Transformer Fault Recognition Based on Kbert Text Clustering Model

The report text of transformer fault analysis contains the description of the equipment fault, but the description language is highly professional, and there may be multiple types of faults in a single fault description text. This paper proposes a machine identification algorithm Kbert (BERT+K-Means++) for transformer fault description to cluster specific fault types. Firstly, the transformer fault text is converted into a batched high-dimensional text matrix. Secondly, the key weight parameters in the Chinese BERT model are iteratively improved by using the fault description text to obtain a global semantic vector. At the same time, in the iterative improvement, according to the difficulty of sample fitting, the Kbert model’s recognition error weight for the sample is dynamically modified. Finally, through the K-Means++ algorithm, Kbert improves the original single BERT model that is difficult to deal with a single fault text containing multiple fault types. The calculation example identifies 782 real transformer fault analysis report texts across the country. The results show that the performance index F1 of the Kbert model is superior to the commonly used Bi-LSTM+Attention method, overcoming the problem of poor machine training caused by the long distance of the transformer fault description text and the uneven classification of sample types, which has achieved high accuracy and rapid cluster recognition of various transformer fault information. © 2022 Science Press. All rights reserved.