3.16 \%$ compared to the other models, respectively. This study provides valuable interpretative insights that can enhance the engineering applicability and reliability of data-driven fault diagnosis methods in fire water system.