Study On Partial Discharge Fault Diagnosis Based On Support Vector Machine And Information Fusion

Along with the increase of power level and capacity in power systems, partial discharge has become a major reason for insulating degradation.80% defects was caused by transformer insulation problems, therefore PD detection has become a major monitoring method. There are several PD types in the transformer insulation system, each of which affects the insulating condition in varied degrees. Therefore correctly recognizing PD models can help the fault to be properly handled. Based on the principle and characteristics of PD, this dissertation carried on the following research.The dual sensor detection system of transformer PD test system was constructed with the application of IEC60270 method with four PD models of suspension discharge, point-plane discharge, surface discharge and gap discharge. Then the PD features are extracted from the raw signal by two methods, Phase Resolved Partial Discharge (PRPD) and Time Resolved Partial Discharge (TRPD). Under these two models, the relevance and difference between the signals captured by different sensors are discussed, based on which the Grey Fractal Features and Cross-Wavelet Spectrum Features are calculated for further recognition processing.A Support Vector Machine based classifier was proposed. The optimal learning algorithm was employed to shorten the learning period of "binary tree" multi-class SVM classifier. Then the kernel parameters are optimized through Bare-Bone Particle Swarm Optimization (BBPSO). Then the evaluating indicators for SVM performance are researched along with the PD recognition system are constructed.An Optimal Radius Hyper-sphere SVM (ORH-SVM) is proposed, which is based on the principle of maximum interval of SVM and multiple classification methods. The ORH-SVM algorithm was proposed to solve the problems of recognition miss in traditional recognition methods. Simulation and experiments prove that the ORH-SVM algorithm for classifier performs better recognition rate and shorter training time than the traditional SVM algorithm.In view of the advantage of multi-information fusion, an information fusion recognition system of partial discharge based on DS evidence theory is constructed. Multi information fusion can resolve the collision problem. Tests show that information fusion inproves the recognition stability and reliability.A Combined Kernel Multiclass SVM (CKM-SVM) is constructed to achieve the direct multi-classification of PD models. Four PD feature spaces are constructed and mapped to different SVM kernel functions individually; then each kernel function is optimized via BBPSO, and then the weight coefficients for CKM-SVM model are calculated. Tests show that CKM-SVM performs good feature spatial fusion. Tests show that the recognition accuracy precedes traditional multiclass SVM and training period is shortened.