The construction of shield machine fault diagnosis knowledge graph based on joint knowledge extraction model

Traditional shield machine fault diagnosis methods rely on engineers’ experience and unstructured maintenance data, lacking a logically clear fault diagnosis knowledge base. Creating a fault knowledge graph can better organise, store, and manage complex fault information, aiding the development of automated fault diagnosis. However, current methods struggle with joint learning tasks for entity recognition and relation extraction, especially with polysemy and relation overlap. This paper proposes a novel XLNet-BiLSTM-LSTM model for knowledge extraction. The pre-trained XLNet model uses dynamic word vectors to serialise the text, making the contextual semantic representation more accurate. The Bi-directional Long Short-Term Memory (BiLSTM) encoding layer captures deep contextual features of the text. The LSTM decoding layer handles complex contextual dependencies and long-distance relationships, enabling joint decoding. Experimental results indicate that this model enhances the joint extraction of fault entities and relations, achieving an F1-score of 86.91%. Additionally, this paper introduces a new method for joint annotation of entities and relations, enabling the model to address the issue of overlapping relationships. Based on this, a construction framework for the shield machine fault diagnosis knowledge graph is proposed, ultimately developing a shield machine fault diagnosis knowledge graph comprising 1,330 triples.