| Metallurgical equipment is under high load during long periods of time,and the working condition is very poor,which is very easy to cause equipment failure.However,equipment failures are mostly recorded by workers in the form of failure investigation sheets and O?M(operation and maintenance)history documents.There are features such as irregular content description,uneven quality of recorded information,and the non-editable image documents.There is no effective way to mine equipment failure O?M information.In turn,it is difficult to reuse the knowledge of O?M history,it affects the reference to the O?M knowledge when formulating the cross-enterprise equipment maintenance strategy,hindering the maintenance process and efficiency.In this paper,a metallurgical enterprise's O?M history documents and fault investigation sheets are used as the research objects,and an improved Mask R-CNN method is proposed.At the same time,combined with the Unet network segmentation table frame line and OCR technology,it is used to structurally identify table texts.A joint entity-relation extraction method is raised to build an O?M knowledge graph.The semantic similarity measurement method of the document of O?M history table is studied.The research work is as follows:(1)For at the problem that the image format of the resume form document of metallurgical equipment O?M makes the text characters in the form non-editable,this paper proposes an improved Mask R-CNN to locate table regions.At the same time,this paper combines the Unet network to segment the cells of the table,and the combination of rules based on the str uctural characteristics of the table and OCR technology,structured to identify the text of the cells.(2)In order to solve the problems of poor semantic description specification of O?M history text,inconsistent text quality and numerous redundant information,which leads to difficulties in extracting O?M information effectively,a joint entity-relationship extraction network is proposed which directly models triples with a sequence labeling strategy.The network is composed of BERT,Bi LSTM,Attention and CRF modules,which can directly obtain triple data.And The O?M knowledge graph is constructed,and the O?M knowledge is structurally related and provides a visual semantic network.(3)To address the problem of low similarity accuracy of reused O?M history documents,a semantic similarity measure for O?M history documents is proposed.In view of the company's existing prior information,the knowledge graph subgraph of the device tree to locate the target device class or instance set is constructed,and the graph network model is re built.Extract semantic feature vectors based on Graph SAGE graph neural network a ggregating node information of fault phenomenon subgraphs.Then use the cosine value to measure the similarity with the semantic feature vector of the trouble ticket.The most similar resume instances are sorted and pushed out.The resume document is presented in the form of a visual knowledge graph to assist the formulation of maintenance strategies.The case verification is based on a large number of driving operation and maintenance history documents and fault investigation sheets accumulated by a steelmaking enterprise as data sources,the document analysis and knowledge extraction were carried out,and the O?M knowledge graph and the reconstruction graph network model were constructed.Then,the semantic feature vectors of different graph networks are extracted to measure the similarity,and the most similar resume documents are pushed out for knowledge reuse.At the same time,this paper develops a resume document retrieval system,which provides cross-enterprise document parsing,knowledge extrac tion and similarity retrieval applications based on the shared nature of cloud storage.It realizes process-based O?M resume document retrieval and cross-enterprise knowledge reuse,which has certain significance for improving the O?M efficiency of metallurgical equipment and reducing the O?M cost. |
PDF Full Text Request