Diagnosis accuracy in electric power apparatus conditions using classification methods

The use of the decision tree method was recommended as a classification tool in diagnosing electric power apparatus because it provides the visible if-then rule, making it possible to connect the physical phenomena with the observed signals. Using a variety of feature variables extracted from the partial discharge patterns and others, the misclassification rates were found to be as small as 2% if results were obtained using training data only. In this paper, we assess the diagnosing accuracy of the classification methods using test data; we have found that the small values of the misclassification rates remain even when test data are applied. The appropriate methods perform fairly well, with misclassification rates of less than 5%. We conclude that although the misclassification rates by the decision tree are not as small as the values obtained by effective ensemble classifiers such as bagging and boosting, the decision tree is still useful and attractive because the method provides explicit rules, and the variability of the misclassification rates is not very large.