Knowledge Graph Construction Of Fault Diagnosis For Turbine Generator Set

As a crucial piece of equipment in the electric power field,once the turbine generator set breaks down,it will affect the whole production process.Intelligent operation and maintenance are vital for ensuring the generator set’s safe,stable and economical operation and also a pivotal link to realizing the “Intelligent Smart Turbine”.For the field of fault diagnosis for turbine generator sets,integrating the experience and valuable knowledge of domain experts to provide a basis for analysis by decision-makers and domain experts is one of the critical tasks in realizing intelligent operation and maintenance of turbine generating sets.Based on the fault maintenance case data and knowledge caused by turbine generator set in the process of operation and maintenance,as well as expert analysis experience,these data and knowledge are processed through Natural Language Processing(NLP)techniques such as semantic matching,named entity recognition,relationship extraction and entity alignment,and a knowledge graph of fault diagnosis for turbine generator set is constructed.So is to realize an intelligent health management system that integrates functions such as fault diagnosis,maintenance suggestion and maintenance decision,and provides auxiliary decision-making for the operation and maintenance of the turbine generator set.The main content and summary of this thesis are listed below.(1)A dataset of fault diagnosis for the turbine generator set is constructed.Based on the publicly available information on fault cases of turbine generator sets,six types of entity type information and six types of relationship type information are identified.The data are manually annotated to construct a named entity recognition dataset and a relationship extraction dataset of fault diagnosis for turbine generator set.This solves the problem of needing more publicly available annotation corpus datasets of entity relationships in constructing the knowledge graph of fault diagnosis for turbine generator set.(2)An algorithm for named entity recognition of turbine generator set is studied.A named entity recognition method combining multi-headed self-attention(MHA)and BERT-Bi LSTMCRF is proposed to address the difficulty of extracting association weight features in long texts in the specialized field of fault diagnosis for turbine generator sets.High quality fault entities are extracted from the fault text,thus providing for the construction of a knowledge graph management system for the sets fault diagnosis.(3)An algorithm for fault entity relationship extraction of turbine generate sets is studied.Considering the complex structure of fault case texts in the field of fault diagnosis for turbine generator sets and the problem that the traditional pipelined entity relationship extraction method cannot well integrate two sub-tasks,an end-to-end method for fault entity relationship extraction based on adversarial training is proposed for the extraction of entity relationships in the field of fault diagnosis for turbine generator set.Joint annotation of fault entity relations is performed to transform the joint extraction task into a sequence annotation problem.A bidirectional recurrent neural network is used as encoder to integrate the contextual features of text sentence sequences to maintain the semantic integrity of the text as far as possible.The decoder layer uses LSTM to generate labeled representations,and adversarial training is performed by introducing adversarial factors to improve the robustness of the model and the extraction effect.(4)A knowledge graph management system is designed and developed for fault diagnosis of the turbine generator sets.The Neo4 j database stores the extracted fault entities and relationships as a triad.A knowledge graph management system of fault diagnosis for turbine generator sets is designed and developed based on a B/S architecture.Based on this system,functions such as fault entity recognition,fault entity query,fault relationship query,fault knowledge graph construction,knowledge Q&A,and intelligent retrieval can be realized.