Feature Extraction And Pattern Recognition Of Partial Discharge Electromagnetic Signals In Transformer

The transformer is the hub of power system, the reliability of power system depends on its reliable and safety operation. Insulation aging is the main reason of transformer faults. Partial discharge is the presymptom and manifestation of insulation deterioration in transformer. The study of partial discharge signals has great significance. Much exploration and research on de-nosing methods, feature extraction methods and pattern recognition methods for partial discharge were discussed in this thesis. The main works of this thesis are as follow:In this thesis, three kinds of partial discharge signals were analyzed, which are air gap discharge, surface discharge and corona discharge. The three kinds of artificial insulation defect models were built, and the electromagnetic signals of the three types of partial discharge were collected by experiments. The measured signals contain noise disturbance, so the noise must be removed at first. In this thesis, due to the deficiency of soft threshold and hard threshold function of the wavelet packet de-nosing algorithm, the threshold function is improved. The de-noising effect of improved wavelet packet threshold function is promoted through the analysis of the simulation of partial discharge signals. But wavelet packet is lacking in adaptability. In addition, the partial discharge electromagnetic signals of the transformer are related to the measuring system and environment, which creates some difficulties in the selection of the wavelet packet basis function. So EMD de-noising algorithm based on the combination of auto correlation function and interval threshold was applied to the de-noising of transformer. Through the analysis of the simulation of partial discharge signals and the actual measurement signals, the de-nosing effect is much better by EMD de-noising algorithm based on the combination of auto correlation function and interval threshold.The partial discharge signals after de-noising were separated into intrinsic mode functions by EMD. The intrinsic mode functions that have main information were chosen by correlation coefficient method. Two characteristic parameters of intrinsic mode functions were extracted, which were fractal dimension and singular value.The fractal dimension and the singular value were used as the classification feature, and the BP neural network were used to recognize the types of partial discharge. The recognition results show that the two kinds of characteristic parameters can recognize the partial discharge signals, and the recognition rate of the singular value is higher than that of the fractal dimension.In addition, the least square support vector machine(LSSVM) was used to recognize the partial discharge in this thesis. Grid search algorithm was adopted to realize parameter optimization by the way from roughly search to fine search. The optimal model is determined by 10-fold cross-validation method. M-ary classification was used as multi-classification. Partial discharge electromagnetic signals were classified by SVM and LSSVM. The recognition results showed that the recognition rate of the two kinds of characteristic quantity were high. In addition, recognition rate by LSSVM is higher than that by SVM. The training time of LSSVM is faster than that of SVM. Finally, comparing the LSSVM with BP neural network algorithm, the recognition rate of the singular value is higher than that of the fractal dimension by these algorithms, LSSVM is more suitable for pattern recognition of partial discharge electromagnetic signals.