Research On Automatic Question Answering Of Wind Power Fault Based On Knowledge Graph

With the depletion of fossil energy,China vigorously develops wind power in the context of seeking green,clean and alternative energy,and the installed capacity of wind turbines continues to grow.However,the environment of the wind site is harsh,resulting in frequent fan failures.The knowledge of wind power failure and wind power operation and maintenance only exists in various fault reports and some professional book texts.Some repetitive knowledge cannot be effectively used,and a large amount of wind power failure knowledge cannot be effectively stored and used.This paper uses natural language processing technology to process wind power fault data,constructs wind power fault knowledge graph,realizes fault data storage and realizes wind power fault knowledge question and answer based on knowledge graph.The specific research and work are as follows:(1)Research and build a knowledge map of wind power failures.By analyzing the wind power fault data,it is obtained that the wind power fault data has two types of structures: structured and unstructured.For the structured data,the knowledge graph can be directly constructed by using its own structural characteristics.For unstructured data,through the bottom-up model analysis of fault data,the entity concept and entity relationship concept of wind power fault knowledge graph are designed,and 6 types of entity types and 6 types of inter-entity relationship types are defined.In order to improve the entity extraction efficiency and accuracy of wind power fault text,this paper proposes a wind power fault text entity extraction method based on BiLSTM-CRF.Using the feature that fault text data is short text,sequence annotation is performed and BiLSTM-CRF is used to achieve automatic entity extraction.Compared with the Hidden Markov(HMM)model and LSTM-CRF,BiLSTM,and Conditional Random Field Model(CRF),the F1 score of the extraction effect has been improved by 8.1%,1.3%,1.9%,and 6.3%,respectively.The R-BERT model,an entity relationship extraction model based on the BERT pre-training model,is proposed.By labeling the entity locations in the wind power fault text after entity extraction,the pre-training model BERT is used to train the wind power fault text data.The comprehensive information extraction of vectors and tail entities improves the success rate of relationship extraction between entities.The model’s extraction success rate for 6 types of relationships is above 85%.Finally,the extracted entities and relationships are stored in Neo4 j in one-to-one correspondence in CSV file format to complete the construction of the wind power fault knowledge graph.The constructed knowledge graph includes 3874 entity nodes and 9769 edges connecting the nodes.(2)Research and realize automatic question answering of wind power failure based on knowledge graph.A convolutional neural network-based user question intent recognition model is designed,which turns user intent recognition into a short text classification problem.The word vector representation of question text based on word2 vec is designed,which improves the classification accuracy.The comparison experiments with SVM and BERT models confirm that this model has significant advantages in accuracy and training and testing time consumption.Using jieba word segmentation and part-of-speech tagging,the text entity extraction of user questions is realized,and the entity linking is realized based on cosine similarity algorithm.Finally,using the python framework to implement a wind power fault question answering system based on knowledge graph.The question-and-answer system can ask questions about wind turbine faults,which can improve the efficiency of users’ acquisition of wind power fault knowledge.