A kind of semi-supervised classifying method research for power transformer fault diagnosis

Abstract

Dissolved gas analysis is one of the most common techniques to detect the faults in the power transformers. Most of existing diagnosis method will need large amount of labeled data sets to construct classifier, while normally ignoring without unlabeled data sets. This paper presents a power transformer fault diagnosis method which based on semi-supervised classifying. In Its learning process, the semi-supervised classifying method can simultaneously use labeled data sets and unlabeled data sets to acquire more information so that make better learning effect. A semi-supervised classifying (SSC) method adopting fuzzy nearest neighborhood label propagation (FNNLP-SSC)is adopted to diagnose the fault of power transformer, in the meantime, the proposed method, based on the similarity connections between a sample and its K nearest data, classifies the unlabeled data by making the labels propagate from the labeled data to unlabeled data. The experiments indicate that method of this paper has been proposed has higher fault diagnosis accuracy compared with C-means (FCM) algorithm and the three ratio method IEC. Also, it verifies the effectiveness and feasibility of the proposed method in the transformer fault diagnosis.