| Oil-immersed transformers are key equipment in power transmission and transformation system,whose operating conditions directly affect the safety and economics of the power grid.A large number of oil-immersed transformers have a long period of operating,hence in certain risk of insulation deterioration.Therefore,accurate state evaluation and fault diagnosis of transformers is of great significance for ensuring reliable operation of the power grid.The traditional methods of transformer fault diagnosis and condition evaluation rely on the experience of operation and maintenance personnel and expert rules,fail to effectively use multi-source heterogeneous power equipment data,so they are deficiencies in objectivity and accuracy.Based on machine learning and natural language processing related methods,this paper aims to perform in-depth mining of transformer online monitoring data and transformer defect text data to optimize the effects of key steps such as data processing,feature extraction,and model building,thereby improving transformer fault diagnosis and status evaluation accuracy.Firstly,in order to overcome traditional transformer fault diagnosis methods’ drawbacks including sensibility to noise,low diagnostic accuracy and difficulty to determine model parameters,a transformer fault diagnosis method based on data preprocessing and Gradient Boosting Decision Tree(GBDT)is proposed.The features that represents working condition of transformers are obtained based on data of dissolved gas in transformer oil.Multistage data preprocessing method is used to identify and replace noisy data to obtain de-noising samples.Then a fault diagnosis model based on GBDT is constructed.Bayesian optimization algorithm is used to further improve the diagnostic accuracy.Secondly,in order to improve the comprehensiveness of transformer condition evaluation,the description text of transformer defect phenomenon is used as an important supplement for equipment fault diagnosis and condition evaluation to realize a wider range of data mining and applications.Based on the convolutional neural network optimized by the attention mechanism,a text classification model is constructed to judge equipment defect levels,which reaches higher classification accuracy than traditional models;based on short text matching algorithms such as BM25,the accurate matching of defect phenomena is achieved and the matching results are used to further enrich the conclusion of condition evaluation;based on entity extraction and attribute value filling technology,key indexes and their attributes in text descriptions are extracted which lay the foundation for heterogeneous data fusion diagnosis.Finally,combined with the research results of the above chapters,a fault diagnosis and status evaluation system of transformer based on machine learning and natural language processing was developed.According to the operation and maintenance needs of a certain region of Zhejiang Province,the preliminary transformation of theoretical research results into engineering practice was achieved.The main functions of the system,development tools,software and hardware architecture are introduced in detail in this chapter. |
