Research On Text-Oriented Fault Model Mining Technology

Chemical accidents will not only bring huge economic losses to enterprises and individuals,but also cause irreparable damage to the environment,and even endanger human life and safety.Accident analysis is an effective measure which has been proved by practice to reduce accident risk step by step.Accident analysis is the process of describing the accident process,analyzing the causes of the accident,summing up experience and putting forward preventive measures.The results of accident analysis are finally stored in the accident report in the form of text.Although the accident report can express the cause of the accident one by one,it can not give the expert intuitive understanding of the cause and effect logic;the fault tree model can clearly show the logical relationship of the accident cause,but only describe the evolution process of the accident,without describing the time,place and environmental state in detail,which may result in the loss of important information;and the process of artificial fault tree constructed method is very error-prone.There may be semantic ambiguity of node information,incomplete node information and mixed information.To solve these problems,this paper will build a text-oriented fault model based on fault tree.It can be used to replace the coarse-grained analysis of traditional fault tree to realize automatic,reasonable,accurate and efficient accident analysis.The main work of this paper is as follows:1.Event extraction.In event extraction,event quaternion is defined in this paper.Natural language processing technology is used to analyze accident report text and fault tree node information,and automatically extract accident elements.In this paper,an event extraction algorithm is designed: the participants and triggers of events are identified by semantic role labeling,the location and timestamp of events are identified by semantic dependency parsing,the subject and object structures of event sentences are identified by dependency parsing,and events are represented in the form of "subject + predicate + object".In the experiment,a total of 2235 events were extracted and 7464 nodes were constructed.According to the semantic features and context features of event entities,this paper implements event knowledge fusion.2.Event relation extraction.In event relation extraction,this paper defines event relationpattern,which is mainly used to extract event causality.In this paper,we extract event elements from event sentences by syntactic and semantic analysis,analyze fault tree logical relationship recognition and extract event explicit relations,and propose an implicit causality extraction method based on bidirectional GRU.Event sentences containing two events are represented by word-embedded sequence.The word-embedded sequence as input of BiGRU neural network model.Attention mechanism is added to weigh sentence vectors,and LN algorithm is introduced to normalize the convergence efficiency of BiGRU neural network.The semantic features of event sentences are extracted,and the semantic features and three kinds of internal structure features are fused into the softmax classifier to achieve implicit causality extraction.Experiments show that the accuracy of this method is 94.88% for fault tree-oriented chemical event relationship extraction,which proves the validity of this method in fault tree event relationship extraction.3.Application of event evolution graph.In displaying the application of the event evolution graph,this paper uses the method of the event evolution graph constructing in this paper to construct and display event evolution graph based on fault tree and text analysis,which is used for deep search,reasoning of accident evolution process and reasoning of relationship.