Feature extraction and pattern recognition of signals radiated from partial discharge

The artificial neural networks based BP algorithm is used to recognize two typical discharge patterns, corona and spark. In order to have a comparison, feature extraction based on waveform parameter and time-frequency analysis were used separately to provide the training input. The results show that the highest average recognition rate based on waveform parameter reaches 92.5%, while this based on time-frequency is 95%. On the contrary, the lowest average recognition rate based on waveform parameter is 70%, while this based on time-frequency is 90%. This indicates that time-frequency analysis is more effective and more suitable for discharge pattern recognition.