Research On The Methods Of The Power Knowledge Graph Construction Based On Crowdsourcing And Deep Active Learning

With the increasing scale of power system,new knowledge is pouring in continuously,the total amount of knowledge in the system is increasing explosively.The existing power knowledge organization and management can not meet the increasing application of power system knowledge.Knowledge mapping is a new and efficient technology for knowledge organization and management.The application of knowledge map in electric power field can organize complex electric power knowledge network in series and improve the information retrieval ability of electric power network.Knowledge extraction technology is needed to construct power knowledge atlas.The construction of knowledge extraction model needs a lot of manual marking data,and the training of the model depends on a lot of computing resources.In order to solve the above technical difficulties,this paper mainly focuses on the power scenario,based on deep learning and machine learning methods to study the low-cost construction method of power knowledge atlas,this paper summarizes the factors that affect the construction of knowledge map for power scene.The main tasks are as follows:(1)Knowledge mapping as a novel and efficient technical means of organizing and managing knowledge information.The power system information is organized in series through core elements such as triads.The construction of knowledge graphs relies on crowdsourced annotated data.However,the quality of crowdsourced annotated data varies and lacks effective quality evaluation methods.In this section,the crowdsourced data quality evaluation algorithms based on KF-Bert(Bidirectional Encoder Representation from Transformers with K Fold Test)and Cart-DT(Classification and Regression Decision Tree)data quality evaluation algorithms.Inspired by the idea of K-fold validation,we realize the effective fusion of different labeled data.Based on the decision tree classification model,the quality evaluation of data annotated by different annotators is completed.It reduces the involvement of expert system in data annotation,reduces the cost of data pre-processing,and provides a high-quality data base for subsequent text classification work.(2)The construction of knowledge graphs is beneficial for grid production,electrical safety and security,fault diagnosis and observability.High-precision text classification algorithm is the key to build the expertise graph of electric power system.However,there are many poorly described and specialized texts in the electric power business system,and the amount of data containing valid labels in these texts is low.This will bring a great challenge to the improvement of text classification model accuracy.To solve the problem,this chapter proposes a classification algorithm for Chinese Text in the Electric Power Industry based on deep active learning(CCTP-DAL).Taking full advantage of Transformer multi-attention mechanism in high-dimensional data processing and automatic feature extraction,combining with the hierarchical confidence active learning mechanism has great potential of in reducing model training data requirements.With less data annotation cost and very low computational resources,we obtain a high-precision power text classification model and do the groundwork for the construction of the power knowledge graph.(3)Based on the above work,this paper designs and realizes information security knowledge map oriented to power field.Firstly,the knowledge representation and knowledge storage in the process of graph construction are introduced,including the generalization of knowledge ontology in security domain,the description language of information security ontology and the batch storage of power information security knowledge,then the dynamic search and visual display function of knowledge map is presented by using “Transmission network” as the key word.Finally,the path query function of power knowledge map is introduced,to help operators understand the power complex coupling business.