Intelligent Classifiers in Distinguishing Transformer Faults Using Frequency Response Analysis

With the expansion of the use of frequency response analysis (FRA) as a reliable tool for fault detection in transformers, more capabilities of this method are discovered every day. So that today the number of transformer faults that can be identified by FRA method has also increased. One of the most critical steps in fault detection with FRA is to distinguish faults and classify them in different classes. In this paper, well-known intelligent classifiers (probabilistic neural network, decision tree, support vector machine, and k-nearest neighbors) are used to classify transformer faults. For this purpose, the necessary measurements are performed on the model transformers under the healthy condition and under different fault conditions (axial displacement, radial deformation, disc space variation, short-circuits, and core deformation). Then, by dividing the frequency ranges of the measured transfer functions of the transformer, a new feature based on numerical and statistical indices for training and validation of classifiers is proposed. After completing the training process, the performance of the classifiers is evaluated and compared by applying the data obtained from real transformers.