On-line diagnosing of GIS (gas insulated switchgears) requires the pattern classification and identification of signals that are emitted from GIS. To classify the patterns correctly, substantial data sets that are emitted by artificially mimicked defects in GIS are needed. Applying the neural networks to the data sets, in general, identification methods of defects in GIS have widely been developed. Some identification system shows a good success such that the misclassification rate is reduced to below 5%; the key features in identification, however, are not obviously revealed in neural networks systems because of nonlinear network structures. The decision tree method that classifies the signals using the feature rules in plain graphical representations can explains the classification rules in clear forms. We applied the decision tree classification method to the signals emitted from the signals by artificially prepared defects in GIS, and find that the method shows a good classification rates over 95% which are comparable to that in neural networks. We also discuss the robustness from noise, and compare the results of the misclassification rates by the two methods.