Study On On-line PD Monitoring Of The Stator Winding Of Large Generators

The large generator is one of the most important devices in power system. Partial discharge (PD) in large generators can lead to deterioration of the insulation, resulting in insulation breakdown and short circuit. Traditional off-line test cannot present the real status of generators in operating station, so it cannot ensure safe-operation of generators. Insulation faults were still reported now and then. It's necessary to develop a complete and practicable on-line monitoring system of partial discharge in large generator insulation.On the base of many documents, three different styles of PD, including the causes of PD and the characteristics of PD, are studied in this paper. Besides CA model, ES model, SL model are set up. A wide frequency band sensor is used to get signal of on-line PD monitoring. This sensor gets the plenty of noise at the same time of collecting PD signal. PD signal is weak signal, so the first problem of on-line monitoring is how to pick up the weak anti-stationary random signal.Due to the variety of interference existing in the field, a measurement aimed to the processing of signals hierarchically is given in this paper. Firstly, due to the distinct difference in frequency domain between periodical narrow-band interferences and PD pulse signals, a threshold-curve method based on self-organizing neural network is adapted in the frequency domain. Secondly, because of the similarity of white noises and PD pulse signals, wavelet transform with good time-frequency analysis performance is adapted to eliminate white noises.PD signals and white noise can be distinguished since their wavelet coefficients are contrary in wavelet transform. Namely, with the increase of the scale, wavelet coefficients of PD signals will rise, and white noise will descend. So PD signals can be extracted by choosing suitable wavelet function. With the development of the technology, the methods processing signals will be more perfect, so there is a practical further with this measurement.In order to reduce the quantity and dimension of PD characteristic, the statistical operators based on the extracted PD signals are used to effectively figure the character of PD signals. And using BP neural network, which its inputs are statistical operators, also recognizes PD patterns. The simulation result shows that BP NN can recognize different PD patterns successfully.