Neuro fuzzy recognition of ultra-high frequency partial discharges in transformers

In this paper, partial discharge (PD) signals in the ultra-high frequency (UHF) range were investigated. A spectrum analyzer functioning in the zero span mode was applied to capture and record the PD signal component at a specific frequency over a time interval. Different PD sources produce different PD patterns, thus it is possible to recognize the PD sources from the captured PD patterns. Here, the PD patterns produced by 3 different laboratory models representing defects in transformer windings (void, floating metal, and surface discharge) are recorded and analyzed. From the PD pattern data, 6 features are extracted using 3 statistical parameters, i.e. mean, skewness and kurtosis for both positive and negative voltage halfcycles. The 6 features were used to recognize the PD sources by applying neuro fuzzy method to classify the PD pattern. ANFIS, a MatLab function, was used to train the fuzzy inference system (FIS). The trained FIS was then used to recognize the source of the PDs. Result shows the trained FIS has a high success rate to recognize and thus classify the PD sources.