基于深度学习的机载传感器领域知识图谱构建研究与实现

ABSTRACTABSTRACTWith the continuous growth of the global aviation maintenance industry (MROMaintenance,Repair Operations)market,China's aviation MRO industry has alsoushered in many development opportunities and challenges.How to do a good job oflayout and planning according to its own development status to meet the new wave ofdevelopment has become a topic worthy of attention in the domestic aviationmaintenance industry.It is very meaningful for my country's civil aviation maintenanceindustry to make good use of the development of advanced technology and rely on thepopular big data to study civil aircraft maintenance programs to find more rationalizedsolutions that enhance aircraft performance and reduce maintenance costs.Most of theaircraft that China's civil aviation flies use fly-by-wire control,and the number andtypes of airborne sensors are numerous.The big data application technology ofprocessing related data represented by the knowledge graph is one of the methods tosolve the better aircraft maintenance program with airborne sensors.The core content ofthe airborne sensor knowledge graph construction includes entity recognition andrelationship extraction.Based on the text information of the airborne sensormaintenance log of the author's MRO company,this article deeply studies the airbornesensor knowledge graph entity recognition and relationship extraction issues,the mainresearch content include:(1)In the problem of airborne sensor entity recognition,a converter-basedbidirectional encoding representation (BERT)combined with two-layer bidirectionallong short-term memory network (BiLSTMs)and conditional random field (CRF)isproposed for airborne sensor entity recognition The method is to label each word of itsPosition embedding for the BERT preprocessing model,and then connect the two-layerBiLSTM model to obtain a better context representation.Combine the Attentionmechanism to obtain word strength information and reduce the huge amount brought bythe three-layer model.Calculate the amount,and finally obtain the sequence annotationresult through the CRF model.The experimental results show that the F value isincreased by 4.03%compared with the traditional CRF method.(2)In the problem of mechanism sensor relationship extraction,the BiLSTM-ATTmodel is used.Firstly,each word vector in the text information is fused with the