Defect Diagnosis Method of Main Transformer Based on Operation and Maintenance Text Mining

Abstract

During the operation and maintenance of electrical equipment like power transformers, the information of defects or faults are usually recorded by text data. However, the method of classifying transformer defects by text data relies on manual at present, which is inefficient and uneconomical. This paper presents a recurrent convolutional neural network with Bayesian optimization, which construct a text classification model can automatically classify the power text data. Firstly, the text is preprocessed, including the establishment of the transformer's dictionary, word segmentation of defect text and then mapping the results to word vectors based on distributed representation. Furthermore, training the CNN and RCNN networks by supervised learning. It is worth mentioning that in RCNN network, the Bi-LSTM structure is used instead of the convolutional layer, which can learn the semantics of the text more effectively. In addition, in order to obtaining a better classification effect after training, the Bayesian method is used to optimize the hyper-parameters of the networks. Finally, On the test dataset, two kinds of network achieved 90% test accuracy and 92% test accuracy, respectively.