An efficient diagnosis method for data mining on single PD pulses of transformer insulation defect models

Reviewing the various Partial Discharges (PD data mining researches which have been reported so far, this study compares the performance of different feature spaces and different classifiers employed for PD classification in insulation condition monitoring of power transformers. In this process, first a knowledge basis is developed through construction of 4 different types of PD models in the high voltage laboratory. Background noise is considered as one class in this knowledge basis. The high frequency time domain current signals of high voltage equipment are captured over one power frequency cycle. The single PD activities within this captured signal are extracted by application of a threshold-based method. Four popular feature extraction methods i.e. Statistical, texture, FFT and Cepstral features are applied on these recorded extracted PD signals. To distinguish the different PD types, three conceptually different classifier types, Neural Network, Decision Tree, and k-nearest neighbours, are applied on the recorded feature spaces. Using Bayesian theory, a performance analysis is carried out to find whether the classifiers are over-fitted or not. Although, the most reliable data mining tool found to be a combination of a Cepstral feature space, and neural network classifier however, since the statistical features can be computed very fast it is employed in this work. Next, it is proposed to use a cascade PD identifier to find whether the detected signal is noise or not. And if it is PD, employing Cepstral feature space knowledge-basis, its type is identified.