Transformer fault diagnosis method based on k-nearest-neighbor SMOTE (Synthetic Minority Oversampling Technique) and deep learning

Abstract

The invention discloses a transformer fault diagnosis method based on k-nearest-neighbor SMOTE (Synthetic Minority Oversampling Technique) and deep learning. The method includes steps of: preprocessing on an initial unbalanced sample set, DNN neural network training and diagnosis testing. According to the method, firstly clustering is carried out on minority samples before interpolation, interpolation is carried out by areas divided by clustering, and the problem that existence of generalization in a process of inserting data leads to data distribution marginalization and positive-negative class boundary blurring is avoided; and generated interpolations are on connecting lines of cluster cores and original minority sample points, thus no situation of data distribution marginalization can exist, and a diagnosis effect is improved.

Description

technical field [0001] The invention relates to a transformer fault diagnosis method, in particular to a transformer fault diagnosis method based on k-adjacent SMOTE and deep learning, and belongs to the technical field of power supply. Background technique [0002] Transformer faults are the result of the combined effect and long-term accumulation of the transformer itself and its application environment. The characteristic quantities of faults are various, and the relationship between fault characteristic quantities and fault mechanisms is also intricate, which makes it very difficult to establish a transformer fault diagnosis model. [0003] In transformer fault diagnosis, since transformer fault is a low-probability event, the real distribution of transformers in abnormal state is very small. At the same time, there are problems of incomplete record information in transformer fault case reports and incomplete record information in the case database. The fault characteriza...

AI Technical Summary

Problems solved by technology

First, SMOTE has a certain degree of blindness when determining the neighbor value k. k is a hyperparameter, that is, a parameter artificially specified, and k is an empirical value, so there is a certain degree of subjectivity in the selection of neighbors, which may lead to poor sampling

Second, after SMOTE oversampling, the marginalization of data distribution is easy to occur, which changes the data distribution of unbalanced data sets, resulting in the problem of fuzzy positive and negative class boundaries.

Images