Semi-supervised ISA: A novel industrial knowledge graph construction method enhanced by the fault log corpus analysis and semi-supervised learning

In industrial systems, knowledge graph-based intelligent fault diagnosis methods utilize extensive textual information, such as accumulated fault logs, to effectively construct domain-specific knowledge graphs. These graphs facilitate the use of unstructured data, thereby enhancing both diagnostic efficiency and accuracy. However, much of the existing research applies general knowledge graph construction methods to industrial fault diagnosis, without adapting them to the specific characteristics of fault logs. This oversight poses challenges in ensuring adequate and accurate model training. To address these challenges, this paper offers a comprehensive analysis of the essential attributes of fault logs, and proposes a semi-supervised industrial-adaptive knowledge graph construction method. The method employs a BiLSTM-BIO-based named entity recognition model, followed by a testing-enhanced self-attention relation extraction model designed for semi-supervised learning patterns. The extracted entities and relationships are organized into triplets to construct the knowledge graph. Finally, the proposed method is evaluated using fault logs from a specific heavy-duty train model. Extensive comparisons with various existing knowledge graph construction methods demonstrate the superior performance of the proposed method.