Research On The Construction Method Of Industrial Robot Fault Diagnosis Knowledge Graph

In the industrial digital age,a large amount of data is generated,and how to convert the data into valuable knowledge for utilization is a problem worth studying.As a key technology in intelligent manufacturing,industrial robots play an important role in the entire automated production process.As the system structure of industrial robots is becoming more and more complex,once a failure occurs,it may affect the entire production cycle.It is crucial to accurately diagnose industrial robots and make th em in the best working condition.With the development of knowledge graph,more and more attention is paid to transforming valuable and experienced data into knowledge for analysis.Therefore,the topic of building a knowledge graph for industrial robot fa ult diagnosis is proposed.It is of great research significance to explore the complex relationship between faults,realize the reuse and analysis of knowledge,and provide semantic search,auxiliary decision-making,diagnostic reasoning and other applications for maintenance personnel.For the process of knowledge graph construction,the research work in this thesis mainly includes:(1)Construct a dataset in the field of industrial robot fault diagnosis.Since there is no data set in this field on the public network,the relevant data sentence information is organized for related documents such as the fault maintenance table.Through expert consultation and guidance,the ontology concept relationship is constructed,and seven kinds of entity information and four kinds of relationship information are divided.The data is marked to form a small-scale data set IRFDC(Industrial Robot Fault Diagnosis Corpus)in the field.(2)Research on joint extraction algorithm of multi-head selection combined with attention mechanism.Aiming at the traditional pipeline method can not well integrate the relationship between the two subtasks,a joint extraction model BABCJ(Bert-AttentionBi LSTM-CRF in Joint Extraction of Entities and Relations)is proposed for entity relationship extraction.On the basis of multi-head selection,the pre-trained language model Bert and attention mechanism are introduced.The model effect of the comprehensive F1 value of 88.21% is achieved on the constructed dataset IRFDC.The effectiveness of the algorithm is verified through comparative experiments and ablation experiments,and the generalization ability of the model was verified on public datasets.(3)Research on entity recognition algorithms based on active learning.Due to the limited annotated text information in a specific field,fully supervised learning will bring about the problem of high annotation cost,so active learning is introduced to carry out research.Active learning can reduce the cost of manual annotation to a certain extent by selectively selecting samples to train the model.Based on active learning,this thesis first conducts research in the field of entity recognition.Through the design of query strategy,combined with the improved learning engine,the effectiveness of active learning is finally verified through experiments.(4)Knowledge storage and visualization based on Neo4 j.By using the Neo4 j graph database to store the triple information obtained in the process of information extraction,the knowledge storage operation is realized,and the visualization operation of the knowledge graph of industrial robot fault diagnosis is carried out in the Neo4 j graph database.