A Probabilistic Classifier for Transformer Dissolved Gas Analysis Using Various Input Variables

Dissolved gas analysis (DGA) has been an important technique for transformer incipient fault diagnosis. A probabilistic classifier using diffusion-based kernel density estimator is proposed for DGA analysis. Two types of input variables, namely ratio variables and hybrid variables, are discussed. A ratio variable is derived from feature vectors used in traditional DGA method while a hybrid variable is a combination of two different ratio variables. Input variable components with dependence and independence are considered for comparison purposes. Experimental results show that the proposed model yields high diagnosis accuracy. Diagnosis accuracy is even better when hybrid variables are adopted with the independence assumption on some ratio components.