Research On Applications Of Unstructured Data Mining In Power Equipment

As the construction and development of intelligence and informatization in China's power grid,a great number of power data in various forms have accumulated through long-term operation,maintenance and test of the power equipment.Among them,compared with the structured data stored mainly in numerical form,unstructured data mainly exist in the form of texts,images,sounds,videos and so on,which have wider application scenarios and higher value density.However,because unstructured data cannot be recognized and processed directly by computers,there are more difficulties in the mining process.Therefore,this thesis takes two types of typical unstructured data of power equipment,text data and image data,as examples,and studies the mining applications of these two types of data in the defect assessment,defect processing and state recognition of power equipment.The achieved results are as follows.1.In view of the fact that a large number of defects in power equipment need to be classified manually,an automatic text classification method for defect records of power equipment based on convolution neural network is proposed.The text representation technology based on word vectors and the text classification technology based on convolution neural networks are used to construct the classification model,and the model structure is improved according to the characteristics of the textual data of power equipment,which improves the accuracy and efficiency of classification and ensures that defects can be treated and reported in time.2.Due to the complexity and variability of power equipment defects,many defect treatment decisions often lack reference.To solve this problem,a text retrieval method of power equipment defect records based on knowledge graph technology is proposed.By constructing the knowledge graph of power equipment defects automatically and using the graph searching technology,the accurate retrieval of similar historical defect records through current defect records is realized,so that the historical defect treatments can be effective reference for the current ones,and provide guidance for those who are relatively deficient in knowledge and experience of defect treatments.3.To detect the categories and positions of various power components in power equipment inspection images automatically,an object detection method for power equipment images based on improved Faster R-CNN is proposed.Taking the inspection images of main transformers as an example,the significant difference in component sizes and the relations between component positions are considered,and the structure of Faster R-CNN is improved,which obtains higher accuracy in category and position recognition for multiple components,and provides a foundation for detecting defect and fault phenomena of different components.4.To solve the problem that the algorithm effect of power equipment state recognition is limited by training samples,an augmentation method for power equipment images combined with three-dimensional spatial information is proposed.Taking the inspection images of disconnecting switches as an example,combined with the prior knowledge about the three-dimensional shapes of the disconnecting switches,the training samples of the disconnecting switch images with different shooting angles are augmented by using the methods including perspective projection transformation and three-dimensional rotation transformation.Compared with the traditional image data augmentation methods,the proposed method can improve the position and state recognition effect of disconnecting switches more significantly.