An Effective Knowledge Mining Method for Compressor Fault Text Data Based on Large Language Model

The fault diagnosis method of compressors determines the reliability of the gas transmission pipeline station. Existing compressor fault diagnosis methods mostly relies on data-driven, which leads to a high application threshold from the mechanism. To address this issue, this paper introduces the knowledge graph into the compressor fault diagnosis for the first time and proposes a compressor fault text data knowledge mining method based on large language model. Firstly, the characteristics and principles of compressor faults are analyzed. Then, a text data knowledge mining model called CFRTE for compressors is constructed. Experimental results show that the Fl score of the CFRTE model can reach 0.98, meeting the requirements of compressor fault knowledge mining. Finally, combined with the results of knowledge mining and the graph database, a new system for the storage and indexing of the compressor fault knowledge graph is proposed. To further verify the role of the large language model in compressor fault knowledge mining, this paper conducts a comparative experiment of CFRTE models based on RNN encoder and BERT encoder. Experimental results show that compared with GRU, BiGRU, LSTM, and BiLSTM as the encoder layer, the Fl score of the CFRTE model with BERT as the encoder layer has increased by 26.78%, 6.18%, 21.89%, and 5.49% respectively. This work provides a systematic feasible scheme for introducing knowledge graphs into compressor fault diagnosis, which can be used for reference in the fault diagnosis of related equipment.