Incremental Learning-based Knowledge Graph Construction Technology For Petrochemical Safety Domain

In the field of petrochemical safety,safety analysis and evaluation information has an important role to play.HAZOP analysis plays an important role in the field of petrochemical safety as a proposed analysis and evaluation method for petrochemical safety.There is a lack of more intuitive and effective organization of petrochemical process safety analysis information resources.There is a lack of effective resource integration and reuse sharing between multiple sources of information data.To address this problem,this study takes HAZOP analysis data as a foothold and carries out research on knowledge mapping construction technology in petrochemical safety field.The HAZOP knowledge ontology(PSHPOntology)is constructed by referring to the national standard ISO15926 for knowledge modeling.Seven basic elements are defined,and each type of element is described in detail and specifically.Finally,the seven steps and Protégé are combined to improve the construction of PSHPOntology.The primary task of knowledge extraction is entity extraction,also known as named entity recognition.In this study,an incremental learning strategy is adopted for entity extraction.Considering the characteristics of HAZOP analysis report such as confidentiality,multiple sources and sustainable update,a model of Cimb NER under the knowledge distillation framework is proposed.The model is able to learn new knowledge without catastrophic forgetting of old knowledge through incremental learning strategies.For the specificity of HAZOP text constructions,the innovation of graph neural network in this study aims to be able to better determine the boundary of entity recognition and improve the efficiency of entity recognition.Another task of knowledge extraction is relationship extraction.In this study,relevant relationships in the field of petrochemical safety are predefined.To avoid the problems of neural network redundancy and inability to long information dependency,the liner embedding method is used to achieve relationship extraction.In the knowledge graph complementation section,this study proposes an incremental learning based knowledge graph dynamic complementation method.The method utilizes frequency-based empirical replay and temporal regularization to improve the performance of the model over current and past time steps.Finally,a relevant case study is presented using the HAZOP meeting analysis record of a node in the oil synthesis unit of an indirect coal liquefaction project as a case study.Firstly,the entity extraction results for this HAZOP meeting analysis record are shown,and then the relationships are added to it.Finally,a petrochemical safety domain knowledge graph(PSKG)is constructed.The PSKG shows the close connection between safety analysis and evaluation information.Compared with the original HAZOP meeting analysis records,it visually shows the various types of information involved in a certain node of the petrochemical process.The organization of information resources between HAZOP data is enhanced,enabling HAZOP reports to be shared and reused.