Research On Intelligent Recommendation Method Of Safety Production Inspection Based On Knowledge Graph

As an effective method to ensure the operation of the city,safety production inspection plays a key role in reducing and preventing production safety accidents.However,the large number and wide variety of hidden entities contradict the limited law enforcement forces,which results in information overload for relevant law enforcement agencies,and leading to repeated or missed inspections.In order to solve this contradiction,this paper constructs the knowledge graph of production safety law enforcement elements through the law enforcement data of production safety,utilizes the rich semantic relations in the knowledge graph as auxiliary information to assist the recommendation model to improve the recommendation performance.The model we proposed recommends the places and devices that need to be inspected in different enterprises for law enforcement departments to help accurate law enforcement and improve the efficiency.The main work of this paper is as follows:(a)Proposed an entity alignment algorithm based on ERNIE-HNSW semantic graph structure retrieval.To solve the problem of non-standard entities in multi-source data and manually records,and to achieve standardization of safety production corpus,this method first uses ERNIE to learn the semantic representation of safety production hazardous entities based on Chinese data sources,and obtains the semantic feature vectors of hazardous entities.A hierarchical network structure is used to construct an index,and the HNSW graph structure retrieval algorithm is used to obtain the alignment results of safety production hazardous entities.By processing 33,004 safety production inspection records,a corpus of places and devices in hazardous chemical fields was formed,including 8 types of hazardous-related industries.At the same time,a safety production enforcement interaction matrix composed of hazardous enterprises and inspection entities is sorted out.(b)Presented a construction method of knowledge graph on safety production inspection elements combining a top-down and bottom-up approach.According to the needs of safety production enforcement business,we start building the ontology from the top-level concepts and complete the ontology design of the pattern layer by integrating structured data,such as standard documents of hazardous field,with hierarchical categories and corresponding relationships.Then,the extracted and aligned knowledge of entities,such as places and devices,is filled into the pattern layer ontology,completing the construction of the data layer,and forming a knowledge graph on safety production inspection elements consisting of 6726 triples of 9 entity types and 8 relationship types.(c)Proposed a hazardous entity recommendation model for safety production inspection based on multi-task learning,named CKR.In order to improve the recommendation of hazardous entities in safety production inspections,a multi-task learning model CMKR was proposed,which enables knowledge graph recommendation tasks and knowledge graph representation tasks to promote each other during training.To address the problems of severe overfitting and sparsity of the safety production dataset,a multi-layer CNN-1D was used instead of the MLP in the low-level,and the ConvKB knowledge graph representation learning method was used in the high-level to improve the model’s generalization ability.CTR prediction and Top-K recommendation prediction experiments are carried out in hazardous entity recommendations for safety production inspection.Results of the experiments illustrate that CMKR performs best in CTR prediction and Top-K recommendation prediction among baselines and variants of CMKR,including MKR+CNN-1D and MKR+ConvKB.During CTR prediction,CMKR obtains the best AUC(0.7112)and ACC(0.7061)in Place dataset and the best ACC(0.6803)in Device dataset.Compared with the previous algorithm model,this method effectively controls the overfitting problem,improves the overall performance,and realizes the intelligent recommendation of the key places and devices in the safety production law enforcement.The application of the safety production enforcement content recommendation in the actual scenario of the "precision enforcement system" shows that the enterprises in the same industry have both similar and different recommended hazardous entities.What’s more,recommended entities and historical records have corresponding hazardous entities.The recommendation results show the applicability and rationality of the recommendation in the actual application scenario.